SESA: GitHub'dan güncel dosyalar aktarıldı - 2026-03-18 23:04

assets/config.json

@@ -2,5 +2,10 @@
   "lang": {
     "override": true,
     "selected_lang": "en_us"
+  },
+  "sharing": {
+    "method": "gradio",
+    "ngrok_token": "",
+    "port": 8142
   }
 }

assets/custom_models.json

@@ -0,0 +1 @@
+{}

assets/i18n/languages/ar_sa.json

@@ -17,11 +17,10 @@
   "overlap": "التداخل",
   "overlap_info": "موصى به: 2-10 (القيم الأعلى تزيد من الجودة ولكن تتطلب المزيد من VRAM)",
   "tta_boost": "تعزيز TTA",
-  "tta_info": "يحسن الجودة ولكنه يبطئ سرعة المعالجة",
   "phase_fix": "تصحيح الطور",
+  "use_phase_correction": "تطبيق تصحيح طور متقدم",
   "phase_fix_info": "تصحيح طور متقدم للمسارات الآلية",
   "instrumental": "آلي",
-  "instrumental_info": "عادةً يعطي مخرجين، لكن هناك أوقات يعطي فيها مخرجًا واحدًا، تحتاج إلى التجربة :)",
   "process": "🚀 معالجة",
   "reset": "🧹 إعادة تعيين",
   "status": "الحالة",
@@ -71,99 +70,210 @@
   "output_preview": "معاينة الإخراج",
   "refresh_output": "🔄 تحديث الإخراج",
   "ensemble_copy_status": "سيتم نسخ مخرجات التجميع هنا...",
-  "guidelines": "<h4>إرشادات اختيار النموذج</h4><ul><li><strong>تجنب الخلط بين الفئات:</strong> قد يؤدي الجمع بين نماذج الأصوات والآلات إلى خلط غير مرغوب فيه</li><li><strong>ملاحظات نموذج خاصة:</strong><ul><li>نماذج الازدواجية (v1/v2) - تخرج كلا الجذعين</li><li>MDX23C Separator - نتائج هجينة</li></ul></li><li><strong>أفضل الممارسات:</strong> استخدم 3-5 نماذج متشابهة من نفس الفئة</li></ul><div>💡 نصيحة احترافية: ابدأ بتركيبة \"VOCALS-MelBand-Roformer BigBeta5e\" + \"VOCALS-BS-Roformer_1297\"</div>",
-  "download_sources_tab": "⬇️ تحميل المصادر",
-  "direct_links": "🌐 روابط مباشرة",
+  "guidelines": "<h4>إرشادات اختيار النموذج</h4><ul><li><strong>تجنب الخلط بين الفئات:</strong> قد يؤدي الجمع بين نماذج الأصوات والآلات إلى خلط غير مرغوب فيه</li><li><strong>ملاحظات نموذج خاصة:</strong><ul><li>نماذج الازدواجية (v1/v2) - تنتج كلا الستيمات</li><li>فاصل MDX23C - نتائج هجينة</li></ul></li><li><strong>أفضل ممارسة:</strong> استخدم 3-5 نماذج مشابهة من نفس الفئة</li></ul><div>💡 نصيحة محترف: ابدأ بمزيج \"VOCALS-MelBand-Roformer BigBeta5e\" + \"VOCALS-BS-Roformer_1297\"</div>",
+  "download_sources_tab": "⬇️ تنزيل المصادر",
+  "direct_links": "🌐 الروابط المباشرة",
   "audio_file_url": "رابط ملف الصوت",
-  "download_from_url": "⬇️ تحميل من الرابط",
-  "download_status": "حالة التحميل",
+  "download_from_url": "⬇️ تنزيل من الرابط",
+  "download_status": "حالة التنزيل",
   "downloaded_file": "الملف المحمل",
   "cookie_management": "🍪 إدارة الكوكيز",
-  "upload_cookies_txt": "رفع ملف Cookies.txt",
-  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 لماذا هو ضروري؟**<br>- الوصول إلى المحتوى المقيد بالعمر<br>- تحميل الفيديوهات الخاصة/غير المدرجة<br>- تجاوز القيود الإقليمية<br>- تجنب حدود تحميل يوتيوب<br><br>**⚠️ ملاحظات هامة**<br>- لا تشارك ملفات الكوكيز الخاصة بك أبدًا!<br>- قم بتحديث الكوكيز عندما:<br>  • تظهر أخطاء \"403 ممنوع\"<br>  • تتوقف التنزيلات فجأة<br>  • تظهر رسائل \"انتهت الجلسة\"<br><br>**🔄 خطوات التجديد**<br>1. قم بتثبيت هذا <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">امتداد كروم</a><br>2. سجل الدخول إلى يوتيوب في كروم<br>3. انقر على أيقونة الامتداد → \"تصدير\"<br>4. ارفع الملف المحمل هنا<br><br>**⏳ عمر الكوكيز**<br>- الجلسات العادية: 24 ساعة<br>- العمليات الحساسة: ساعة واحدة<br>- تغييرات كلمة المرور: إبطال فوري</div>",
+  "upload_cookies_txt": "رفع Cookies.txt",
+  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 لماذا مطلوب؟**<br>- الوصول إلى المحتوى المقيد بالعمر<br>- تنزيل الفيديوهات الخاصة/غير المدرجة<br>- تجاوز القيود الإقليمية<br>- تجنب حدود تنزيل يوتيوب<br><br>**⚠️ ملاحظات مهمة**<br>- لا تشارك ملفات الكوكيز الخاصة بك أبدًا!<br>- قم بتحديث الكوكيز عند:<br>  • الحصول على أخطاء \"403 ممنوع\"<br>  • توقف التنزيلات فجأة<br>  • رؤية رسائل \"انتهت صلاحي�� الجلسة\"<br><br>**🔄 خطوات التحديث**<br>1. قم بتثبيت <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">إضافة كروم هذه</a><br>2. تسجيل الدخول إلى يوتيوب في كروم<br>3. انقر على أيقونة الإضافة → \"تصدير\"<br>4. قم برفع الملف المحمل هنا<br><br>**⏳ عمر الكوكيز**<br>- الجلسات العادية: 24 ساعة<br>- العمليات الحساسة: 1 ساعة<br>- تغيير كلمة المرور: تحديث فوري</div>",
+  "ensemble_output_copied": "✅ تم نسخ إخراج التجميع إلى {}",
+  "error_copying_ensemble_output": "❌ خطأ في نسخ إخراج التجميع: {}",
+  "drive_mounted_copying_ensemble": "تم تحميل Google Drive. نسخ إخراج التجميع...",
+  "drive_already_mounted_copying_ensemble": "Google Drive محمل بالفعل. نسخ إخراج التجميع...",
+  "no_ensemble_output_files_found": "❌ لم يتم العثور على ملفات إخراج تجميع.",
+  "drive_mounted_copying_files": "تم تحميل Google Drive. نسخ الملفات...",
+  "drive_already_mounted_copying_files": "Google Drive محمل بالفعل. نسخ الملفات...",
+  "files_copied_to_drive": "✅ تم نسخ الملفات إلى {}",
+  "error_copying_files": "❌ خطأ في نسخ الملفات: {}",
+  "mounting_drive": "تحميل Google Drive...",
+  "settings_saved": "تم حفظ الإعدادات في config.json: اللغة={lang}, الطريقة={method}, المنفذ={port}",
+  "selected_language": "اللغة المختارة: {lang}",
+  "available_languages": "اللغات المتاحة: {langs}",
+  "select_language_prompt": "اختر لغة (افتراضي: {default}): ",
+  "invalid_language": "اختيار لغة غير صالح! استخدام اللغة الافتراضية ({default}).",
+  "available_sharing_methods": "طرق المشاركة المتاحة: gradio، localtunnel، ngrok",
+  "select_sharing_method_prompt": "اختر طريقة مشاركة (الافتراضي: {default}): ",
+  "enter_ngrok_token_prompt": "أدخل رمز Ngrok الخاص بك (الافتراضي: {default}): ",
+  "ngrok_token_required": "خطأ: رمز Ngrok مطلوب لطريقة ngrok!",
+  "enter_port_prompt": "أدخل رقم المنفذ (الافتراضي: {default}): ",
+  "opening_previous_url": "فتح الرابط السابق: {url}",
+  "open_url_manually": "افتح الرابط يدويًا في Colab: {url}",
+  "invalid_previous_url": "رابط سابق غير صالح: {url}، لم يتم فتحه.",
+  "starting_method": "بدء {method} على المنفذ {port}...",
+  "process_stopped": "\n🛑 تم إيقاف العملية من قبل المستخدم",
+  "selected_port": "المنفذ المختار: {port}",
+  "starting_gradio_with_sharing": "بدء Gradio مع المشاركة المدمجة...",
+  "starting_localtunnel": "بدء Localtunnel على المنفذ {port}...",
+  "share_link": "رابط المشاركة: {url}",
+  "password_ip": "كلمة مرور IP: {ip}",
+  "starting_ngrok": "بدء Ngrok على المنفذ {port}...",
+  "ngrok_url": "رابط Ngrok: {url}",
+  "ngrok_error": "خطأ في بدء Ngrok: {error}",
+  "auto_apollo_chunk_size": "حجم قطعة Apollo",
+  "auto_apollo_chunk_size_info": "حجم القطعة لتحسين Apollo (موصى به: 19 للنموذج العام)",
+  "auto_apollo_overlap": "تداخل Apollo",
+  "auto_enhancing_with_apollo": "تحسين مع Apollo ({}/{} ملفات)...",
+  "auto_apollo_processing_completed": "اكتملت معالجة Apollo!",
+  "auto_apollo_overlap_info": "التداخل لتحسين Apollo (موصى به: 2)",
+  "apollo_chunk_size": "حجم قطعة Apollo",
+  "apollo_chunk_size_info": "حجم القطعة لتحسين Apollo (موصى به: 19 للنموذج العام)",
+  "apollo_overlap": "تداخل Apollo",
+  "enhance_with_apollo": "تحسين مع Apollo",
+  "apollo_enhancement_info": "تحسين جودة الصوت",
+  "enhancing_with_apollo": "تحسين مع Apollo ({}/{} ملفات)...",
+  "apollo_processing_completed": "اكتملت معالجة Apollo!",
+  "apollo_overlap_info": "التداخل لتحسين Apollo (موصى به: 2)",
+  "selected_models": "النماذج المختارة",
+  "save_preset": "حفظ الإعداد المسبق",
+  "delete_preset": "حذف الإعداد المسبق",
+  "refresh_presets": "تحديث الإعدادات المسبقة",
+  "preset_name": "اسم الإعداد المسبق",
+  "select_preset": "اختيار إعداد مسبق",
+  "add_favorite": "إضافة إلى المفضلة",
+  "apply_matchering": "تطبيق Matchering",
+  "matchering_info": "تطبيق Matchering لتحسين جودة الصوت وتوازنه",
+  "matchering_passes": "تمريرات Matchering",
+  "matchering_passes_info": "عدد التكر��رات لـ Matchering (1-5، القيم الأعلى تحسن الجودة ولكن تستغرق وقتًا أطول)",
+  "matchering_processing": "جارٍ تطبيق Matchering... ({:.1f}%)",
+  "matchering_completed": "✅ اكتمل معالجة Matchering!",
+  "matchering_error": "❌ حدث خطأ أثناء Matchering: {}",
+  "invalid_method": "خطأ: طريقة غير صالحة! استخدم 'gradio'، 'localtunnel'، أو 'ngrok'.",
+  "language_changed_message": "تم تغيير اللغة. يرجى إعادة تشغيل الخلية.",
+  "backend_settings": "🚀 إعدادات الخلفية",
+  "inference_backend": "خلفية الاستدلال",
+  "optimization_mode": "وضع التحسين",
+  "optimization_mode_info": "اختيار وضع تحسين PyTorch",
+  "mixed_precision_amp": "الدقة المختلطة (AMP)",
+  "mixed_precision_info": "استدلال أسرع 2x - موصى به",
+  "tf32_acceleration": "تسريع TF32",
+  "tf32_acceleration_info": "زيادة سرعة إضافية لـ RTX 30xx+",
+  "cudnn_benchmark": "معيار cuDNN",
+  "cudnn_benchmark_info": "أبطأ في التشغيل الأول، أسرع بكثير في التشغيلات اللاحقة",
+  "ultra_optimized_pytorch": "خلفية PyTorch فائقة التحسين",
+  "default_active_max_speed": "نشط افتراضيًا - تحسين سرعة قصوى",
+  "channels_last_mode": "الأسرع لـ RTX GPUs",
+  "compile_mode": "سرعة إضافية لـ PyTorch 2.0+",
+  "default_mode": "قياسي",
+  "tta_info": "يزيد من الجودة لكنه يبطئ المعالجة",
+  "direct_download_error": "❌ خطأ في التنزيل المباشر: {}",
+  "invalid_google_drive_url": "❌ رابط Google Drive غير صالح",
+  "continuing_without_google_drive": "فشل الاتصال بـ Google Drive، الاستمرار بدونه...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive غير محمل، تخطي النسخ إلى Drive...",
+  "direct_download_failed": "فشل تنزيل الملف مباشرة",
+  "instrumental_info": "عادةً ينتج 2 ستيمات، لكن أحيانًا واحدة فقط - تحتاج إلى التجربة! :)",
   "manual_ensemble_tab": "🎚️ التجميع اليدوي",
+  "phase_fixer_tab": "🔧 مصحح الطور",
+  "batch_processing_tab": "📦 المعالجة الدفعية",
+  "batch_description": "معالجة ملفات متعددة في وقت واحد بنفس النموذج",
+  "batch_input_folder": "مجلد الإدخال",
+  "batch_input_folder_placeholder": "أدخل مسار المجلد الذي يحتوي على ملفات الصوت",
+  "batch_output_folder": "مجلد الإخراج",
+  "batch_output_folder_placeholder": "أدخل مسار المجلد لملفات الإخراج",
+  "batch_file_list": "الملفات المراد معالجتها",
+  "batch_add_files": "إضافة ملفات",
+  "batch_clear_list": "مسح القائمة",
+  "batch_start": "🚀 بدء المعالجة الدفعية",
+  "batch_stop": "⏹️ إيقاف",
+  "batch_progress": "تقدم المعالجة الدفعية",
+  "batch_current_file": "يُعالَج حاليًا",
+  "batch_completed": "✅ اكتملت المعالجة الدفعية!",
+  "batch_stopped": "⏹️ تم إيقاف المعالجة الدفعية",
+  "batch_no_files": "❌ لم يتم اختيار ملفات للمعالجة",
+  "source_model": "النموذج المصدر (أصوات)",
+  "source_model_info": "نموذج لاستخراج الأصوات/بيانات الطور",
+  "target_model": "النموذج الهدف (آلي)",
+  "target_model_info": "النموذج الذي يتم تطبيق تصحيح الطور عليه",
   "input_sources": "📂 مصادر الإدخال",
   "refresh": "🔄 تحديث",
   "ensemble_algorithm": "خوارزمية التجميع",
-  "select_audio_files": "اختر ملفات الصوت",
+  "select_audio_files": "اختيار ملفات الصوت",
   "available_files": "الملفات المتاحة",
-  "custom_weights": "أوزان مخصصة (مفصولة بفواصل)",
-  "custom_weights_placeholder": "مثال: 0.8، 1.2، 1.0، ...",
-  "custom_weights_info": "اتركه فارغًا لأوزان متساوية",
+  "custom_weights": "أوزان مخصصة (مفصولة بفاصلة)",
+  "custom_weights_placeholder": "مثال: 0.8, 1.2, 1.0, ...",
+  "custom_weights_info": "اتركه فارغًا للأوزان المتساوية",
   "result_preview_tab": "🎧 معاينة النتيجة",
-  "ensembled_output": "مخرج التجميع",
+  "ensembled_output": "الإخراج المجمَّع",
   "processing_log_tab": "📋 سجل المعالجة",
   "processing_details": "تفاصيل المعالجة",
   "process_ensemble": "⚡ معالجة التجميع",
   "languages_tab": "🌐 اللغات",
   "Vocal Models": "نماذج الأصوات",
-  "Instrumental Models": "نماذج الآلات",
-  "4-Stem Models": "نماذج 4 جذور",
+  "Instrumental Models": "النماذج الآلية",
+  "4-Stem Models": "نماذج 4 مسارات",
   "Denoise Models": "نماذج إزالة الضوضاء",
   "Dereverb Models": "نماذج إزالة الصدى",
-  "Other Models": "نماذج أخرى",
-  "ensemble_files_description": "يجمع عدة ملفات صوتية في مخرج واحد باستخدام طرق التجميع المحددة.",
+  "Other Shapes": "نماذج أخرى",
+  "Custom Models": "النماذج المخصصة",
+  "ensemble_files_description": "يجمع ملفات صوتية متعددة في إخراج واحد باستخدام طرق تجميع محددة.",
   "ensemble_files_help": "مسار جميع ملفات الصوت للتجميع",
-  "ensemble_type_help": "واحد من avg_wave، median_wave، min_wave، max_wave، avg_fft، median_fft، min_fft، max_fft",
-  "ensemble_weights_help": "الأوزان لإنشاء التجميع. يجب أن يتساوى عدد الأوزان مع عدد الملفات",
-  "ensemble_output_help": "مسار ملف WAV حيث سيتم حفظ نتيجة التجميع",
+  "ensemble_type_help": "أحد: avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
+  "ensemble_weights_help": "الأوزان لإنشاء التجميع. يجب أن يساوي عدد الأوزان عدد الملفات",
+  "ensemble_output_help": "مسار ملف WAV حيث سيتم تخزين نتيجة التجميع",
   "ensemble_type_print": "نوع التجميع: {}",
   "num_input_files_print": "عدد ملفات الإدخال: {}",
   "weights_print": "الأوزان: {}",
   "output_file_print": "ملف الإخراج: {}",
   "duration_mismatch_error": "يجب أن تكون جميع الملفات بنفس المدة",
-  "file_not_found_error": "خطأ. لا يمكن العثور على الملف: {}. تحقق من المسارات.",
+  "file_not_found_error": "خطأ. تعذر العثور على الملف: {}. تحقق من المسارات.",
   "reading_chunk_print": "قراءة قطعة من الملف: {} (البداية: {}ث، المدة: {}ث)",
   "chunk_result_shape_print": "شكل نتيجة القطعة: {}",
   "ensemble_completed_print": "اكتمل التجميع. تم حفظ الإخراج في: {}",
-  "file_deletion_error": "لا يمكن حذف {}: {}",
-  "directory_not_exist_warning": "⚠️ الدليل غير موجود: {}",
-  "not_a_directory_warning": "⚠️ المسار ليس دليلًا: {}",
+  "file_deletion_error": "تعذر حذف {}: {}",
+  "directory_not_exist_warning": "⚠️ المجلد غير موجود: {}",
+  "not_a_directory_warning": "⚠️ المسار ليس مجلدًا: {}",
   "item_deletion_error": "⚠️ خطأ في حذف {}: {}",
+  "critical_error": "⛔ خطأ فادح: {}",
   "old_output_not_exist": "❌ مجلد الإخراج القديم غير موجود",
-  "old_outputs_cleared": "✅ تم مسح المخرجات القديمة بنجاح!",
+  "old_outputs_cleared": "✅ تم مسح الإخراجات القديمة بنجاح!",
   "error": "🔥 خطأ: {}",
   "ffmpeg_error": "خطأ FFmpeg ({}): {}",
   "file_saved_successfully": "تم حفظ الملف بنجاح: {}",
-  "total_files_found": "إجمالي الملفات الموجودة: {}. باستخدام معدل العينة: {}",
-  "total_progress": "التقدم الإجمالي",
-  "detailed_pbar_enabled": "تم تفعيل شريط التقدم التفصيلي: {}",
-  "loaded_audio": "الصوت المحمل: {}، الشكل: {}",
-  "cannot_read_track": "لا يمكن قراءة المسار: {}",
+  "total_files_found": "إجمالي الملفات الموجودة: {}. معدل العينات المستخدم: {}",
+  "total_progress": "التقدم الكلي",
+  "detailed_pbar_enabled": "شريط التقدم التفصيلي مُفعَّل: {}",
+  "loaded_audio": "تم تحميل الصوت: {}، الشكل: {}",
+  "cannot_read_track": "تعذر قراءة المسار: {}",
   "error_message": "رسالة الخطأ: {}",
   "demudding_track": "تنظيف المسار (إعادة مزج الطور - آلي): {}",
   "elapsed_time": "الوقت المنقضي: {:.2f} ثانية.",
-  "proc_folder_description": "يعالج ملفات الصوت في مجلد باستخدام نموذج محدد.",
-  "model_type_help": "نوع النموذج (bandit، bs_roformer، mdx23c، إلخ.)",
+  "proc_folder_description": "معالجة ملفات الصوت في مجلد باستخدام نموذج محدد.",
+  "model_type_help": "نوع النموذج (bandit, bs_roformer, mdx23c، إلخ)",
   "config_path_help": "مسار ملف التكوين",
-  "demud_phaseremix_help": "تفعيل إعادة مزج الطور للآلي",
-  "start_checkpoint_help": "نقطة التفتيش الأولية للأوزان الصالحة",
-  "input_folder_help": "مجلد يحتوي على الخلطات للمعالجة",
+  "demud_phaseremix_help": "تمكين إعادة مزج الطور للآلات الموسيقية",
+  "start_checkpoint_help": "نقطة التفتيش الأولية للأوزان الصحيحة",
+  "input_folder_help": "المجلد الذي يحتوي على المزيجات للمعالجة",
   "audio_path_help": "مسار ملف صوتي واحد للمعالجة",
-  "store_dir_help": "مسار لتخزين النتائج",
+  "store_dir_help": "مسار تخزين النتائج",
   "device_ids_help": "قائمة معرفات GPU",
-  "extract_instrumental_help": "قلب الأصوات للحصول على الآلي إذا تم توفيره",
-  "disable_detailed_pbar_help": "تعطيل شريط التقدم التفصيلي في الفصل",
-  "force_cpu_help": "فرض استخدام CPU حتى لو كان CUDA متاحًا",
+  "extract_instrumental_help": "عكس الأصوات للحصول على الآلي إذا كان متاحًا",
+  "disable_detailed_pbar_help": "تعطيل شريط التقدم التفصيلي في demix",
+  "force_cpu_help": "إجبار استخدام CPU حتى لو كانت CUDA متاحة",
   "flac_file_help": "إخراج ملف FLAC بدلاً من WAV",
-  "export_format_help": "صيغة التصدير ونوع PCM",
+  "export_format_help": "تنسيق التصدير ونوع PCM",
   "pcm_type_help": "نوع PCM لملفات FLAC",
-  "use_tta_help": "تفعيل التكبير في وقت الاختبار",
+  "use_tta_help": "تمكين زيادة وقت الاختبار",
   "lora_checkpoint_help": "نقطة التفتيش الأولية لأوزان LoRA",
   "cuda_available": "CUDA متاح، استخدم --force_cpu لتعطيله.",
-  "using_device": "استخدام الجهاز: {}",
-  "instruments_print": "الآلات: {}",
+  "using_device": "الجهاز المستخدم: {}",
+  "instruments_print": "الأدوات الموسيقية: {}",
   "model_load_time": "وقت تحميل النموذج: {:.2f} ثانية",
   "invalid_url": "❌ رابط غير صالح",
   "cookie_file_updated": "✅ تم تحديث ملف الكوكيز!",
   "cookie_installation_error": "⚠️ خطأ في تثبيت الكوكيز: {}",
+  "google_drive_mount_error": "❌ خطأ في ربط Google Drive: {}",
   "file_size_zero_error": "حجم الملف صفر أو لم يتم إنشاء الملف",
-  "google_drive_error": "❌ خطأ في تحميل Google Drive: {}",
+  "google_drive_error": "❌ خطأ في تنزيل Google Drive: {}",
   "wav_conversion_failed": "فشل تحويل WAV",
-  "download_error": "❌ خطأ في التحميل: {}",
-  "download_success": "🎉 تم التحميل بنجاح!",
-  "download_failed": "❌ فشل التحميل",
-  "no_output_files_found": "❌ لم يتم العثور على ملفات إخراج.",
+  "download_error": "❌ خطأ في التنزيل: {}",
+  "file_copied_to_drive": "تم نسخ الملف إلى Google Drive: {}",
+  "copy_to_drive_error": "❌ خطأ في النسخ إلى Google Drive: {}",
+  "download_success": "🎉 تم التنزيل بنجاح!",
+  "download_failed": "❌ فشل التنزيل",
+  "no_output_files_found": "❌ لم يتم العثور على ملفات الإخراج.",
   "output_refreshed_successfully": "✅ تم تحديث الإخراج بنجاح!",
   "error_refreshing_output": "❌ خطأ في تحديث الإخراج: {}",
   "starting_audio_separation": "بدء فصل الصوت...",
@@ -172,109 +282,64 @@
   "separation_complete": "اكتمل الفصل!",
   "progress_parsing_error": "خطأ في تحليل التقدم: {}",
   "error_occurred": "حدث خطأ: {}",
-  "separation_process_completed": "اكتمل عملية الفصل!",
-  "no_audio_file_error": "❌ لم يتم توفير ملف صوتي ولا يوجد ملف موجود في دليل الإدخال.",
-  "no_input_progress_label": "خطأ: لم يتم توفير إدخال -- 0.0%",
+  "separation_process_completed": "اكتملت عملية الفصل!",
+  "no_audio_file_error": "❌ لم يتم تقديم ملف صوتي ولا يوجد ملف في مجلد الإدخال.",
+  "no_input_progress_label": "خطأ: لم يتم تقديم إدخال -- 0.0%",
   "processing_audio_print": "معالجة الصوت من: {} باستخدام النموذج: {}",
   "starting_audio_separation_progress_label": "بدء فصل الصوت... -- 0.0%",
   "separating_audio_progress_label": "فصل الصوت... -- {}.0%",
   "audio_processing_completed_progress_label": "اكتملت معالجة الصوت! -- 100.0%",
   "audio_processing_completed": "✅ اكتملت معالجة الصوت!",
   "error_occurred_progress_label": "حدث خطأ -- 0.0%",
-  "minimum_files_required": "⚠️ الحد الأدنى 2 ملف مطلوب",
+  "minimum_files_required": "⚠️ مطلوب ملفان على الأقل",
   "valid_files_not_found": "❌ لم يتم العثور على ملفات صالحة",
   "starting_ensemble_process": "بدء عملية التجميع...",
-  "ensembling_progress": "التجميع... ({:.1f}%)",
+  "ensembling_progress": "جارٍ التجميع... ({:.1f}%)",
   "finalizing_ensemble_output": "إنهاء إخراج التجميع...",
   "success_log": "✅ نجاح!\n{}",
   "error_log": "❌ خطأ!\n{}",
-  "critical_error": "⛔ خطأ حرج: {}",
   "ensemble_process_completed": "اكتملت عملية التجميع!",
-  "no_models_selected": "❌ لم يتم اختيار نماذج",
-  "no_input_audio_provided": "❌ لم يتم توفير صوت إدخال",
-  "loading_model": "تحميل النموذج {}/{}: {}...",
-  "loading_model_progress_label": "تحميل النموذج {}/{}: {}... -- {}.0%",
+  "no_models_selected": "❌ لم يتم اختيار أي نماذج",
+  "no_input_audio_provided": "❌ لم يتم تقديم صوت إدخال",
+  "loading_model": "جارٍ تحميل النموذج {}/{}: {}...",
+  "loading_model_progress_label": "جارٍ تحميل النموذج {}/{}: {}... -- {}.0%",
   "completed_model": "اكتمل النموذج {}/{}: {}",
   "completed_model_progress_label": "اكتمل النموذج {}/{}: {} -- {}.0%",
   "running_command": "تشغيل الأمر: {}",
   "model_failed": "فشل النموذج {}: {}",
-  "critical_error_with_model": "خطأ حرج مع {}: {}",
-  "model_output_failed": "فشل {} في إنتاج مخرج",
+  "critical_error_with_model": "خطأ فادح مع {}: {}",
+  "model_output_failed": "فشل {} في إنتاج الإخراج",
   "waiting_for_files": "في انتظار جاهزية جميع الملفات...",
   "waiting_for_files_progress_label": "في انتظار جاهزية جميع الملفات... -- 90.0%",
-  "performing_ensemble": "تنفيذ التجميع...",
-  "performing_ensemble_progress_label": "تنفيذ التجميع... -- 92.0%",
+  "performing_ensemble": "جارٍ التجميع...",
+  "performing_ensemble_progress_label": "جارٍ التجميع... -- 92.0%",
   "memory_usage_before_ensemble": "استخدام الذاكرة قبل التجميع: {}%",
   "memory_usage_after_ensemble": "استخدام الذاكرة بعد التجميع: {}%",
   "finalizing_ensemble_output_progress_label": "إنهاء إخراج التجميع... -- 98.0%",
-  "ensemble_file_creation_failed": "لم يتم إنشاء ملف التجميع: {}",
+  "ensemble_file_creation_failed": "تعذر إنشاء ملف التجميع: {}",
   "ensemble_completed_progress_label": "اكتمل التجميع بنجاح! -- 100.0%",
   "success_output_created": "✅ نجاح! تم إنشاء ملف الإخراج.",
-  "drive_mounted_copying_ensemble": "تم ربط Google Drive. يتم نسخ مخرج التجميع...",
-  "drive_already_mounted_copying_ensemble": "Google Drive مربوط بالفعل. يتم نسخ مخرج التجميع...",
-  "no_ensemble_output_files_found": "❌ لم يتم العثور على ملفات إخراج التجميع.",
-  "ensemble_output_copied": "✅ تم نسخ مخرج التجميع إلى {}",
-  "error_copying_ensemble_output": "❌ خطأ في نسخ مخرج التجميع: {}",
-  "drive_mounted_copying_files": "تم ربط Google Drive. يتم نسخ الملفات...",
-  "drive_already_mounted_copying_files": "Google Drive مربوط بالفعل. يتم نسخ الملفات...",
-  "files_copied_to_drive": "✅ تم نسخ الملفات إلى {}",
-  "error_copying_files": "❌ خطأ في نسخ الملفات: {}",
-  "mounting_drive": "ربط Google Drive...",
-  "settings_saved": "تم حفظ الإعدادات في config.json: اللغة={lang}، الطريقة={method}، المنفذ={port}",
-  "selected_language": "اللغة المختارة: {lang}",
-  "available_languages": "اللغات المتاحة: {langs}",
-  "select_language_prompt": "اختر لغة (الافتراضي: {default}): ",
-  "invalid_language": "اختيار لغة غير صالح! استخدام اللغة الافتراضية ({default}).",
-  "available_sharing_methods": "طرق المشاركة المتاحة: gradio، localtunnel، ngrok",
-  "select_sharing_method_prompt": "اختر طريقة مشاركة (الافتراضي: {default}): ",
-  "enter_ngrok_token_prompt": "أدخل رمز Ngrok الخاص بك (الافتراضي: {default}): ",
-  "ngrok_token_required": "خطأ: رمز Ngrok مطلوب لطريقة ngrok!",
-  "enter_port_prompt": "أدخل رقم المنفذ (الافتراضي: {default}): ",
-  "opening_previous_url": "فتح الرابط السابق: {url}",
-  "open_url_manually": "افتح الرابط يدويًا في Colab: {url}",
-  "invalid_previous_url": "رابط سابق غير صالح: {url}، لم يتم فتحه.",
-  "starting_method": "بدء {method} على المنفذ {port}...",
-  "process_stopped": "\n🛑 تم إيقاف العملية من قبل المستخدم",
-  "selected_port": "المنفذ المختار: {port}",
-  "starting_gradio_with_sharing": "بدء Gradio مع المشاركة المدمجة...",
-  "starting_localtunnel": "بدء Localtunnel على المنفذ {port}...",
-  "share_link": "رابط المشاركة: {url}",
-  "password_ip": "كلمة مرور IP: {ip}",
-  "starting_ngrok": "بدء Ngrok على المنفذ {port}...",
-  "ngrok_url": "رابط Ngrok: {url}",
-  "ngrok_error": "خطأ في بدء Ngrok: {error}",
-  "apollo_chunk_size": "حجم قطعة Apollo",
-  "apollo_chunk_size_info": "حجم القطعة لتحسين Apollo (موصى به: 19 للنموذج العام)",
-  "apollo_overlap": "تداخل Apollo",
-  "enhancing_with_apollo": "تحسين مع Apollo ({}/{} ملفات)...",
-  "apollo_processing_completed": "اكتملت معالجة Apollo!",
-  "apollo_overlap_info": "التداخل لتحسين Apollo (موصى به: 2)",
-  "invalid_method": "خطأ: طريقة غير صالحة! استخدم 'gradio'، 'localtunnel'، أو 'ngrok'.",
-  "apollo_enhancement_settings": "إعدادات تحسين Apollo",
-  "enhance_with_apollo": "تحسين مع Apollo",
-  "enhance_with_apollo_info": "تفعيل Apollo لتحسين الصوت بعد الفصل",
-  "apollo_method": "طريقة Apollo",
-  "apollo_method_info": "اختر طريقة المعالجة لـ Apollo",
-  "normal_method": "الطريقة العادية",
-  "mid_side_method": "طريقة الوسط/الجانب",
-  "apollo_normal_model": "نموذج Apollo العادي",
-  "apollo_normal_model_info": "النموذج المستخدم لمعالجة Apollo العادية",
-  "apollo_midside_model": "نموذج Apollo الوسط/الجانب",
-  "apollo_enhancement_info": "تحسين جودة الصوت",
-  "selected_models": "النماذج المختارة",
-  "save_preset": "حفظ الإعداد المسبق",
-  "delete_preset": "حذف الإعداد المسبق",
-  "refresh_presets": "تحديث الإعدادات المسبقة",
-  "preset_name": "اسم الإعداد المسبق",
-  "select_preset": "اختيار إعداد مسبق",
-  "add_favorite": "إضافة إلى المفضلة",
-  "apollo_midside_model_info": "النموذج المستخدم لمعالجة الوسط/الجانب (اختياري)",
-  "language_changed_message": "تم تغيير اللغة. يرجى إعادة تشغيل الخلية.",
-  "apply_matchering": "تطبيق Matchering",
-  "matchering_info": "تطبيق Matchering لتحسين جودة الصوت وتوازنه",
-  "matchering_passes": "تمريرات Matchering",
-  "matchering_passes_info": "عدد التكرارات لـ Matchering (1-5، القيم الأعلى تحسن الجودة ولكن تستغرق وقتًا أطول)",
-  "matchering_processing": "جارٍ تطبيق Matchering... ({:.1f}%)",
-  "matchering_completed": "✅ اكتمل معالجة Matchering!",
-  "matchering_error": "❌ حدث خطأ أثناء Matchering: {}"
+  "custom_models_tab": "⚡ النماذج المخصصة",
+  "add_custom_model": "➕ إضافة نموذج مخصص",
+  "custom_model_name": "اسم النموذج",
+  "custom_model_name_placeholder": "أدخل اسمًا لنموذجك",
+  "checkpoint_url": "رابط نقطة التفتيش (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "رابط التكوين (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "رابط ملف Python المخصص (اختياري)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "الكشف التلقائي عن نوع النموذج",
+  "model_type": "نوع النموذج",
+  "add_model_btn": "➕ إضافة نموذج",
+  "custom_models_list": "📋 قائمة النماذج المخصصة",
+  "delete_model": "🗑️ حذف",
+  "no_custom_models": "لم تتم إضافة نماذج مخصصة بعد",
+  "model_added_success": "✅ تم إضافة النموذج بنجاح!",
+  "model_deleted_success": "✅ تم حذف النموذج بنجاح!",
+  "model_add_error": "❌ خطأ في إضافة النموذج: {}",
+  "model_delete_error": "❌ خطأ في حذف النموذج: {}",
+  "refresh_models": "🔄 تحديث",
+  "custom_model_info": "أضف نماذج مخصصة عن طريق توفير روابط التنزيل. سيتم تنزيل النماذج تلقائيًا عند استخدامها.",
+  "select_model_to_delete": "اختر نموذجًا للحذف"
 }

assets/i18n/languages/de_de.json

@@ -1,7 +1,7 @@
 {
   "select_language": "Sprache auswählen",
   "gecekondu_production": "Gecekondu Production",
-  "ultimate_audio_separation": "Deine ultimative Audio-Trennung",
+  "ultimate_audio_separation": "Ihre ultimative Audio-Trennung",
   "presented_by": "Präsentiert von Gecekondu Production © 2025",
   "audio_separation_tab": "🎙️ Audio-Trennung",
   "input_model": "📥 Eingabe & Modell",
@@ -13,15 +13,16 @@
   "settings": "⚙ Einstellungen",
   "format": "Format",
   "chunk_size": "Chunk-Größe",
-  "chunk_size_info": "Nicht ändern, es sei denn, du hast spezifische Anforderungen",
+  "chunk_size_info": "Nicht ändern, es sei denn, Sie haben spezifische Anforderungen",
   "overlap": "Überlappung",
-  "overlap_info": "Empfohlen: 2-10 (Höhere Werte verbessern die Qualität, benötigen aber mehr VRAM)",
-  "tta_boost": "TTA Boost",
-  "tta_info": "Verbessert die Qualität, verlangsamt aber die Verarbeitungsgeschwindigkeit",
+  "overlap_info": "Empfohlen: 2-10 (höhere Werte erhöhen die Qualität, erfordern aber mehr VRAM)",
+  "tta_boost": "TTA-Boost",
+  "tta_info": "Erhöht die Qualität, verlangsamt aber die Verarbeitung",
   "phase_fix": "Phasenkorrektur",
-  "phase_fix_info": "Fortgeschrittene Phasenkorrektur für Instrumentalspuren",
+  "use_phase_correction": "Erweiterte Phasenkorrektur anwenden",
+  "phase_fix_info": "Erweiterte Phasenkorrektur für Instrumentalspuren",
   "instrumental": "Instrumental",
-  "instrumental_info": "Liefert normalerweise 2 Ausgaben, manchmal aber nur eine einzige Ausgabe, du musst es ausprobieren :)",
+  "instrumental_info": "Gibt normalerweise 2 Ausgaben, manchmal nur eine - Sie müssen es ausprobieren! :)",
   "process": "🚀 Verarbeiten",
   "reset": "🧹 Zurücksetzen",
   "status": "Status",
@@ -37,28 +38,28 @@
   "drums": "Schlagzeug",
   "bass": "Bass",
   "effects": "Effekte",
-  "advanced_tab": "⚙ Fortgeschritten",
+  "advanced_tab": "⚙ Erweitert",
   "phase_remix": "Phasen-Remix",
   "dry": "Trocken",
   "music": "Musik",
   "karaoke": "Karaoke",
-  "bleed": "Überlauf",
-  "copy_to_drive": "📂 Auf Laufwerk kopieren",
+  "bleed": "Übersprechen",
+  "copy_to_drive": "📂 Auf Drive kopieren",
   "copy_status": "Kopierstatus",
   "files_will_be_copied": "Dateien werden hierher kopiert...",
-  "processing_tip": "<b>🔈 Verarbeitungstipp:</b> Für verrauschte Ergebnisse verwende <code>bleed_suppressor_v1</code> oder <code>denoisedebleed</code> Modelle in der Kategorie <i>\"Rauschunterdrückung & Effektentfernung\"</i>, um die Ausgabe zu bereinigen",
+  "processing_tip": "<b>🔈 Verarbeitungstipp:</b> Für verrauschte Ergebnisse verwenden Sie die Modelle <code>bleed_suppressor_v1</code> oder <code>denoisedebleed</code> in der Kategorie <i>\"Entrauschen & Effektentfernung\"</i>, um die Ausgabe zu bereinigen",
   "waiting_for_processing": "Warten auf Verarbeitung...",
   "auto_ensemble_tab": "🤖 Auto-Ensemble",
   "upload_file": "Datei hochladen",
   "enter_file_path": "Oder Dateipfad eingeben",
   "file_path_placeholder": "Vollständigen Pfad zur Audiodatei eingeben",
-  "advanced_settings": "⚙️ Fortgeschrittene Einstellungen",
+  "advanced_settings": "⚙️ Erweiterte Einstellungen",
   "use_tta": "TTA verwenden",
   "instrumental_only": "Nur Instrumental",
   "auto_overlap": "Überlappung",
   "auto_chunk_size": "Chunk-Größe",
   "output_format": "Ausgabeformat",
-  "model_selection": "🧠 Modellwahl",
+  "model_selection": "🧠 Modellauswahl",
   "model_category": "Modellkategorie",
   "select_models": "Modelle aus Kategorie auswählen",
   "ensemble_settings": "⚡ Ensemble-Einstellungen",
@@ -68,10 +69,10 @@
   "original_audio_tab": "🔊 Original-Audio",
   "original_audio": "Original-Audio",
   "ensemble_result_tab": "🎚️ Ensemble-Ergebnis",
-  "output_preview": "Ausgabe-Vorschau",
+  "output_preview": "Ausgabevorschau",
   "refresh_output": "🔄 Ausgabe aktualisieren",
   "ensemble_copy_status": "Ensemble-Ausgabe wird hierher kopiert...",
-  "guidelines": "<h4>Richtlinien zur Modellwahl</h4><ul><li><strong>Vermeide das Mischen von Kategorien:</strong> Die Kombination von Gesangs- und Instrumentalmodellen kann unerwünschte Mischungen erzeugen</li><li><strong>Besondere Modellhinweise:</strong><ul><li>Dualitätsmodelle (v1/v2) - Geben beide Stems aus</li><li>MDX23C Separator - Hybride Ergebnisse</li></ul></li><li><strong>Beste Praxis:</strong> Verwende 3-5 ähnliche Modelle aus derselben Kategorie</li></ul><div>💡 Profi-Tipp: Beginne mit der Kombination \"VOCALS-MelBand-Roformer BigBeta5e\" + \"VOCALS-BS-Roformer_1297\"</div>",
+  "guidelines": "<h4>Richtlinien zur Modellauswahl</h4><ul><li><strong>Kategorien nicht mischen:</strong> Das Kombinieren von Gesangs- und Instrumentalmodellen kann unerwünschte Mischungen erzeugen</li><li><strong>Hinweise zu Spezialmodellen:</strong><ul><li>Dualitätsmodelle (v1/v2) - Geben beide Stems aus</li><li>MDX23C Separator - Hybride Ergebnisse</li></ul></li><li><strong>Best Practice:</strong> Verwenden Sie 3-5 ähnliche Modelle aus derselben Kategorie</li></ul><div>💡 Pro-Tipp: Starten Sie mit der Kombination \"VOCALS-MelBand-Roformer BigBeta5e\" + \"VOCALS-BS-Roformer_1297\"</div>",
   "download_sources_tab": "⬇️ Quellen herunterladen",
   "direct_links": "🌐 Direkte Links",
   "audio_file_url": "Audio-Datei-URL",
@@ -80,7 +81,85 @@
   "downloaded_file": "Heruntergeladene Datei",
   "cookie_management": "🍪 Cookie-Verwaltung",
   "upload_cookies_txt": "Cookies.txt hochladen",
-  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 Warum nötig?**<br>- Zugriff auf altersbeschränkte Inhalte<br>- Herunterladen privater/nicht gelisteter Videos<br>- Umgehen regionaler Einschränkungen<br>- Vermeiden von YouTube-Download-Limits<br><br>**⚠️ Wichtige Hinweise**<br>- TEILE NIEMALS deine Cookie-Dateien!<br>- Aktualisiere Cookies, wenn:<br>  • \"403 Forbidden\"-Fehler auftreten<br>  • Downloads plötzlich stoppen<br>  • \"Sitzung abgelaufen\"-Meldungen erscheinen<br><br>**🔄 Erneuerungsschritte**<br>1. Installiere diese <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">Chrome-Erweiterung</a><br>2. Melde dich bei YouTube in Chrome an<br>3. Klicke auf das Erweiterungssymbol → \"Exportieren\"<br>4. Lade die heruntergeladene Datei hier hoch<br><br>**⏳ Cookie-Lebensdauer**<br>- Normale Sitzungen: 24 Stunden<br>- Sensible Operationen: 1 Stunde<br>- Passwortänderungen: Sofortige Ungültigkeit</div>",
+  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 Warum notwendig?**<br>- Zugriff auf altersbeschränkte Inhalte<br>- Herunterladen privater/nicht gelisteter Videos<br>- Umgehung regionaler Einschränkungen<br>- Vermeidung von YouTube-Download-Limits<br><br>**⚠️ Wichtige Hinweise**<br>- Teilen Sie Ihre Cookie-Dateien NIEMALS!<br>- Aktualisieren Sie Cookies, wenn:<br>  • \"403 Verboten\"-Fehler auftreten<br>  • Downloads plötzlich stoppen<br>  • \"Sitzung abgelaufen\"-Meldungen erscheinen<br><br>**🔄 Aktualisierungsschritte**<br>1. Installieren Sie diese <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">Chrome-Erweiterung</a><br>2. Melden Sie sich in Chrome bei YouTube an<br>3. Klicken Sie auf das Erweiterungs-Symbol → \"Exportieren\"<br>4. Laden Sie die heruntergeladene Datei hier hoch<br><br>**⏳ Cookie-Lebensdauer**<br>- Normale Sitzungen: 24 Stunden<br>- Sensible Operationen: 1 Stunde<br>- Passwortänderung: Sofortige Aktualisierung</div>",
+  "ensemble_output_copied": "✅ Ensemble-Ausgabe nach {} kopiert",
+  "error_copying_ensemble_output": "❌ Fehler beim Kopieren der Ensemble-Ausgabe: {}",
+  "drive_mounted_copying_ensemble": "Google Drive gemountet. Kopiere Ensemble-Ausgabe...",
+  "drive_already_mounted_copying_ensemble": "Google Drive bereits gemountet. Kopiere Ensemble-Ausgabe...",
+  "no_ensemble_output_files_found": "❌ Keine Ensemble-Ausgabedateien gefunden.",
+  "drive_mounted_copying_files": "Google Drive gemountet. Kopiere Dateien...",
+  "drive_already_mounted_copying_files": "Google Drive bereits gemountet. Kopiere Dateien...",
+  "files_copied_to_drive": "✅ Dateien nach {} kopiert",
+  "error_copying_files": "❌ Fehler beim Kopieren der Dateien: {}",
+  "mounting_drive": "Google Drive wird gemountet...",
+  "settings_saved": "Einstellungen in config.json gespeichert: Sprache={lang}, Methode={method}, Port={port}",
+  "selected_language": "Ausgewählte Sprache: {lang}",
+  "available_languages": "Verfügbare Sprachen: {langs}",
+  "select_language_prompt": "Wählen Sie eine Sprache (Standard: {default}): ",
+  "invalid_language": "Ungültige Sprachauswahl! Verwende Standardsprache ({default}).",
+  "available_sharing_methods": "Verfügbare Freigabemethoden: gradio, localtunnel, ngrok",
+  "select_sharing_method_prompt": "Wählen Sie eine Freigabemethode (Standard: {default}): ",
+  "enter_ngrok_token_prompt": "Geben Sie Ihren Ngrok-Token ein (Standard: {default}): ",
+  "ngrok_token_required": "Fehler: Ngrok-Token für ngrok-Methode erforderlich!",
+  "enter_port_prompt": "Geben Sie die Portnummer ein (Standard: {default}): ",
+  "opening_previous_url": "Öffne vorherige URL: {url}",
+  "open_url_manually": "Öffnen Sie die URL manuell in Colab: {url}",
+  "invalid_previous_url": "Ungültige vorherige URL: {url}, nicht geöffnet.",
+  "starting_method": "Starte {method} auf Port {port}...",
+  "process_stopped": "\n🛑 Prozess vom Benutzer gestoppt",
+  "selected_port": "Ausgewählter Port: {port}",
+  "starting_gradio_with_sharing": "Starte Gradio mit integrierter Freigabe...",
+  "starting_localtunnel": "Starte Localtunnel auf Port {port}...",
+  "share_link": "Freigabelink: {url}",
+  "password_ip": "Passwort-IP: {ip}",
+  "starting_ngrok": "Starte Ngrok auf Port {port}...",
+  "ngrok_url": "Ngrok-URL: {url}",
+  "ngrok_error": "Fehler beim Starten von Ngrok: {error}",
+  "auto_apollo_chunk_size": "Apollo-Chunk-Größe",
+  "auto_apollo_chunk_size_info": "Chunk-Größe für Apollo-Verbesserung (empfohlen: 19 für Universal-Modell)",
+  "auto_apollo_overlap": "Apollo-Überlappung",
+  "auto_enhancing_with_apollo": "Verbesserung mit Apollo ({}/{} Dateien)...",
+  "auto_apollo_processing_completed": "Apollo-Verarbeitung abgeschlossen!",
+  "auto_apollo_overlap_info": "Überlappung für Apollo-Verbesserung (empfohlen: 2)",
+  "apollo_chunk_size": "Apollo-Chunk-Größe",
+  "apollo_chunk_size_info": "Chunk-Größe für Apollo-Verbesserung (empfohlen: 19 für Universal-Modell)",
+  "apollo_overlap": "Apollo-Überlappung",
+  "enhance_with_apollo": "Mit Apollo verbessern",
+  "apollo_enhancement_info": "Audioqualitätsverbesserung",
+  "enhancing_with_apollo": "Verbesserung mit Apollo ({}/{} Dateien)...",
+  "apollo_processing_completed": "Apollo-Verarbeitung abgeschlossen!",
+  "apollo_overlap_info": "Überlappung für Apollo-Verbesserung (empfohlen: 2)",
+  "selected_models": "Ausgewählte Modelle",
+  "save_preset": "Preset speichern",
+  "delete_preset": "Preset löschen",
+  "refresh_presets": "Presets aktualisieren",
+  "preset_name": "Preset-Name",
+  "select_preset": "Preset auswählen",
+  "add_favorite": "Zu Favoriten hinzufügen",
+  "apply_matchering": "Matchering anwenden",
+  "matchering_info": "Matchering anwenden, um Audio-Mastering und -Balance zu verbessern",
+  "matchering_passes": "Matchering-Durchläufe",
+  "matchering_passes_info": "Anzahl der Matchering-Iterationen (1-5, höhere Werte verbessern die Qualität, dauern aber länger)",
+  "matchering_processing": "Matchering wird angewendet... ({:.1f}%)",
+  "matchering_completed": "✅ Matchering-Verarbeitung abgeschlossen!",
+  "matchering_error": "❌ Fehler während Matchering: {}",
+  "invalid_method": "Fehler: Ungültige Methode! Verwenden Sie 'gradio', 'localtunnel' oder 'ngrok'.",
+  "language_changed_message": "Sprache geändert. Bitte starten Sie die Zelle neu.",
+  "backend_settings": "🚀 Backend-Einstellungen",
+  "inference_backend": "Inference-Backend",
+  "optimization_mode": "Optimierungsmodus",
+  "optimization_mode_info": "Auswahl des PyTorch-Optimierungsmodus",
+  "mixed_precision_amp": "Gemischte Präzision (AMP)",
+  "mixed_precision_info": "2x schnellere Inference - empfohlen",
+  "tf32_acceleration": "TF32-Beschleunigung",
+  "tf32_acceleration_info": "Zusätzlicher Geschwindigkeitsboost für RTX 30xx+",
+  "cudnn_benchmark": "cuDNN-Benchmark",
+  "cudnn_benchmark_info": "Langsamer beim ersten Lauf, viel schneller bei nachfolgenden Läufen",
+  "ultra_optimized_pytorch": "ULTRA-OPTIMIERTES PyTorch-Backend",
+  "default_active_max_speed": "Standardmäßig aktiv - Maximale Geschwindigkeitsoptimierung",
+  "channels_last_mode": "Schnellstes für RTX-GPUs",
+  "compile_mode": "Zusätzliche Geschwindigkeit für PyTorch 2.0+",
+    "default_mode": "Standard",
   "manual_ensemble_tab": "🎚️ Manuelles Ensemble",
   "input_sources": "📂 Eingabequellen",
   "refresh": "🔄 Aktualisieren",
@@ -99,23 +178,24 @@
   "Vocal Models": "Gesangsmodelle",
   "Instrumental Models": "Instrumentalmodelle",
   "4-Stem Models": "4-Stem-Modelle",
-  "Denoise Models": "Rauschunterdrückungsmodelle",
-  "Dereverb Models": "Hallentfernungsmodelle",
+  "Denoise Models": "Entrauschungsmodelle",
+  "Dereverb Models": "Nachhallentfernungsmodelle",
   "Other Models": "Andere Modelle",
-  "ensemble_files_description": "Kombiniert mehrere Audiodateien zu einer einzigen Ausgabe unter Verwendung spezifizierter Ensemble-Methoden.",
+  "Other Shapes": "Andere Modelle",
+  "ensemble_files_description": "Kombiniert mehrere Audiodateien zu einer einzigen Ausgabe unter Verwendung angegebener Ensemble-Methoden.",
   "ensemble_files_help": "Pfad zu allen Audiodateien für das Ensemble",
-  "ensemble_type_help": "Einer von avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
-  "ensemble_weights_help": "Gewichte zur Erstellung des Ensembles. Die Anzahl der Gewichte muss der Anzahl der Dateien entsprechen",
+  "ensemble_type_help": "Einer von: avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
+  "ensemble_weights_help": "Gewichte zum Erstellen des Ensembles. Anzahl der Gewichte muss der Anzahl der Dateien entsprechen",
   "ensemble_output_help": "Pfad zur WAV-Datei, in der das Ensemble-Ergebnis gespeichert wird",
   "ensemble_type_print": "Ensemble-Typ: {}",
   "num_input_files_print": "Anzahl der Eingabedateien: {}",
   "weights_print": "Gewichte: {}",
   "output_file_print": "Ausgabedatei: {}",
   "duration_mismatch_error": "Alle Dateien müssen die gleiche Dauer haben",
-  "file_not_found_error": "Fehler. Datei nicht gefunden: {}. Überprüfe die Pfade.",
+  "file_not_found_error": "Fehler. Datei nicht gefunden: {}. Überprüfen Sie die Pfade.",
   "reading_chunk_print": "Lese Chunk aus Datei: {} (Start: {}s, Dauer: {}s)",
   "chunk_result_shape_print": "Chunk-Ergebnisform: {}",
-  "ensemble_completed_print": "Ensemble abgeschlossen. Ausgabe gespeichert unter: {}",
+  "ensemble_completed_print": "Ensemble abgeschlossen. Ausgabe gespeichert in: {}",
   "file_deletion_error": "{} konnte nicht gelöscht werden: {}",
   "directory_not_exist_warning": "⚠️ Verzeichnis existiert nicht: {}",
   "not_a_directory_warning": "⚠️ Pfad ist kein Verzeichnis: {}",
@@ -125,42 +205,50 @@
   "error": "🔥 Fehler: {}",
   "ffmpeg_error": "FFmpeg-Fehler ({}): {}",
   "file_saved_successfully": "Datei erfolgreich gespeichert: {}",
-  "total_files_found": "Gesamtzahl gefundener Dateien: {}. Verwende Abtastrate: {}",
+  "total_files_found": "Gesamtzahl gefundener Dateien: {}. Verwendete Abtastrate: {}",
   "total_progress": "Gesamtfortschritt",
-  "detailed_pbar_enabled": "Detaillierter Fortschrittsbalken aktiviert: {}",
-  "loaded_audio": "Geladenes Audio: {}, Form: {}",
-  "cannot_read_track": "Spur kann nicht gelesen werden: {}",
+  "detailed_pbar_enabled": "Detaillierte Fortschrittsanzeige aktiviert: {}",
+  "loaded_audio": "Audio geladen: {}, Form: {}",
+  "cannot_read_track": "Track kann nicht gelesen werden: {}",
   "error_message": "Fehlermeldung: {}",
-  "demudding_track": "Spur entmischen (Phasen-Remix - Instrumental): {}",
+  "demudding_track": "Track wird bereinigt (Phasen-Remix - Instrumental): {}",
   "elapsed_time": "Verstrichene Zeit: {:.2f} Sekunden.",
-  "proc_folder_description": "Verarbeitet Audiodateien in einem Ordner mit einem spezifizierten Modell.",
+  "proc_folder_description": "Verarbeitet Audiodateien in einem Ordner mit einem angegebenen Modell.",
   "model_type_help": "Modelltyp (bandit, bs_roformer, mdx23c, etc.)",
   "config_path_help": "Pfad zur Konfigurationsdatei",
   "demud_phaseremix_help": "Phasen-Remix für Instrumental aktivieren",
-  "start_checkpoint_help": "Initialer Checkpoint für gültige Gewichte",
-  "input_folder_help": "Ordner mit Mischungen zur Verarbeitung",
-  "audio_path_help": "Pfad zu einer einzelnen Audiodatei zur Verarbeitung",
+  "start_checkpoint_help": "Anfangs-Checkpoint für gültige Gewichte",
+  "input_folder_help": "Ordner mit zu verarbeitenden Mischungen",
+  "audio_path_help": "Pfad zu einer einzelnen zu verarbeitenden Audiodatei",
   "store_dir_help": "Pfad zum Speichern der Ergebnisse",
   "device_ids_help": "Liste der GPU-IDs",
   "extract_instrumental_help": "Gesang invertieren, um Instrumental zu erhalten, falls vorhanden",
-  "disable_detailed_pbar_help": "Detaillierten Fortschrittsbalken in Demix deaktivieren",
-  "force_cpu_help": "CPU erzwingen, selbst wenn CUDA verfügbar ist",
-  "flac_file_help": "FLAC-Datei anstelle von WAV ausgeben",
+  "disable_detailed_pbar_help": "Detaillierte Fortschrittsanzeige beim Demix deaktivieren",
+  "force_cpu_help": "CPU-Nutzung erzwingen, auch wenn CUDA verfügbar ist",
+  "flac_file_help": "FLAC-Datei statt WAV ausgeben",
   "export_format_help": "Exportformat und PCM-Typ",
   "pcm_type_help": "PCM-Typ für FLAC-Dateien",
-  "use_tta_help": "Testzeit-Augmentation aktivieren",
-  "lora_checkpoint_help": "Initialer Checkpoint für LoRA-Gewichte",
-  "cuda_available": "CUDA ist verfügbar, verwende --force_cpu, um es zu deaktivieren.",
+  "use_tta_help": "Test-Time-Augmentation aktivieren",
+  "lora_checkpoint_help": "Anfangs-Checkpoint für LoRA-Gewichte",
+  "cuda_available": "CUDA ist verfügbar, verwenden Sie --force_cpu zum Deaktivieren.",
   "using_device": "Verwendetes Gerät: {}",
   "instruments_print": "Instrumente: {}",
   "model_load_time": "Modell-Ladezeit: {:.2f} Sek.",
   "invalid_url": "❌ Ungültige URL",
   "cookie_file_updated": "✅ Cookie-Datei aktualisiert!",
-  "cookie_installation_error": "⚠️ Fehler bei der Cookie-Installation: {}",
-  "file_size_zero_error": "Dateigröße null oder Datei nicht erstellt",
-  "google_drive_error": "❌ Google Drive Download-Fehler: {}",
+  "cookie_installation_error": "⚠️ Cookie-Installationsfehler: {}",
+  "google_drive_mount_error": "❌ Google Drive-Verbindungsfehler: {}",
+  "file_size_zero_error": "Dateigröße Null oder Datei nicht erstellt",
+  "google_drive_error": "❌ Google Drive-Download-Fehler: {}",
   "wav_conversion_failed": "WAV-Konvertierung fehlgeschlagen",
   "download_error": "❌ Download-Fehler: {}",
+  "file_copied_to_drive": "Datei nach Google Drive kopiert: {}",
+  "copy_to_drive_error": "❌ Fehler beim Kopieren nach Google Drive: {}",
+  "direct_download_failed": "Direkter Download der Datei fehlgeschlagen",
+  "direct_download_error": "❌ Direkter Download-Fehler: {}",
+  "invalid_google_drive_url": "❌ Ungültige Google Drive-URL",
+  "continuing_without_google_drive": "Google Drive-Verbindung fehlgeschlagen, fahre ohne fort...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive nicht gemountet, überspringe Drive-Kopie...",
   "download_success": "🎉 Erfolgreich heruntergeladen!",
   "download_failed": "❌ Download fehlgeschlagen",
   "no_output_files_found": "❌ Keine Ausgabedateien gefunden.",
@@ -170,19 +258,19 @@
   "processing_audio": "Audio wird verarbeitet",
   "separating_audio": "Audio wird getrennt... ({:.1f}%)",
   "separation_complete": "Trennung abgeschlossen!",
-  "progress_parsing_error": "Fortschrittsanalyse-Fehler: {}",
+  "progress_parsing_error": "Fortschritts-Parsing-Fehler: {}",
   "error_occurred": "Ein Fehler ist aufgetreten: {}",
   "separation_process_completed": "Trennungsprozess abgeschlossen!",
-  "no_audio_file_error": "❌ Keine Audiodatei bereitgestellt und keine vorhandene Datei im Eingabeordner.",
+  "no_audio_file_error": "❌ Keine Audiodatei bereitgestellt und keine vorhandene Datei im Eingabeverzeichnis.",
   "no_input_progress_label": "Fehler: Keine Eingabe bereitgestellt -- 0.0%",
   "processing_audio_print": "Verarbeite Audio von: {} mit Modell: {}",
   "starting_audio_separation_progress_label": "Audio-Trennung wird gestartet... -- 0.0%",
   "separating_audio_progress_label": "Audio wird getrennt... -- {}.0%",
-  "audio_processing_completed_progress_label": "Audioverarbeitung abgeschlossen! -- 100.0%",
-  "audio_processing_completed": "✅ Audioverarbeitung abgeschlossen!",
+  "audio_processing_completed_progress_label": "Audio-Verarbeitung abgeschlossen! -- 100.0%",
+  "audio_processing_completed": "✅ Audio-Verarbeitung abgeschlossen!",
   "error_occurred_progress_label": "Fehler aufgetreten -- 0.0%",
   "minimum_files_required": "⚠️ Mindestens 2 Dateien erforderlich",
-  "valid_files_not_found": "❌ Gültige Dateien nicht gefunden",
+  "valid_files_not_found": "❌ Keine gültigen Dateien gefunden",
   "starting_ensemble_process": "Ensemble-Prozess wird gestartet...",
   "ensembling_progress": "Ensemble wird erstellt... ({:.1f}%)",
   "finalizing_ensemble_output": "Ensemble-Ausgabe wird finalisiert...",
@@ -194,87 +282,65 @@
   "no_input_audio_provided": "❌ Kein Eingabe-Audio bereitgestellt",
   "loading_model": "Lade Modell {}/{}: {}...",
   "loading_model_progress_label": "Lade Modell {}/{}: {}... -- {}.0%",
-  "completed_model": "Abgeschlossenes Modell {}/{}: {}",
-  "completed_model_progress_label": "Abgeschlossenes Modell {}/{}: {} -- {}.0%",
+  "completed_model": "Modell abgeschlossen {}/{}: {}",
+  "completed_model_progress_label": "Modell abgeschlossen {}/{}: {} -- {}.0%",
   "running_command": "Führe Befehl aus: {}",
   "model_failed": "Modell {} fehlgeschlagen: {}",
   "critical_error_with_model": "Kritischer Fehler mit {}: {}",
   "model_output_failed": "{} konnte keine Ausgabe erzeugen",
-  "waiting_for_files": "Warten auf die Bereitschaft aller Dateien...",
-  "waiting_for_files_progress_label": "Warten auf die Bereitschaft aller Dateien... -- 90.0%",
+  "waiting_for_files": "Warten auf Bereitstellung aller Dateien...",
+  "waiting_for_files_progress_label": "Warten auf Bereitstellung aller Dateien... -- 90.0%",
   "performing_ensemble": "Ensemble wird durchgeführt...",
   "performing_ensemble_progress_label": "Ensemble wird durchgeführt... -- 92.0%",
-  "memory_usage_before_ensemble": "Speicherauslastung vor Ensemble: {}%",
-  "memory_usage_after_ensemble": "Speicherauslastung nach Ensemble: {}%",
+  "memory_usage_before_ensemble": "Speichernutzung vor Ensemble: {}%",
+  "memory_usage_after_ensemble": "Speichernutzung nach Ensemble: {}%",
   "finalizing_ensemble_output_progress_label": "Ensemble-Ausgabe wird finalisiert... -- 98.0%",
   "ensemble_file_creation_failed": "Ensemble-Datei konnte nicht erstellt werden: {}",
   "ensemble_completed_progress_label": "Ensemble erfolgreich abgeschlossen! -- 100.0%",
   "success_output_created": "✅ Erfolg! Ausgabedatei erstellt.",
-  "drive_mounted_copying_ensemble": "Google Drive eingebunden. Ensemble-Ausgabe wird kopiert...",
-  "drive_already_mounted_copying_ensemble": "Google Drive bereits eingebunden. Ensemble-Ausgabe wird kopiert...",
-  "no_ensemble_output_files_found": "❌ Keine Ensemble-Ausgabedateien gefunden.",
-  "ensemble_output_copied": "✅ Ensemble-Ausgabe kopiert nach {}",
-  "error_copying_ensemble_output": "❌ Fehler beim Kopieren der Ensemble-Ausgabe: {}",
-  "drive_mounted_copying_files": "Google Drive eingebunden. Dateien werden kopiert...",
-  "drive_already_mounted_copying_files": "Google Drive bereits eingebunden. Dateien werden kopiert...",
-  "files_copied_to_drive": "✅ Dateien kopiert nach {}",
-  "error_copying_files": "❌ Fehler beim Kopieren der Dateien: {}",
-  "mounting_drive": "Google Drive wird eingebunden...",
-  "settings_saved": "Einstellungen in config.json gespeichert: Sprache={lang}, Methode={method}, Port={port}",
-  "selected_language": "Ausgewählte Sprache: {lang}",
-  "available_languages": "Verfügbare Sprachen: {langs}",
-  "select_language_prompt": "Wähle eine Sprache (Standard: {default}): ",
-  "invalid_language": "Ungültige Sprachauswahl! Verwende Standardsprache ({default}).",
-  "available_sharing_methods": "Verfügbare Freigabemethoden: gradio, localtunnel, ngrok",
-  "select_sharing_method_prompt": "Wähle eine Freigabemethode (Standard: {default}): ",
-  "enter_ngrok_token_prompt": "Gib deinen Ngrok-Token ein (Standard: {default}): ",
-  "ngrok_token_required": "Fehler: Ngrok-Token ist für die ngrok-Methode erforderlich!",
-  "enter_port_prompt": "Gib die Portnummer ein (Standard: {default}): ",
-  "opening_previous_url": "Öffne vorherige URL: {url}",
-  "open_url_manually": "Öffne die URL manuell in Colab: {url}",
-  "invalid_previous_url": "Ungültige vorherige URL: {url}, nicht geöffnet.",
-  "starting_method": "Starte {method} auf Port {port}...",
-  "process_stopped": "\n🛑 Prozess vom Benutzer gestoppt",
-  "selected_port": "Ausgewählter Port: {port}",
-  "starting_gradio_with_sharing": "Starte Gradio mit integrierter Freigabe...",
-  "starting_localtunnel": "Starte Localtunnel auf Port {port}...",
-  "share_link": "Freigabe-Link: {url}",
-  "password_ip": "Passwort-IP: {ip}",
-  "starting_ngrok": "Starte Ngrok auf Port {port}...",
-  "ngrok_url": "Ngrok-URL: {url}",
-  "ngrok_error": "Fehler beim Starten von Ngrok: {error}",
-  "apollo_chunk_size": "Apollo Chunk-Größe",
-  "apollo_chunk_size_info": "Chunk-Größe für Apollo-Verbesserung (empfohlen: 19 für Universalmodell)",
-  "apollo_overlap": "Apollo-Überlappung",
-  "enhancing_with_apollo": "Verbessere mit Apollo ({}/{} Dateien)...",
-  "apollo_processing_completed": "Apollo-Verarbeitung abgeschlossen!",
-  "apollo_overlap_info": "Überlappung für Apollo-Verbesserung (empfohlen: 2)",
-  "invalid_method": "Fehler: Ungültige Methode! Verwende 'gradio', 'localtunnel' oder 'ngrok'.",
-  "apollo_enhancement_settings": "Apollo-Verbesserungseinstellungen",
-  "enhance_with_apollo": "Mit Apollo verbessern",
-  "enhance_with_apollo_info": "Apollo für Audioverbesserung nach der Trennung aktivieren",
-  "apollo_method": "Apollo-Methode",
-  "apollo_method_info": "Wähle die Verarbeitungsmethode für Apollo",
-  "normal_method": "Normale Methode",
-  "mid_side_method": "Mitte/Seite-Methode",
-  "apollo_normal_model": "Apollo Normalmodell",
-  "apollo_normal_model_info": "Modell für normale Apollo-Verarbeitung",
-  "apollo_midside_model": "Apollo Mitte/Seite-Modell",
-  "apollo_enhancement_info": "Verbesserung der Klangqualität",
-  "selected_models": "Ausgewählte Modelle",
-  "save_preset": "Voreinstellung Speichern",
-  "delete_preset": "Voreinstellung Löschen",
-  "refresh_presets": "Voreinstellungen Aktualisieren",
-  "preset_name": "Name der Voreinstellung",
-  "select_preset": "Voreinstellung Auswählen",
-  "add_favorite": "Zu Favoriten Hinzufügen",
-  "apollo_midside_model_info": "Modell für Mitte/Seite-Verarbeitung (optional)",
-  "language_changed_message": "Sprache geändert. Bitte starte die Zelle neu.",
-  "apply_matchering": "Matchering anwenden",
-  "matchering_info": "Matchering anwenden, um Audio-Mastering und -Balance zu verbessern",
-  "matchering_passes": "Matchering-Durchgänge",
-  "matchering_passes_info": "Anzahl der Matchering-Iterationen (1-5, höhere Werte verbessern die Qualität, dauern aber länger)",
-  "matchering_processing": "Matchering wird angewendet... ({:.1f}%)",
-  "matchering_completed": "✅ Matchering-Verarbeitung abgeschlossen!",
-  "matchering_error": "❌ Fehler während Matchering: {}"
-}
+  "phase_fixer_tab": "🔧 Phasenkorrektor",
+  "batch_processing_tab": "📦 Stapelverarbeitung",
+  "batch_description": "Mehrere Dateien gleichzeitig mit demselben Modell verarbeiten",
+  "batch_input_folder": "Eingabeordner",
+  "batch_input_folder_placeholder": "Pfad zum Ordner mit Audiodateien eingeben",
+  "batch_output_folder": "Ausgabeordner",
+  "batch_output_folder_placeholder": "Pfad zum Ausgabeordner eingeben",
+  "batch_file_list": "Zu verarbeitende Dateien",
+  "batch_add_files": "Dateien hinzufügen",
+  "batch_clear_list": "Liste leeren",
+  "batch_start": "🚀 Stapelverarbeitung starten",
+  "batch_stop": "⏹️ Stopp",
+  "batch_progress": "Stapelverarbeitungsfortschritt",
+  "batch_current_file": "Aktuell in Bearbeitung",
+  "batch_completed": "✅ Stapelverarbeitung abgeschlossen!",
+  "batch_stopped": "⏹️ Stapelverarbeitung gestoppt",
+  "batch_no_files": "❌ Keine Dateien für die Verarbeitung ausgewählt",
+  "source_model": "Quellmodell (Gesang)",
+  "source_model_info": "Modell zur Extraktion von Gesang/Phasendaten",
+  "target_model": "Zielmodell (Instrumental)",
+  "target_model_info": "Modell zur Anwendung der Phasenkorrektur",
+  "custom_models_tab": "⚡ Benutzerdefinierte Modelle",
+  "Custom Models": "Benutzerdefinierte Modelle",
+  "add_custom_model": "➕ Benutzerdefiniertes Modell hinzufügen",
+  "custom_model_name": "Modellname",
+  "custom_model_name_placeholder": "Geben Sie einen Namen für Ihr Modell ein",
+  "checkpoint_url": "Checkpoint-URL (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "Konfigurations-URL (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "Benutzerdefinierte Python-Datei URL (optional)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "Modelltyp automatisch erkennen",
+  "model_type": "Modelltyp",
+  "add_model_btn": "➕ Modell hinzufügen",
+  "custom_models_list": "📋 Liste benutzerdefinierter Modelle",
+  "delete_model": "🗑️ Löschen",
+  "no_custom_models": "Noch keine benutzerdefinierten Modelle hinzugefügt",
+  "model_added_success": "✅ Modell erfolgreich hinzugefügt!",
+  "model_deleted_success": "✅ Modell erfolgreich gelöscht!",
+  "model_add_error": "❌ Fehler beim Hinzufügen des Modells: {}",
+  "model_delete_error": "❌ Fehler beim Löschen des Modells: {}",
+  "refresh_models": "🔄 Aktualisieren",
+  "custom_model_info": "Fügen Sie benutzerdefinierte Modelle durch Angabe von Download-Links hinzu. Modelle werden bei Verwendung automatisch heruntergeladen.",
+  "select_model_to_delete": "Wählen Sie ein Modell zum Löschen aus",
+  "downloading_model_file": "Modell wird heruntergeladen: {}",

assets/i18n/languages/en_us.json

@@ -81,6 +81,27 @@
   "upload_cookies_txt": "Upload Cookies.txt",
   "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 Why Needed?**<br>- Access age-restricted content<br>- Download private/unlisted videos<br>- Bypass regional restrictions<br>- Avoid YouTube download limits<br><br>**⚠️ Important Notes**<br>- NEVER share your cookie files!<br>- Refresh cookies when:<br>  • Getting \"403 Forbidden\" errors<br>  • Downloads suddenly stop<br>  • Seeing \"Session expired\" messages<br><br>**🔄 Renewal Steps**<br>1. Install this <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">Chrome extension</a><br>2. Login to YouTube in Chrome<br>3. Click extension icon → \"Export\"<br>4. Upload the downloaded file here<br><br>**⏳ Cookie Lifespan**<br>- Normal sessions: 24 hours<br>- Sensitive operations: 1 hour<br>- Password changes: Immediate invalidation</div>",
   "manual_ensemble_tab": "🎚️ Manual Ensemble",
+  "phase_fixer_tab": "🔧 Phase Fixer",
+  "batch_processing_tab": "📦 Batch Processing",
+  "batch_description": "Process multiple files at once with the same model",
+  "batch_input_folder": "Input Folder",
+  "batch_input_folder_placeholder": "Enter folder path containing audio files",
+  "batch_output_folder": "Output Folder",
+  "batch_output_folder_placeholder": "Enter folder path for output files",
+  "batch_file_list": "Files to Process",
+  "batch_add_files": "Add Files",
+  "batch_clear_list": "Clear List",
+  "batch_start": "🚀 Start Batch Processing",
+  "batch_stop": "⏹️ Stop",
+  "batch_progress": "Batch Progress",
+  "batch_current_file": "Currently Processing",
+  "batch_completed": "✅ Batch processing completed!",
+  "batch_stopped": "⏹️ Batch processing stopped",
+  "batch_no_files": "❌ No files selected for processing",
+  "source_model": "Source Model (Vocals)",
+  "source_model_info": "Model for extracting vocals/phase data",
+  "target_model": "Target Model (Instrumental)",
+  "target_model_info": "Model to apply phase correction",
   "input_sources": "📂 Input Sources",
   "refresh": "🔄 Refresh",
   "ensemble_algorithm": "Ensemble Algorithm",
@@ -165,6 +186,10 @@
   "file_copied_to_drive": "File copied to Google Drive: {}",
   "copy_to_drive_error": "❌ Error copying to Google Drive: {}",
   "direct_download_failed": "Failed to download the file directly",
+  "direct_download_error": "❌ Direct download error: {}",
+  "invalid_google_drive_url": "❌ Invalid Google Drive URL",
+  "continuing_without_google_drive": "Google Drive connection failed, continuing without it...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive not mounted, skipping Drive copy...",
   "download_success": "🎉 Downloaded successfully!",
   "download_failed": "❌ Download failed",
   "no_output_files_found": "❌ No output files found.",
@@ -275,5 +300,79 @@
   "matchering_completed": "✅ Matchering processing completed!",
   "matchering_error": "❌ Error during Matchering: {}",
   "invalid_method": "Error: Invalid method! Use 'gradio', 'localtunnel', or 'ngrok'.",
-  "language_changed_message": "Language changed. please restart the cell."
+  "language_changed_message": "Language changed. please restart the cell.",
+  "backend_settings": "🚀 Backend Settings",
+  "inference_backend": "Inference Backend",
+  "optimization_mode": "Optimization Mode",
+  "optimization_mode_info": "PyTorch optimization mode selection",
+  "mixed_precision_amp": "Mixed Precision (AMP)",
+  "mixed_precision_info": "2x faster inference - recommended",
+  "tf32_acceleration": "TF32 Acceleration",
+  "tf32_acceleration_info": "Extra speed boost for RTX 30xx+",
+  "cudnn_benchmark": "cuDNN Benchmark",
+  "cudnn_benchmark_info": "Slower at first run, much faster on subsequent runs",
+  "ultra_optimized_pytorch": "ULTRA-OPTIMIZED PyTorch Backend",
+  "default_active_max_speed": "Active by default - Maximum speed optimization",
+  "channels_last_mode": "Fastest for RTX GPUs",
+  "compile_mode": "Extra speed for PyTorch 2.0+",
+  "default_mode": "Standard",
+  "tta_info": "Increases quality but slows down processing",
+  "instrumental_info": "Usually outputs 2 stems, but sometimes only one - you need to try! :)",
+  "custom_models_tab": "⚡ Custom Models",
+  "Custom Models": "Custom Models",
+  "add_custom_model": "➕ Add Custom Model",
+  "custom_model_name": "Model Name",
+  "custom_model_name_placeholder": "Enter a name for your model",
+  "checkpoint_url": "Checkpoint URL (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "Config URL (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "Custom Python File URL (optional)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "Auto-detect model type",
+  "model_type": "Model Type",
+  "add_model_btn": "➕ Add Model",
+  "custom_models_list": "📋 Custom Models List",
+  "delete_model": "🗑️ Delete",
+  "no_custom_models": "No custom models added yet",
+  "model_added_success": "✅ Model added successfully!",
+  "model_deleted_success": "✅ Model deleted successfully!",
+  "model_add_error": "❌ Error adding model: {}",
+  "model_delete_error": "❌ Error deleting model: {}",
+  "refresh_models": "🔄 Refresh",
+  "custom_model_info": "Add custom models by providing download links. Models will be downloaded automatically when used.",
+  "select_model_to_delete": "Select a model to delete",
+  "downloading_model_file": "Downloading model: {}",
+  "downloading_file_progress": "Downloading: {} - {}%",
+  "chunk_size_mode": "Chunk Size Mode",
+  "chunk_size_mode_info": "base: standard presets | custom: enter your own value | yaml: read from model YAML after download",
+  "chunk_size_custom_label": "Custom Chunk Size",
+  "chunk_size_custom_info": "Enter a custom chunk size value",
+  "chunk_size_yaml_label": "YAML Chunk Size",
+  "chunk_size_yaml_display_info": "Will be read from the model's YAML config after download",
+  "chunk_size_yaml_not_downloaded": "⏳ Model YAML not downloaded yet — value will be read during processing",
+  "chunk_size_yaml_detected": "✅ From YAML: {}",
+  "remove_favorite": "Remove from Favorites",
+  "normal_method": "Normal method",
+  "mid_side_method": "Mid-side method",
+  "apollo_processing_method": "Apollo Processing Method",
+  "apollo_normal_model": "Apollo Normal Model",
+  "apollo_mid_side_model": "Apollo Mid-Side Model",
+  "scale_factor": "Scale Factor",
+  "scale_factor_info": "Phase scaling factor (recommended: 1.4)",
+  "source_file_label": "Source File (Vocals)",
+  "target_file_label": "Target File (Instrumental)",
+  "run_phase_fixer": "🔧 Run Phase Fixer",
+  "phase_fixed_output": "Phase-Fixed Output",
+  "phase_fixer_settings": "Phase Fixer Settings",
+  "low_cutoff": "Low Cutoff (Hz)",
+  "high_cutoff": "High Cutoff (Hz)",
+  "low_cutoff_info": "Low frequency cutoff for phase correction",
+  "high_cutoff_info": "High frequency cutoff for phase correction",
+  "no_preset_name_provided": "⚠️ No preset name provided",
+  "no_models_selected_for_preset": "⚠️ No models selected for preset",
+  "preset_saved": "✅ Preset '{}' saved!",
+  "please_upload_both_files": "⚠️ Please upload both source and target files",
+  "processing_log_placeholder": "Processing log will appear here...",
+  "enter_preset_name": "Enter preset name..."
 }

assets/i18n/languages/es_es.json

@@ -17,11 +17,10 @@
   "overlap": "Solapamiento",
   "overlap_info": "Recomendado: 2-10 (valores más altos mejoran la calidad pero requieren más VRAM)",
   "tta_boost": "Impulso TTA",
-  "tta_info": "Mejora la calidad pero ralentiza la velocidad de procesamiento",
   "phase_fix": "Corrección de fase",
+  "use_phase_correction": "Aplicar Corrección de Fase Avanzada",
   "phase_fix_info": "Corrección avanzada de fase para pistas instrumentales",
   "instrumental": "Instrumental",
-  "instrumental_info": "Normalmente da 2 salidas, pero a veces da una sola salida, ¡tienes que probar! :)",
   "process": "🚀 Procesar",
   "reset": "🧹 Restablecer",
   "status": "Estado",
@@ -80,7 +79,87 @@
   "downloaded_file": "Archivo descargado",
   "cookie_management": "🍪 Gestión de cookies",
   "upload_cookies_txt": "Subir Cookies.txt",
-  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 ¿Por qué es necesario?**<br>- Acceso a contenido restringido por edad<br>- Descarga de videos privados/no listados<br>- Evitar restricciones regionales<br>- Evitar límites de descarga de YouTube<br><br>**⚠️ Notas importantes**<br>- ¡NUNCA compartas tus archivos de cookies!<br>- Actualiza las cookies cuando:<br>  • Recibes errores \"403 Prohibido\"<br>  • Las descargas se detienen repentinamente<br>  • Ves mensajes de \"Sesión expirada\"<br><br>**🔄 Pasos para renovar**<br>1. Instala esta <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">extensión de Chrome</a><br>2. Inicia sesión en YouTube en Chrome<br>3. Haz clic en el ícono de la extensión → \"Exportar\"<br>4. Sube el archivo descargado aquí<br><br>**⏳ Vida útil de las cookies**<br>- Sesiones normales: 24 horas<br>- Operaciones sensibles: 1 hora<br>- Cambios de contraseña: Invalidación inmediata</div>",
+  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 ¿Por qué es necesario?**<br>- Acceso a contenido con restricción de edad<br>- Descarga de videos privados/no listados<br>- Evitar restricciones regionales<br>- Evitar límites de descarga de YouTube<br><br>**⚠️ Notas Importantes**<br>- ¡NUNCA compartas tus archivos de cookies!<br>- Actualiza cookies cuando:<br>  • Obtengas errores \"403 Prohibido\"<br>  • Las descargas se detengan de repente<br>  • Veas mensajes \"Sesión expirada\"<br><br>**🔄 Pasos de Renovación**<br>1. Instala esta <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">extensión de Chrome</a><br>2. Inicia sesión en YouTube en Chrome<br>3. Haz clic en el icono de la extensión → \"Exportar\"<br>4. Sube el archivo descargado aquí<br><br>**⏳ Vida Útil de Cookies**<br>- Sesiones normales: 24 horas<br>- Operaciones sensibles: 1 hora<br>- Cambio de contraseña: Actualización inmediata</div>",
+  "ensemble_output_copied": "✅ Salida del ensamblaje copiada a {}",
+  "error_copying_ensemble_output": "❌ Error al copiar salida del ensamblaje: {}",
+  "drive_mounted_copying_ensemble": "Google Drive montado. Copiando salida del ensamblaje...",
+  "drive_already_mounted_copying_ensemble": "Google Drive ya montado. Copiando salida del ensamblaje...",
+  "no_ensemble_output_files_found": "❌ No se encontraron archivos de salida del ensamblaje.",
+  "drive_mounted_copying_files": "Google Drive montado. Copiando archivos...",
+  "drive_already_mounted_copying_files": "Google Drive ya montado. Copiando archivos...",
+  "files_copied_to_drive": "✅ Archivos copiados a {}",
+  "error_copying_files": "❌ Error al copiar archivos: {}",
+  "mounting_drive": "Montando Google Drive...",
+  "settings_saved": "Configuración guardada en config.json: Idioma={lang}, Método={method}, Puerto={port}",
+  "selected_language": "Idioma seleccionado: {lang}",
+  "available_languages": "Idiomas disponibles: {langs}",
+  "select_language_prompt": "Selecciona un idioma (predeterminado: {default}): ",
+  "invalid_language": "¡Selección de idioma inválida! Usando idioma predeterminado ({default}).",
+  "available_sharing_methods": "Métodos de compartición disponibles: gradio, localtunnel, ngrok",
+  "select_sharing_method_prompt": "Elige un método de compartición (predeterminado: {default}): ",
+  "enter_ngrok_token_prompt": "Ingresa tu token de Ngrok (predeterminado: {default}): ",
+  "ngrok_token_required": "Error: ¡Se requiere un token de Ngrok para el método ngrok!",
+  "enter_port_prompt": "Ingresa el número de puerto (predeterminado: {default}): ",
+  "opening_previous_url": "Abriendo URL anterior: {url}",
+  "open_url_manually": "Abre la URL manualmente en Colab: {url}",
+  "invalid_previous_url": "URL anterior inválida: {url}, no abierta.",
+  "starting_method": "Iniciando {method} en el puerto {port}...",
+  "process_stopped": "\n🛑 Proceso detenido por el usuario",
+  "selected_port": "Puerto seleccionado: {port}",
+  "starting_gradio_with_sharing": "Iniciando Gradio con compartición integrada...",
+  "starting_localtunnel": "Iniciando Localtunnel en el puerto {port}...",
+  "share_link": "Enlace para compartir: {url}",
+  "password_ip": "Contraseña IP: {ip}",
+  "starting_ngrok": "Iniciando Ngrok en el puerto {port}...",
+  "ngrok_url": "URL de Ngrok: {url}",
+  "ngrok_error": "Error al iniciar Ngrok: {error}",
+  "auto_apollo_chunk_size": "Tamaño de fragmento Apollo",
+  "auto_apollo_chunk_size_info": "Tamaño de fragmento para mejora Apollo (recomendado: 19 para modelo universal)",
+  "auto_apollo_overlap": "Solapamiento Apollo",
+  "auto_enhancing_with_apollo": "Mejorando con Apollo ({}/{} archivos)...",
+  "auto_apollo_processing_completed": "¡Procesamiento Apollo completado!",
+  "auto_apollo_overlap_info": "Solapamiento para mejora Apollo (recomendado: 2)",
+  "apollo_chunk_size": "Tamaño de fragmento Apollo",
+  "apollo_chunk_size_info": "Tamaño de fragmento para mejora Apollo (recomendado: 19 para modelo universal)",
+  "apollo_overlap": "Solapamiento Apollo",
+  "enhance_with_apollo": "Mejorar con Apollo",
+  "apollo_enhancement_info": "Mejora de la calidad del sonido",
+  "enhancing_with_apollo": "Mejorando con Apollo ({}/{} archivos)...",
+  "apollo_processing_completed": "¡Procesamiento Apollo completado!",
+  "apollo_overlap_info": "Solapamiento para mejora Apollo (recomendado: 2)",
+  "selected_models": "Modelos seleccionados",
+  "save_preset": "Guardar preajuste",
+  "delete_preset": "Eliminar preajuste",
+  "refresh_presets": "Actualizar preajustes",
+  "preset_name": "Nombre del preajuste",
+  "select_preset": "Seleccionar preajuste",
+  "add_favorite": "Añadir a favoritos",
+  "apply_matchering": "Aplicar Matchering",
+  "matchering_info": "Aplicar Matchering para mejorar el masterizado y el balance del audio",
+  "matchering_passes": "Pasadas de Matchering",
+  "matchering_passes_info": "Número de iteraciones de Matchering (1-5, valores más altos mejoran la calidad pero toman más tiempo)",
+  "matchering_processing": "Aplicando Matchering... ({:.1f}%)",
+  "matchering_completed": "✅ ¡Procesamiento de Matchering completado!",
+  "matchering_error": "❌ Error durante Matchering: {}",
+  "invalid_method": "Error: ¡Método inválido! Usa 'gradio', 'localtunnel' o 'ngrok'.",
+  "language_changed_message": "Idioma cambiado. Por favor, reinicia la celda.",
+  "backend_settings": "🚀 Configuración de backend",
+  "inference_backend": "Backend de inferencia",
+  "optimization_mode": "Modo de optimización",
+  "optimization_mode_info": "Selección de modo de optimización PyTorch",
+  "mixed_precision_amp": "Precisión mixta (AMP)",
+  "mixed_precision_info": "Inferencia 2x más rápida - recomendado",
+  "tf32_acceleration": "Aceleración TF32",
+  "tf32_acceleration_info": "Aumento de velocidad extra para RTX 30xx+",
+  "cudnn_benchmark": "Benchmark cuDNN",
+  "cudnn_benchmark_info": "Más lento en la primera ejecución, mucho más rápido en ejecuciones posteriores",
+  "ultra_optimized_pytorch": "Backend PyTorch ULTRA-OPTIMIZADO",
+  "default_active_max_speed": "Activo por defecto - Optimización de velocidad máxima",
+  "channels_last_mode": "El más rápido para GPUs RTX",
+  "compile_mode": "Velocidad extra para PyTorch 2.0+",
+  "default_mode": "Estándar",
+  "tta_info": "Aumenta la calidad pero ralentiza el procesamiento",
+    "instrumental_info": "Normalmente produce 2 tallos, pero a veces solo uno - ¡tienes que probar! :)",
   "manual_ensemble_tab": "🎚️ Ensamblaje manual",
   "input_sources": "📂 Fuentes de entrada",
   "refresh": "🔄 Actualizar",
@@ -93,19 +172,20 @@
   "result_preview_tab": "🎧 Vista previa del resultado",
   "ensembled_output": "Salida ensamblada",
   "processing_log_tab": "📋 Registro de procesamiento",
-  "processing_details": "Detalles de procesamiento",
+  "processing_details": "Detalles del procesamiento",
   "process_ensemble": "⚡ Procesar ensamblaje",
   "languages_tab": "🌐 Idiomas",
   "Vocal Models": "Modelos vocales",
   "Instrumental Models": "Modelos instrumentales",
-  "4-Stem Models": "Modelos de 4 stems",
+  "4-Stem Models": "Modelos de 4 tallos",
   "Denoise Models": "Modelos de eliminación de ruido",
   "Dereverb Models": "Modelos de eliminación de reverberación",
   "Other Models": "Otros modelos",
-  "ensemble_files_description": "Combina varios archivos de audio en una sola salida usando métodos de ensamblaje especificados.",
-  "ensemble_files_help": "Ruta a todos los archivos de audio para el ensamblaje",
-  "ensemble_type_help": "Uno de avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
-  "ensemble_weights_help": "Pesos para crear el ensamblaje. El número de pesos debe igualar el número de archivos",
+  "Other Shapes": "Otros modelos",
+  "ensemble_files_description": "Combina múltiples archivos de audio en una sola salida usando métodos de ensamblaje especificados.",
+  "ensemble_files_help": "Ruta a todos los archivos de audio para ensamblaje",
+  "ensemble_type_help": "Uno de: avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
+  "ensemble_weights_help": "Pesos para crear el ensamblaje. El número de pesos debe ser igual al número de archivos",
   "ensemble_output_help": "Ruta al archivo WAV donde se almacenará el resultado del ensamblaje",
   "ensemble_type_print": "Tipo de ensamblaje: {}",
   "num_input_files_print": "Número de archivos de entrada: {}",
@@ -116,7 +196,7 @@
   "reading_chunk_print": "Leyendo fragmento del archivo: {} (inicio: {}s, duración: {}s)",
   "chunk_result_shape_print": "Forma del resultado del fragmento: {}",
   "ensemble_completed_print": "Ensamblaje completado. Salida guardada en: {}",
-  "file_deletion_error": "{} no pudo ser eliminado: {}",
+  "file_deletion_error": "{} no se pudo eliminar: {}",
   "directory_not_exist_warning": "⚠️ El directorio no existe: {}",
   "not_a_directory_warning": "⚠️ La ruta no es un directorio: {}",
   "item_deletion_error": "⚠️ Error al eliminar {}: {}",
@@ -125,67 +205,75 @@
   "error": "🔥 Error: {}",
   "ffmpeg_error": "Error de FFmpeg ({}): {}",
   "file_saved_successfully": "Archivo guardado con éxito: {}",
-  "total_files_found": "Total de archivos encontrados: {}. Usando tasa de muestreo: {}",
+  "total_files_found": "Total de archivos encontrados: {}. Tasa de muestreo utilizada: {}",
   "total_progress": "Progreso total",
-  "detailed_pbar_enabled": "Barra de progreso detallada activada: {}",
+  "detailed_pbar_enabled": "Barra de progreso detallada habilitada: {}",
   "loaded_audio": "Audio cargado: {}, forma: {}",
   "cannot_read_track": "No se puede leer la pista: {}",
   "error_message": "Mensaje de error: {}",
-  "demudding_track": "Desm Mezclando pista (remezcla de fase - instrumental): {}",
+  "demudding_track": "Limpiando pista (remezcla de fase - instrumental): {}",
   "elapsed_time": "Tiempo transcurrido: {:.2f} segundos.",
   "proc_folder_description": "Procesa archivos de audio en una carpeta usando un modelo especificado.",
   "model_type_help": "Tipo de modelo (bandit, bs_roformer, mdx23c, etc.)",
   "config_path_help": "Ruta al archivo de configuración",
   "demud_phaseremix_help": "Habilitar remezcla de fase para instrumental",
   "start_checkpoint_help": "Punto de control inicial para pesos válidos",
-  "input_folder_help": "Carpeta con mezclas para procesar",
-  "audio_path_help": "Ruta a un solo archivo de audio para procesar",
-  "store_dir_help": "Ruta para almacenar los resultados",
+  "input_folder_help": "Carpeta con mezclas a procesar",
+  "audio_path_help": "Ruta a un solo archivo de audio a procesar",
+  "store_dir_help": "Ruta para almacenar resultados",
   "device_ids_help": "Lista de IDs de GPU",
-  "extract_instrumental_help": "Invertir las voces para obtener instrumental si se proporciona",
-  "disable_detailed_pbar_help": "Deshabilitar barra de progreso detallada en la desmezcla",
+  "extract_instrumental_help": "Invertir voces para obtener instrumental si se proporciona",
+  "disable_detailed_pbar_help": "Desactivar barra de progreso detallada en demix",
   "force_cpu_help": "Forzar el uso de CPU incluso si CUDA está disponible",
   "flac_file_help": "Generar archivo FLAC en lugar de WAV",
   "export_format_help": "Formato de exportación y tipo PCM",
   "pcm_type_help": "Tipo PCM para archivos FLAC",
   "use_tta_help": "Habilitar aumento en tiempo de prueba",
   "lora_checkpoint_help": "Punto de control inicial para pesos LoRA",
-  "cuda_available": "CUDA está disponible, usa --force_cpu para deshabilitarlo.",
+  "cuda_available": "CUDA está disponible, use --force_cpu para desactivarlo.",
   "using_device": "Usando dispositivo: {}",
   "instruments_print": "Instrumentos: {}",
   "model_load_time": "Tiempo de carga del modelo: {:.2f} seg",
   "invalid_url": "❌ URL inválida",
   "cookie_file_updated": "✅ ¡Archivo de cookies actualizado!",
-  "cookie_installation_error": "⚠️ Error en la instalación de cookies: {}",
-  "file_size_zero_error": "Tamaño del archivo cero o archivo no creado",
+  "cookie_installation_error": "⚠️ Error de instalación de cookies: {}",
+  "google_drive_mount_error": "❌ Error de conexión de Google Drive: {}",
+  "file_size_zero_error": "Tamaño de archivo cero o archivo no creado",
   "google_drive_error": "❌ Error de descarga de Google Drive: {}",
-  "wav_conversion_failed": "Fallo en la conversión a WAV",
+  "wav_conversion_failed": "Conversión WAV fallida",
   "download_error": "❌ Error de descarga: {}",
+  "file_copied_to_drive": "Archivo copiado a Google Drive: {}",
+  "copy_to_drive_error": "❌ Error al copiar a Google Drive: {}",
+  "direct_download_failed": "Falló la descarga directa del archivo",
+  "direct_download_error": "❌ Error de descarga directa: {}",
+  "invalid_google_drive_url": "❌ URL de Google Drive no válida",
+  "continuing_without_google_drive": "Falló la conexión con Google Drive, continuando sin él...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive no montado, omitiendo copia a Drive...",
   "download_success": "🎉 ¡Descargado con éxito!",
-  "download_failed": "❌ Fallo en la descarga",
+  "download_failed": "❌ Descarga fallida",
   "no_output_files_found": "❌ No se encontraron archivos de salida.",
   "output_refreshed_successfully": "✅ ¡Salida actualizada con éxito!",
-  "error_refreshing_output": "❌ Error al actualizar la salida: {}",
+  "error_refreshing_output": "❌ Error al actualizar salida: {}",
   "starting_audio_separation": "Iniciando separación de audio...",
   "processing_audio": "Procesando audio",
   "separating_audio": "Separando audio... ({:.1f}%)",
-  "separation_complete": "¡Separación completada!",
-  "progress_parsing_error": "Error al analizar el progreso: {}",
+  "separation_complete": "¡Separación completa!",
+  "progress_parsing_error": "Error de análisis de progreso: {}",
   "error_occurred": "Ocurrió un error: {}",
   "separation_process_completed": "¡Proceso de separación completado!",
   "no_audio_file_error": "❌ No se proporcionó archivo de audio y no hay archivo existente en el directorio de entrada.",
   "no_input_progress_label": "Error: No se proporcionó entrada -- 0.0%",
-  "processing_audio_print": "Procesando audio desde: {} usando el modelo: {}",
+  "processing_audio_print": "Procesando audio desde: {} usando modelo: {}",
   "starting_audio_separation_progress_label": "Iniciando separación de audio... -- 0.0%",
   "separating_audio_progress_label": "Separando audio... -- {}.0%",
   "audio_processing_completed_progress_label": "¡Procesamiento de audio completado! -- 100.0%",
   "audio_processing_completed": "✅ ¡Procesamiento de audio completado!",
   "error_occurred_progress_label": "Ocurrió un error -- 0.0%",
-  "minimum_files_required": "⚠️ Se requieren al menos 2 archivos",
+  "minimum_files_required": "⚠️ Se requieren mínimo 2 archivos",
   "valid_files_not_found": "❌ No se encontraron archivos válidos",
   "starting_ensemble_process": "Iniciando proceso de ensamblaje...",
   "ensembling_progress": "Ensamblando... ({:.1f}%)",
-  "finalizing_ensemble_output": "Finalizando salida del ensamblaje...",
+  "finalizing_ensemble_output": "Finalizando salida de ensamblaje...",
   "success_log": "✅ ¡Éxito!\n{}",
   "error_log": "❌ ¡Error!\n{}",
   "critical_error": "⛔ Error crítico: {}",
@@ -197,84 +285,95 @@
   "completed_model": "Modelo completado {}/{}: {}",
   "completed_model_progress_label": "Modelo completado {}/{}: {} -- {}.0%",
   "running_command": "Ejecutando comando: {}",
-  "model_failed": "El modelo {} falló: {}",
+  "model_failed": "Modelo {} falló: {}",
   "critical_error_with_model": "Error crítico con {}: {}",
-  "model_output_failed": "{} falló al producir salida",
-  "waiting_for_files": "Esperando a que todos los archivos estén listos...",
-  "waiting_for_files_progress_label": "Esperando a que todos los archivos estén listos... -- 90.0%",
+  "model_output_failed": "{} no pudo producir salida",
+  "waiting_for_files": "Esperando que todos los archivos estén listos...",
+  "waiting_for_files_progress_label": "Esperando que todos los archivos estén listos... -- 90.0%",
   "performing_ensemble": "Realizando ensamblaje...",
   "performing_ensemble_progress_label": "Realizando ensamblaje... -- 92.0%",
   "memory_usage_before_ensemble": "Uso de memoria antes del ensamblaje: {}%",
   "memory_usage_after_ensemble": "Uso de memoria después del ensamblaje: {}%",
-  "finalizing_ensemble_output_progress_label": "Finalizando salida del ensamblaje... -- 98.0%",
+  "finalizing_ensemble_output_progress_label": "Finalizando salida de ensamblaje... -- 98.0%",
   "ensemble_file_creation_failed": "No se pudo crear el archivo de ensamblaje: {}",
   "ensemble_completed_progress_label": "¡Ensamblaje completado con éxito! -- 100.0%",
   "success_output_created": "✅ ¡Éxito! Archivo de salida creado.",
-  "drive_mounted_copying_ensemble": "Google Drive montado. Copiando salida del ensamblaje...",
-  "drive_already_mounted_copying_ensemble": "Google Drive ya montado. Copiando salida del ensamblaje...",
-  "no_ensemble_output_files_found": "❌ No se encontraron archivos de salida de ensamblaje.",
-  "ensemble_output_copied": "✅ Salida del ensamblaje copiada a {}",
-  "error_copying_ensemble_output": "❌ Error al copiar la salida del ensamblaje: {}",
-  "drive_mounted_copying_files": "Google Drive montado. Copiando archivos...",
-  "drive_already_mounted_copying_files": "Google Drive ya montado. Copiando archivos...",
-  "files_copied_to_drive": "✅ Archivos copiados a {}",
-  "error_copying_files": "❌ Error al copiar archivos: {}",
-  "mounting_drive": "Montando Google Drive...",
-  "settings_saved": "Configuración guardada en config.json: Idioma={lang}, Método={method}, Puerto={port}",
-  "selected_language": "Idioma seleccionado: {lang}",
-  "available_languages": "Idiomas disponibles: {langs}",
-  "select_language_prompt": "Selecciona un idioma (predeterminado: {default}): ",
-  "invalid_language": "¡Selección de idioma inválida! Usando idioma predeterminado ({default}).",
-  "available_sharing_methods": "Métodos de compartición disponibles: gradio, localtunnel, ngrok",
-  "select_sharing_method_prompt": "Elige un método de compartición (predeterminado: {default}): ",
-  "enter_ngrok_token_prompt": "Ingresa tu token de Ngrok (predeterminado: {default}): ",
-  "ngrok_token_required": "Error: ¡Se requiere un token de Ngrok para el método ngrok!",
-  "enter_port_prompt": "Ingresa el número de puerto (predeterminado: {default}): ",
-  "opening_previous_url": "Abriendo URL anterior: {url}",
-  "open_url_manually": "Abre la URL manualmente en Colab: {url}",
-  "invalid_previous_url": "URL anterior inválida: {url}, no abierta.",
-  "starting_method": "Iniciando {method} en el puerto {port}...",
-  "process_stopped": "\n🛑 Proceso detenido por el usuario",
-  "selected_port": "Puerto seleccionado: {port}",
-  "starting_gradio_with_sharing": "Iniciando Gradio con compartición integrada...",
-  "starting_localtunnel": "Iniciando Localtunnel en el puerto {port}...",
-  "share_link": "Enlace para compartir: {url}",
-  "password_ip": "Contraseña IP: {ip}",
-  "starting_ngrok": "Iniciando Ngrok en el puerto {port}...",
-  "ngrok_url": "URL de Ngrok: {url}",
-  "ngrok_error": "Error al iniciar Ngrok: {error}",
-  "apollo_chunk_size": "Tamaño de fragmento Apollo",
-  "apollo_chunk_size_info": "Tamaño de fragmento para mejora Apollo (recomendado: 19 para modelo universal)",
-  "apollo_overlap": "Solapamiento Apollo",
-  "enhancing_with_apollo": "Mejorando con Apollo ({}/{} archivos)...",
-  "apollo_processing_completed": "¡Procesamiento Apollo completado!",
-  "apollo_overlap_info": "Solapamiento para mejora Apollo (recomendado: 2)",
-  "invalid_method": "Error: ¡Método inválido! Usa 'gradio', 'localtunnel' o 'ngrok'.",
-  "apollo_enhancement_settings": "Configuración de mejora Apollo",
-  "enhance_with_apollo": "Mejorar con Apollo",
-  "enhance_with_apollo_info": "Habilitar Apollo para mejora de audio después de la separación",
-  "apollo_method": "Método Apollo",
-  "apollo_method_info": "Selecciona el método de procesamiento para Apollo",
+  "phase_fixer_tab": "🔧 Corrector de fase",
+  "batch_processing_tab": "📦 Procesamiento por lotes",
+  "batch_description": "Procesa múltiples archivos a la vez con el mismo modelo",
+  "batch_input_folder": "Carpeta de entrada",
+  "batch_input_folder_placeholder": "Ingresa la ruta de la carpeta con archivos de audio",
+  "batch_output_folder": "Carpeta de salida",
+  "batch_output_folder_placeholder": "Ingresa la ruta de la carpeta de salida",
+  "batch_file_list": "Archivos a procesar",
+  "batch_add_files": "Añadir archivos",
+  "batch_clear_list": "Limpiar lista",
+  "batch_start": "🚀 Iniciar procesamiento por lotes",
+  "batch_stop": "⏹️ Detener",
+  "batch_progress": "Progreso del lote",
+  "batch_current_file": "Procesando actualmente",
+  "batch_completed": "✅ ¡Procesamiento por lotes completado!",
+  "batch_stopped": "⏹️ Procesamiento por lotes detenido",
+  "batch_no_files": "❌ No se seleccionaron archivos para procesar",
+  "source_model": "Modelo fuente (Voces)",
+  "source_model_info": "Modelo para extraer voces/datos de fase",
+  "target_model": "Modelo objetivo (Instrumental)",
+  "target_model_info": "Modelo al que se aplica la corrección de fase",
+  "custom_models_tab": "⚡ Modelos personalizados",
+  "Custom Models": "Modelos personalizados",
+  "add_custom_model": "➕ Añadir modelo personalizado",
+  "custom_model_name": "Nombre del modelo",
+  "custom_model_name_placeholder": "Ingresa un nombre para tu modelo",
+  "checkpoint_url": "URL del Checkpoint (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "URL de configuración (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "URL de archivo Python personalizado (opcional)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "Detectar tipo de modelo automáticamente",
+  "model_type": "Tipo de modelo",
+  "add_model_btn": "➕ Añadir modelo",
+  "custom_models_list": "📋 Lista de modelos personalizados",
+  "delete_model": "🗑️ Eliminar",
+  "no_custom_models": "Aún no se han añadido modelos personalizados",
+  "model_added_success": "✅ ¡Modelo añadido con éxito!",
+  "model_deleted_success": "✅ ¡Modelo eliminado con éxito!",
+  "model_add_error": "❌ Error al añadir el modelo: {}",
+  "model_delete_error": "❌ Error al eliminar el modelo: {}",
+  "refresh_models": "🔄 Actualizar",
+  "custom_model_info": "Añade modelos personalizados proporcionando enlaces de descarga. Los modelos se descargarán automáticamente cuando se utilicen.",
+  "select_model_to_delete": "Selecciona un modelo para eliminar",
+  "downloading_model_file": "Descargando modelo: {}",
+  "downloading_file_progress": "Descargando: {} - {}%",
+  "chunk_size_mode": "Modo de Tamaño de Fragmento",
+  "chunk_size_mode_info": "base: presets estándar | custom: introduce tu propio valor | yaml: leer del YAML del modelo",
+  "chunk_size_custom_label": "Tamaño de Fragmento Personalizado",
+  "chunk_size_custom_info": "Introduce un valor personalizado de tamaño de fragmento",
+  "chunk_size_yaml_label": "Tamaño de Fragmento YAML",
+  "chunk_size_yaml_display_info": "Se leerá del archivo de configuración YAML del modelo tras la descarga",
+  "chunk_size_yaml_not_downloaded": "⏳ YAML del modelo aún no descargado — el valor se leerá durante el procesamiento",
+  "chunk_size_yaml_detected": "✅ Del YAML: {}",
+  "remove_favorite": "Eliminar de Favoritos",
   "normal_method": "Método normal",
-  "mid_side_method": "Método centro/lateral",
-  "apollo_normal_model": "Modelo Apollo normal",
-  "apollo_normal_model_info": "Modelo a usar para procesamiento Apollo normal",
-  "apollo_midside_model": "Modelo Apollo centro/lateral",
-  "apollo_enhancement_info": "Mejora de la calidad del sonido",
-  "selected_models": "Modelos Seleccionados",
-  "save_preset": "Guardar Preajuste",
-  "delete_preset": "Eliminar Preajuste",
-  "refresh_presets": "Actualizar Preajustes",
-  "preset_name": "Nombre del Preajuste",
-  "select_preset": "Seleccionar Preajuste",
-  "add_favorite": "Añadir a Favoritos",
-  "apollo_midside_model_info": "Modelo a usar para procesamiento centro/lateral (opcional)",
-  "language_changed_message": "Idioma cambiado. Por favor, reinicia la celda.",
-  "apply_matchering": "Aplicar Matchering",
-  "matchering_info": "Aplicar Matchering para mejorar el masterizado y el balance del audio",
-  "matchering_passes": "Pasadas de Matchering",
-  "matchering_passes_info": "Número de iteraciones de Matchering (1-5, valores más altos mejoran la calidad pero toman más tiempo)",
-  "matchering_processing": "Aplicando Matchering... ({:.1f}%)",
-  "matchering_completed": "✅ ¡Procesamiento de Matchering completado!",
-  "matchering_error": "❌ Error durante Matchering: {}"
+  "mid_side_method": "Método Mid-Side",
+  "apollo_processing_method": "Método de Procesamiento Apollo",
+  "apollo_normal_model": "Modelo Normal Apollo",
+  "apollo_mid_side_model": "Modelo Mid-Side Apollo",
+  "scale_factor": "Factor de Escala",
+  "scale_factor_info": "Factor de escala de fase (recomendado: 1.4)",
+  "source_file_label": "Archivo Fuente (Voces)",
+  "target_file_label": "Archivo Destino (Instrumental)",
+  "run_phase_fixer": "🔧 Ejecutar Corrector de Fase",
+  "phase_fixed_output": "Salida con Fase Corregida",
+  "phase_fixer_settings": "Ajustes del Corrector de Fase",
+  "low_cutoff": "Corte Bajo (Hz)",
+  "high_cutoff": "Corte Alto (Hz)",
+  "low_cutoff_info": "Frecuencia de corte baja para la corrección de fase",
+  "high_cutoff_info": "Frecuencia de corte alta para la corrección de fase",
+  "no_preset_name_provided": "⚠️ No se proporcionó nombre de preset",
+  "no_models_selected_for_preset": "⚠️ No se seleccionaron modelos para el preset",
+  "preset_saved": "✅ ¡Preset '{}' guardado!",
+  "please_upload_both_files": "⚠️ Por favor sube tanto el archivo fuente como el destino",
+  "processing_log_placeholder": "El registro de procesamiento aparecerá aquí...",
+  "enter_preset_name": "Introduce el nombre del preset..."
 }

assets/i18n/languages/fr_fr.json

@@ -209,7 +209,6 @@
   "finalizing_ensemble_output_progress_label": "Finalisation de la sortie de l'assemblage... -- 98.0%",
   "ensemble_file_creation_failed": "Le fichier d'assemblage n'a pas pu être créé : {}",
   "ensemble_completed_progress_label": "Assemblage terminé avec succès ! -- 100.0%",
-  "success_output_created": "✅ Succès ! Fichier de sortie créé.",
   "drive_mounted_copying_ensemble": "Google Drive monté. Copie de la sortie de l'assemblage...",
   "drive_already_mounted_copying_ensemble": "Google Drive déjà monté. Copie de la sortie de l'assemblage...",
   "no_ensemble_output_files_found": "❌ Aucun fichier de sortie d'assemblage trouvé.",
@@ -269,5 +268,90 @@
   "select_preset": "Sélectionner un Préréglage",
   "add_favorite": "Ajouter aux Favoris",
   "apollo_midside_model_info": "Modèle à utiliser pour le traitement centre/côté (optionnel)",
-  "language_changed_message": "Langue modifiée. Veuillez redémarrer la cellule."
-}
+    "language_changed_message": "Langue modifiée. Veuillez redémarrer la cellule.",
+  "Other Shapes": "Autres modèles",
+  "use_phase_correction": "Appliquer la correction de phase avancée",
+  "google_drive_mount_error": "❌ Erreur de liaison Google Drive : {}",
+  "file_copied_to_drive": "Fichier copié sur Google Drive : {}",
+  "copy_to_drive_error": "❌ Erreur lors de la copie sur Google Drive : {}",
+  "direct_download_failed": "Échec du téléchargement direct du fichier",
+  "direct_download_error": "❌ Erreur de téléchargement direct : {}",
+  "invalid_google_drive_url": "❌ URL Google Drive non valide",
+  "continuing_without_google_drive": "La connexion Google Drive a échoué, poursuite sans...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive non monté, omission de la copie sur Drive...",
+  "auto_apollo_chunk_size": "Taille de fragment Apollo",
+  "auto_apollo_chunk_size_info": "Taille de fragment pour l'amélioration Apollo (recommandé : 19 pour le modèle universel)",
+  "auto_apollo_overlap": "Chevauchement Apollo",
+  "auto_enhancing_with_apollo": "Amélioration avec Apollo ({}/{} fichiers)...",
+  "auto_apollo_processing_completed": "Traitement Apollo terminé !",
+  "auto_apollo_overlap_info": "Chevauchement pour l'amélioration Apollo (recommandé : 2)",
+  "apply_matchering": "Appliquer Matchering",
+  "matchering_info": "Appliquer Matchering pour améliorer le mastering et l'équilibre audio",
+  "matchering_passes": "Passages Matchering",
+  "matchering_passes_info": "Nombre d'itérations Matchering (1-5, des valeurs plus élevées améliorent la qualité mais prennent plus de temps)",
+  "matchering_processing": "Application de Matchering... ({:.1f}%)",
+  "matchering_completed": "✅ Traitement Matchering terminé !",
+  "matchering_error": "❌ Erreur pendant Matchering : {}",
+  "backend_settings": "🚀 Paramètres Backend",
+  "inference_backend": "Backend d'Inference",
+  "optimization_mode": "Mode d'Optimisation",
+  "optimization_mode_info": "Sélection du mode d'optimisation PyTorch",
+  "mixed_precision_amp": "Précision Mixte (AMP)",
+  "mixed_precision_info": "Inférence 2x plus rapide - recommandé",
+  "tf32_acceleration": "Accélération TF32",
+  "tf32_acceleration_info": "Augmentation de vitesse supplémentaire pour RTX 30xx+",
+  "cudnn_benchmark": "Benchmark cuDNN",
+  "cudnn_benchmark_info": "Plus lent lors de la première exécution, beaucoup plus rapide lors des exécutions suivantes",
+  "ultra_optimized_pytorch": "Backend PyTorch ULTRA-OPTIMISÉ",
+  "default_active_max_speed": "Actif par défaut - Optimisation de vitesse maximale",
+  "channels_last_mode": "Le plus rapide pour les GPU RTX",
+  "compile_mode": "Vitesse supplémentaire pour PyTorch 2.0+",
+  "default_mode": "Standard",
+  "phase_fixer_tab": "🔧 Correcteur de phase",
+  "batch_processing_tab": "📦 Traitement par lots",
+  "batch_description": "Traiter plusieurs fichiers à la fois avec le même modèle",
+  "batch_input_folder": "Dossier d'entrée",
+  "batch_input_folder_placeholder": "Entrez le chemin du dossier contenant les fichiers audio",
+  "batch_output_folder": "Dossier de sortie",
+  "batch_output_folder_placeholder": "Entrez le chemin du dossier de sortie",
+  "batch_file_list": "Fichiers à traiter",
+  "batch_add_files": "Ajouter des fichiers",
+  "batch_clear_list": "Vider la liste",
+  "batch_start": "🚀 Démarrer le traitement par lots",
+  "batch_stop": "⏹️ Arrêter",
+  "batch_progress": "Progression du lot",
+  "batch_current_file": "En cours de traitement",
+  "batch_completed": "✅ Traitement par lots terminé !",
+  "batch_stopped": "⏹️ Traitement par lots arrêté",
+  "batch_no_files": "❌ Aucun fichier sélectionné pour le traitement",
+  "source_model": "Modèle source (Voix)",
+  "source_model_info": "Modèle pour extraire les voix/données de phase",
+  "target_model": "Modèle cible (Instrumental)",
+  "target_model_info": "Modèle auquel appliquer la correction de phase",
+  "custom_models_tab": "⚡ Modèles personnalisés",
+  "Custom Models": "Modèles personnalisés",
+  "add_custom_model": "➕ Ajouter un modèle personnalisé",
+  "custom_model_name": "Nom du modèle",
+  "custom_model_name_placeholder": "Entrez un nom pour votre modèle",
+  "checkpoint_url": "URL du Checkpoint (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "URL de configuration (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "URL du fichier Python personnalisé (optionnel)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "Détecter automatiquement le type de modèle",
+  "model_type": "Type de modèle",
+  "add_model_btn": "➕ Ajouter le modèle",
+  "custom_models_list": "📋 Liste des modèles personnalisés",
+  "delete_model": "🗑️ Supprimer",
+  "no_custom_models": "Aucun modèle personnalisé ajouté pour l'instant",
+  "model_added_success": "✅ Modèle ajouté avec succès !",
+  "model_deleted_success": "✅ Modèle supprimé avec succès !",
+  "model_add_error": "❌ Erreur lors de l'ajout du modèle : {}",
+  "model_delete_error": "❌ Erreur lors de la suppression du modèle : {}",
+  "refresh_models": "🔄 Actualiser",
+  "custom_model_info": "Ajoutez des modèles personnalisés en fournissant des liens de téléchargement. Les modèles seront téléchargés automatiquement lors de leur utilisation.",
+  "select_model_to_delete": "Sélectionnez un modèle à supprimer",
+  "success_output_created": "✅ Succès ! Fichier de sortie créé.",
+  "downloading_model_file": "Téléchargement du modèle : {}",
+  "downloading_file_progress": "Téléchargement : {} - {}%",

assets/i18n/languages/hi_in.json

@@ -17,11 +17,10 @@
   "overlap": "ओवरलैप",
   "overlap_info": "अनुशंसित: 2-10 (उच्च मान गुणवत्ता बढ़ाते हैं लेकिन अधिक VRAM की आवश्यकता होती है)",
   "tta_boost": "TTA बूस्ट",
-  "tta_info": "गुणवत्ता में सुधार करता है लेकिन प्रसंस्करण गति को धीमा करता है",
   "phase_fix": "फेज सुधार",
+  "use_phase_correction": "उन्नत फेज सुधार लागू करें",
   "phase_fix_info": "वाद्य ट्रैक के लिए उन्नत फेज सुधार",
   "instrumental": "वाद्य",
-  "instrumental_info": "आमतौर पर 2 आउटपुट देता है, लेकिन कभी-कभी एक ही आउटपुट देता है, आपको कोशिश करनी होगी :)",
   "process": "🚀 प्रक्रिया करें",
   "reset": "🧹 रीसेट करें",
   "status": "स्थिति",
@@ -73,215 +72,274 @@
   "ensemble_copy_status": "संयोजन आउटपुट यहाँ कॉपी किया जाएगा...",
   "guidelines": "<h4>मॉडल चयन दिशानिर्देश</h4><ul><li><strong>क्रॉस-श्रेणी मिश्रण से बचें:</strong> स्वर और वाद्य मॉडल का संयोजन अवांछित मिश्रण बना सकता है</li><li><strong>विशेष मॉडल नोट्स:</strong><ul><li>द्वैत मॉडल (v1/v2) - दोनों स्टेम्स आउटपुट करते हैं</li><li>MDX23C सेपरेटर - हाइब्रिड परिणाम</li></ul></li><li><strong>सर्वोत्तम अभ्यास:</strong> एक ही श्रेणी से 3-5 समान मॉडल का उपयोग करें</li></ul><div>💡 प्रो टिप: \"VOCALS-MelBand-Roformer BigBeta5e\" + \"VOCALS-BS-Roformer_1297\" संयोजन से शुरू करें</div>",
   "download_sources_tab": "⬇️ स्रोत डाउनलोड करें",
-  "direct_links": "🌐 सीधे लिंक",
+  "direct_links": "🌐 डायरेक्ट लिंक्स",
   "audio_file_url": "ऑडियो फाइल URL",
   "download_from_url": "⬇️ URL से डाउनलोड करें",
   "download_status": "डाउनलोड स्थिति",
   "downloaded_file": "डाउनलोड की गई फाइल",
   "cookie_management": "🍪 कुकी प्रबंधन",
-  "upload_cookies_txt": "Cookies.txt अपलोड करें",
-  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 क्यों जरूरी?**<br>- आयु-प्रतिबंधित सामग्री तक पहुंच<br>- निजी/असूचीबद्ध वीडियो डाउनलोड करें<br>- क्षेत्रीय प्रतिबंधों को बायपास करें<br>- यूट्यूब डाउनलोड सीमाओं से बचें<br><br>**⚠️ महत्वपूर्ण नोट्स**<br>- अपनी कुकी फाइलें कभी साझा न करें!<br>- कुकीज़ को ताज़ा करें जब:<br>  • \"403 निषिद्ध\" त्रुटियाँ मिलें<br>  • डाउनलोड अचानक रुक जाएँ<br>  • \"सत्र समाप्त\" संदेश दिखें<br><br>**🔄 नवीकरण चरण**<br>1. इस <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">क्रोम एक्सटेंशन</a> को स्थापित करें<br>2. क्रोम में यूट्यूब में लॉगिन करें<br>3. एक्सटेंशन आइकन पर क्लिक करें → \"निर्यात\"<br>4. डाउनलोड की गई फाइल यहाँ अपलोड करें<br><br>**⏳ कुकी जीवनकाल**<br>- सामान्य सत्र: 24 घंटे<br>- संवेदनशील संचालन: 1 घंटा<br>- पासवर्ड परिवर्तन: तत्काल अमान्यता</div>",
-  "manual_ensemble_tab": "🎚️ मैनुअल संयोजन",
+  "upload_cookies_txt": "कुकीज.txt अपलोड करें",
+  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 क्यों आवश्यक?**<br>- आयु-प्रतिबंधित सामग्री तक पहुंच<br>- न��जी/अप्रकाशित वीडियो डाउनलोड<br>- क्षेत्रीय प्रतिबंधों को बायपास करें<br>- यूट्यूब डाउनलोड सीमाओं से बचें<br><br>**⚠️ महत्वपूर्ण नोट्स**<br>- कभी भी अपनी कुकी फाइलें साझा न करें!<br>- जब रिफ्रेश करें:<br>  • \"403 निषिद्ध\" त्रुटियां प्राप्त हो रही हों<br>  • डाउनलोड अचानक रुक जाएं<br>  • \"सत्र समाप्त\" संदेश देखें<br><br>**🔄 नवीनीकरण चरण**<br>1. इस <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">क्रोम एक्सटेंशन</a> को इंस्टॉल करें<br>2. क्रोम में यूट्यूब पर लॉगिन करें<br>3. एक्सटेंशन आइकन पर क्लिक करें → \"एक्सपोर्ट\"<br>4. यहां डाउनलोड की गई फाइल अपलोड करें<br><br>**⏳ कुकी जीवनकाल**<br>- सामान्य सत्र: 24 घंटे<br>- संवेदनशील संचालन: 1 घंटा<br>- पासवर्ड परिवर्तन: तत्काल रिफ्रेश</div>",
+  "ensemble_output_copied": "✅ संयोजन आउटपुट {} में कॉपी किया गया",
+  "error_copying_ensemble_output": "❌ संयोजन आउटपुट कॉपी करने में त्रुटि: {}",
+  "drive_mounted_copying_ensemble": "गूगल ड्राइव माउंटेड। संयोजन आउटपुट कॉपी कर रहा है...",
+  "drive_already_mounted_copying_ensemble": "गूगल ड्राइव पहले से माउंटेड। संयोजन आउटपुट कॉपी कर रहा है...",
+  "no_ensemble_output_files_found": "❌ कोई संयोजन आउटपुट फाइल नहीं मिली।",
+  "drive_mounted_copying_files": "गूगल ड्राइव माउंटेड। फाइलें कॉपी कर रहा है...",
+  "drive_already_mounted_copying_files": "गूगल ड्राइव पहले से माउंटेड। फाइलें कॉपी कर रहा है...",
+  "files_copied_to_drive": "✅ फाइलें {} में कॉपी की गईं",
+  "error_copying_files": "❌ फाइलें कॉपी करने में त्रुटि: {}",
+  "mounting_drive": "गूगल ड्राइव माउंट कर रहा है...",
+  "settings_saved": "सेटिंग्स config.json में सहेजी गईं: भाषा={lang}, विधि={method}, पोर्ट={port}",
+  "selected_language": "चयनित भाषा: {lang}",
+  "available_languages": "उपलब्ध भाषाएं: {langs}",
+  "select_language_prompt": "भाषा चुनें (डिफ़ॉल्ट: {default}): ",
+  "invalid_language": "अमान्य भाषा चयन! डिफ़ॉल्ट भाषा ({default}) का उपयोग कर रहा है।",
+  "available_sharing_methods": "उपलब्ध साझाकरण विधियां: gradio, localtunnel, ngrok",
+  "select_sharing_method_prompt": "साझाकरण विधि चुनें (डिफ़ॉल्ट: {default}): ",
+  "enter_ngrok_token_prompt": "अपना Ngrok टोकन दर्ज करें (डिफ़ॉल्ट: {default}): ",
+  "ngrok_token_required": "त्रुटि: ngrok विधि के लिए Ngrok टोकन आवश्यक है!",
+  "enter_port_prompt": "पोर्ट नंबर दर्ज करें (डिफ़ॉल्ट: {default}): ",
+  "opening_previous_url": "पिछला URL खोल रहा है: {url}",
+  "open_url_manually": "Colab में URL मैन्युअल रूप से खोलें: {url}",
+  "invalid_previous_url": "अमान्य पिछला URL: {url}, नहीं खोला गया।",
+  "starting_method": "{method} को पोर्ट {port} पर शुरू कर रहा है...",
+  "process_stopped": "\n🛑 उपयोगकर्ता द्वारा प्रक्रिया रोक दी गई",
+  "selected_port": "चयनित पोर्ट: {port}",
+  "starting_gradio_with_sharing": "Gradio को अंतर्निहित साझाकरण के साथ शुरू कर रहा है...",
+  "starting_localtunnel": "Localtunnel को पोर्ट {port} प��� शुरू कर रहा है...",
+  "share_link": "साझा लिंक: {url}",
+  "password_ip": "पासवर्ड IP: {ip}",
+  "starting_ngrok": "Ngrok को पोर्ट {port} पर शुरू कर रहा है...",
+  "ngrok_url": "Ngrok URL: {url}",
+  "ngrok_error": "Ngrok शुरू करने में त्रुटि: {error}",
+  "auto_apollo_chunk_size": "Apollo खंड आकार",
+  "auto_apollo_chunk_size_info": "Apollo सुधार के लिए खंड आकार (यूनिवर्सल मॉडल के लिए अनुशंसित: 19)",
+  "auto_apollo_overlap": "Apollo ओवरलैप",
+  "auto_enhancing_with_apollo": "Apollo के साथ सुधार कर रहा है ({}/{} फाइलें)...",
+  "auto_apollo_processing_completed": "Apollo प्रसंस्करण पूरा!",
+  "auto_apollo_overlap_info": "Apollo सुधार के लिए ओवरलैप (अनुशंसित: 2)",
+  "apollo_chunk_size": "Apollo खंड आकार",
+  "apollo_chunk_size_info": "Apollo सुधार के लिए खंड आकार (यूनिवर्सल मॉडल के लिए अनुशंसित: 19)",
+  "apollo_overlap": "Apollo ओवरलैप",
+  "enhance_with_apollo": "Apollo के साथ सुधार करें",
+  "apollo_enhancement_info": "ध्वनि गुणवत्ता सुधार",
+  "enhancing_with_apollo": "Apollo के साथ सुधार कर रहा है ({}/{} फाइलें)...",
+  "apollo_processing_completed": "Apollo प्रसंस्करण पूरा!",
+  "apollo_overlap_info": "Apollo सुधार के लिए ओवरलैप (अनुशंसित: 2)",
+  "selected_models": "चयनित मॉडल",
+  "save_preset": "प्रीसेट सहेजें",
+  "delete_preset": "प्रीसेट हटाएं",
+  "refresh_presets": "प्रीसेट रिफ्रेश करें",
+  "preset_name": "प्रीसेट का नाम",
+  "select_preset": "प्रीसेट चुनें",
+  "add_favorite": "पसंदीदा में जोड़ें",
+  "apply_matchering": "मैचरिंग लागू करें",
+  "matchering_info": "ऑडियो मास्टरिंग और संतुलन को बढ़ाने के लिए मैचरिंग लागू करें",
+  "matchering_passes": "मैचरिंग पास",
+  "matchering_passes_info": "मैचरिंग की पुनरावृत्तियों की संख्या (1-5, उच्च मान गुणवत्ता को बेहतर बनाते हैं लेकिन अधिक समय लेते हैं)",
+  "matchering_processing": "मैचरिंग लागू किया जा रहा है... ({:.1f}%)",
+  "matchering_completed": "✅ मैचरिंग प्रक्रिया पूरी हुई!",
+  "matchering_error": "❌ मैचरिंग के दौरान त्रुटि: {}",
+  "invalid_method": "त्रुटि: अमान्य विधि! 'gradio', 'localtunnel', या 'ngrok' का उपयोग करें।",
+  "language_changed_message": "भाषा बदली गई। कृपया सेल को पुनः शुरू करें।",
+  "backend_settings": "🚀 बैकएंड सेटिंग्स",
+  "inference_backend": "इन्फरेंस बैकएंड",
+  "optimization_mode": "ऑप्टिमाइजेशन मोड",
+  "optimization_mode_info": "PyTorch ऑप्टिमाइजेशन मोड चयन",
+  "mixed_precision_amp": "मिश्रित प्रेसिजन (AMP)",
+  "mixed_precision_info": "2x तेज इन्फरेंस - अनुशंसित",
+  "tf32_acceleration": "TF32 असेलरेशन",
+  "tf32_acceleration_info": "RTX 30xx+ के लिए अतिरिक्त गति बूस्ट",
+  "cudnn_benchmark": "cuDNN बेंचमार्क",
+  "cudnn_benchmark_info": "पहले रन में धीमा, बाद के रनों में बहुत तेज",
+  "ultra_optimized_pytorch": "अल्ट्रा-ऑप्टिमाइज्ड PyTorch बैकएंड",
+  "default_active_max_speed": "डिफ़ॉल्ट रूप से सक्रिय - अधिकतम गति ऑप्टिमाइजेशन",
+  "channels_last_mode": "RTX GPUs के लिए सबसे तेज",
+  "compile_mode": "PyTorch 2.0+ के लिए अतिरिक्त गति",
+  "default_mode": "मानक",
+  "tta_info": "गुणवत्ता बढ़ाता है लेकिन प्रसंस्करण को धीमा करता है",
+  "instrumental_info": "आमतौर पर 2 स���टेम आउटपुट करता है, लेकिन कभी-कभी केवल एक - आपको कोशिश करनी होगी! :)",
+  "direct_download_error": "❌ सीधा डाउनलोड त्रुटि: {}",
+  "invalid_google_drive_url": "❌ अमान्य Google Drive URL",
+  "continuing_without_google_drive": "Google Drive कनेक्शन विफल, इसके बिना जारी रखते हुए...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive माउंट नहीं है, Drive पर कॉपी छोड़ रहे हैं...",
+  "direct_download_failed": "फ़ाइल को सीधे डाउनलोड करना विफल रहा",
+  "manual_ensemble_tab": "🎚️ मैनुअल एन्सेम्बल",
+  "phase_fixer_tab": "🔧 फेज़ फिक्सर",
+  "batch_processing_tab": "📦 बैच प्रोसेसिंग",
+  "batch_description": "एक ही मॉडल से एक साथ कई फाइलें प्रोसेस करें",
+  "batch_input_folder": "इनपुट फोल्डर",
+  "batch_input_folder_placeholder": "ऑडियो फाइलों वाले फोल्डर का पथ दर्ज करें",
+  "batch_output_folder": "आउटपुट फोल्डर",
+  "batch_output_folder_placeholder": "आउटपुट फाइलों के लिए फोल्डर पथ दर्ज करें",
+  "batch_file_list": "प्रोसेस करने के लिए फाइलें",
+  "batch_add_files": "फाइलें जोड़ें",
+  "batch_clear_list": "सूची साफ करें",
+  "batch_start": "🚀 बैच प्रोसेसिंग शुरू करें",
+  "batch_stop": "⏹️ रोकें",
+  "batch_progress": "बैच प्रगति",
+  "batch_current_file": "वर्तमान में प्रक्रिया हो रही है",
+  "batch_completed": "✅ बैच प्रोसेसिंग पूर्ण!",
+  "batch_stopped": "⏹️ बैच प्रोसेसिंग रोकी गई",
+  "batch_no_files": "❌ प्रोसेसिंग के लिए कोई फाइल नहीं चुनी गई",
+  "source_model": "स्रोत मॉडल (स्वर)",
+  "source_model_info": "स्वर/फेज़ डेटा निकालने के लिए मॉडल",
+  "target_model": "लक्ष्य मॉडल (वाद्य)",
+  "target_model_info": "फेज़ सुधार लागू करने के लिए मॉडल",
   "input_sources": "📂 इनपुट स्रोत",
-  "refresh": "🔄 ताज़ा करें",
-  "ensemble_algorithm": "संयोजन एल्गोरिदम",
+  "refresh": "🔄 रिफ्रेश",
+  "ensemble_algorithm": "एन्सेम्बल एल्गोरिदम",
   "select_audio_files": "ऑडियो फाइलें चुनें",
   "available_files": "उपलब्ध फाइलें",
-  "custom_weights": "कस्टम वेट्स (अल्पविराम से अलग)",
+  "custom_weights": "कस्टम वज़न (अल्पविराम से अलग)",
   "custom_weights_placeholder": "उदाहरण: 0.8, 1.2, 1.0, ...",
-  "custom_weights_info": "समान वेट्स के लिए खाली छोड़ें",
+  "custom_weights_info": "समान वज़न के लिए खाली छोड़ें",
   "result_preview_tab": "🎧 परिणाम पूर्वावलोकन",
-  "ensembled_output": "संयोजित आउटपुट",
-  "processing_log_tab": "📋 प्रसंस्करण लॉग",
-  "processing_details": "प्रसंस्करण विवरण",
-  "process_ensemble": "⚡ संयोजन प्रक्रिया करें",
-  "languages_tab": "🌐 भाषाएँ",
+  "ensembled_output": "एन्सेम्बल आउटपुट",
+  "processing_log_tab": "📋 प्रोसेसिंग लॉग",
+  "processing_details": "प्रोसेसिंग विवरण",
+  "process_ensemble": "⚡ एन्सेम्बल प्रोसेस करें",
+  "languages_tab": "🌐 भाषाएं",
   "Vocal Models": "स्वर मॉडल",
   "Instrumental Models": "वाद्य मॉडल",
   "4-Stem Models": "4-स्टेम मॉडल",
-  "Denoise Models": "शोर हटाने के मॉडल",
-  "Dereverb Models": "रिवर्ब हटाने के मॉडल",
-  "Other Models": "अन्य मॉडल",
-  "ensemble_files_description": "निर्दिष्ट संयोजन विधियों का उपयोग करके कई ऑडियो फाइलों को एक ��ी आउटपुट में जोड़ता है।",
-  "ensemble_files_help": "संयोजन के लिए सभी ऑडियो फाइलों का पथ",
-  "ensemble_type_help": "avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft में से एक",
-  "ensemble_weights_help": "संयोजन बनाने के लिए वेट्स। वेट्स की संख्या फाइलों की संख्या के बराबर होनी चाहिए",
-  "ensemble_output_help": "WAV फाइल का पथ जहां संयोजन परिणाम संग्रहीत होगा",
-  "ensemble_type_print": "संयोजन प्रकार: {}",
+  "Denoise Models": "डीनॉइज़ मॉडल",
+  "Dereverb Models": "डीरिवर्ब मॉडल",
+  "Other Shapes": "अन्य मॉडल",
+  "Custom Models": "कस्टम मॉडल",
+  "ensemble_files_description": "निर्दिष्ट एन्सेम्बल विधियों का उपयोग करके कई ऑडियो फाइलों को एक आउटपुट में जोड़ता है।",
+  "ensemble_files_help": "एन्सेम्बल के लिए सभी ऑडियो फाइलों का पथ",
+  "ensemble_type_help": "इनमें से एक: avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
+  "ensemble_weights_help": "एन्सेम्बल बनाने के लिए वज़न। वज़न की संख्या फाइलों की संख्या के बराबर होनी चाहिए",
+  "ensemble_output_help": "WAV फाइल का पथ जहाँ एन्सेम्बल परिणाम संग्रहीत होगा",
+  "ensemble_type_print": "एन्सेम्बल प्रकार: {}",
   "num_input_files_print": "इनपुट फाइलों की संख्या: {}",
-  "weights_print": "वेट्स: {}",
+  "weights_print": "वज़न: {}",
   "output_file_print": "आउटपुट फाइल: {}",
   "duration_mismatch_error": "सभी फाइलों की अवधि समान होनी चाहिए",
-  "file_not_found_error": "त्रुटि। फाइल नहीं मिली: {}। पथ जांचें।",
-  "reading_chunk_print": "फाइल से खंड पढ़ रहा है: {} (शुरुआत: {}सेकंड, अवधि: {}सेकंड)",
-  "chunk_result_shape_print": "खंड परिणाम आकार: {}",
-  "ensemble_completed_print": "संयोजन पूरा हुआ। आउटपुट यहाँ संग्रहीत: {}",
+  "file_not_found_error": "त्रुटि। फाइल नहीं मिली: {}। पथ जाँचें।",
+  "reading_chunk_print": "फाइल से खंड पढ़ रहे हैं: {} (शुरू: {}s, अवधि: {}s)",
+  "chunk_result_shape_print": "खंड परिणाम का आकार: {}",
+  "ensemble_completed_print": "एन्सेम्बल पूर्ण। आउटपुट यहाँ सहेजा गया: {}",
   "file_deletion_error": "{} हटाया नहीं जा सका: {}",
-  "directory_not_exist_warning": "⚠️ डायरेक्टरी मौजूद नहीं है: {}",
-  "not_a_directory_warning": "⚠️ पथ डायरेक्टरी नहीं है: {}",
+  "directory_not_exist_warning": "⚠️ डायरेक्टरी मौजूद नहीं: {}",
+  "not_a_directory_warning": "⚠️ पथ एक डायरेक्टरी नहीं है: {}",
   "item_deletion_error": "⚠️ {} हटाने में त्रुटि: {}",
-  "old_output_not_exist": "❌ पुराना आउटपुट फ़ोल्डर मौजूद नहीं है",
-  "old_outputs_cleared": "✅ पुराने आउटपुट सफलतापूर्वक हटा दिए गए!",
+  "critical_error": "⛔ गंभीर त्रुटि: {}",
+  "old_output_not_exist": "❌ पुराना आउटपुट फोल्डर मौजूद नहीं है",
+  "old_outputs_cleared": "✅ पुराने आउटपुट सफलतापूर्वक साफ किए गए!",
   "error": "🔥 त्रुटि: {}",
   "ffmpeg_error": "FFmpeg त्रुटि ({}): {}",
-  "file_saved_successfully": "फाइल सफलतापूर्वक संग्रहीत: {}",
-  "total_files_found": "कुल फाइलें मिलीं: {}। सैंपल रेट का उपयोग: {}",
+  "file_saved_successfully": "फाइल सफलतापूर्वक सहेजी गई: {}",
+  "total_files_found": "कुल फा���लें मिलीं: {}। उपयोग की गई सैंपल रेट: {}",
   "total_progress": "कुल प्रगति",
   "detailed_pbar_enabled": "विस्तृत प्रगति बार सक्षम: {}",
-  "loaded_audio": "लोड किया गया ऑडियो: {}, आकार: {}",
-  "cannot_read_track": "ट्रैक पढ़ा नहीं जा सकता: {}",
+  "loaded_audio": "ऑडियो लोड हुआ: {}, आकार: {}",
+  "cannot_read_track": "ट्रैक नहीं पढ़ सकते: {}",
   "error_message": "त्रुटि संदेश: {}",
-  "demudding_track": "ट्रैक को डिमड कर रहा है (फेज रीमिक्स - वाद्य): {}",
+  "demudding_track": "ट्रैक साफ कर रहे हैं (फेज़ रीमिक्स - वाद्य): {}",
   "elapsed_time": "बीता हुआ समय: {:.2f} सेकंड।",
-  "proc_folder_description": "एक निर्दिष्ट मॉडल का उपयोग करके फ़ोल्डर में ऑडियो फाइलों को प्रोसेस करता है।",
+  "proc_folder_description": "निर्दिष्ट मॉडल का उपयोग करके फोल्डर में ऑडियो फाइलें प्रोसेस करता है।",
   "model_type_help": "मॉडल प्रकार (bandit, bs_roformer, mdx23c, आदि)",
   "config_path_help": "कॉन्फ़िग फाइल का पथ",
-  "demud_phaseremix_help": "वाद्य के लिए फेज रीमिक्स सक्षम करें",
-  "start_checkpoint_help": "मान्य वेट्स के लिए प्रारंभिक चेकपॉइंट",
-  "input_folder_help": "प्रोसेस करने के लिए मिश्रणों वाला फ़ोल्डर",
-  "audio_path_help": "प्रोसेस करने के लिए एकल ऑडियो फाइल का पथ",
+  "demud_phaseremix_help": "वाद्य के लिए फेज़ रीमिक्स सक्षम करें",
+  "start_checkpoint_help": "वैध वज़न के लिए प्रारंभिक चेकपॉइंट",
+  "input_folder_help": "प्रक्रिया करने के लिए मिश्रण वाला फोल्डर",
+  "audio_path_help": "एकल ऑडियो फाइल का पथ",
   "store_dir_help": "परिणाम संग्रहीत करने का पथ",
-  "device_ids_help": "GPU IDs की सूची",
-  "extract_instrumental_help": "यदि प्रदान किया गया हो तो स्वर को उलटकर वाद्य प्राप्त करें",
-  "disable_detailed_pbar_help": "डिमिक्स में विस्तृत प्रगति बार अक्षम करें",
-  "force_cpu_help": "CUDA उपलब्ध होने पर भी CPU का उपयोग करें",
+  "device_ids_help": "GPU आईडी की सूची",
+  "extract_instrumental_help": "यदि प्रदान किया गया हो तो वाद्य प्राप्त करने के लिए स्वर को उलटें",
+  "disable_detailed_pbar_help": "डीमिक्स में विस्तृत प्रगति बार अक्षम करें",
+  "force_cpu_help": "CUDA उपलब्ध होने पर भी CPU का उपयोग करने के लिए बाध्य करें",
   "flac_file_help": "WAV के बजाय FLAC फाइल आउटपुट करें",
   "export_format_help": "निर्यात प्रारूप और PCM प्रकार",
   "pcm_type_help": "FLAC फाइलों के लिए PCM प्रकार",
-  "use_tta_help": "टेस्ट टाइम ऑगमेंटेशन सक्षम करें",
-  "lora_checkpoint_help": "LoRA वेट्स के लिए प्रारंभिक चेकपॉइंट",
+  "use_tta_help": "टेस्ट टाइम ऑग्मेंटेशन सक्षम करें",
+  "lora_checkpoint_help": "LoRA वज़न के लिए प्रारंभिक चेकपॉइंट",
   "cuda_available": "CUDA उपलब्ध है, इसे अक्षम करने के लिए --force_cpu का उपयोग करें।",
-  "using_device": "उपयोग किया जा रहा डिवाइस: {}",
-  "instruments_print": "वाद्य: {}",
+  "using_device": "डिवाइस उपयोग: {}",
+  "instruments_print": "वाद्ययंत्र: {}",
   "model_load_time": "मॉडल लोड समय: {:.2f} सेकंड",
   "invalid_url": "❌ अमान्य URL",
-  "cookie_file_updated": "✅ कुकी फाइल अपडेट की गई!",
-  "cookie_installation_error": "⚠️ कुकी स्थापना त्रुटि: {}",
-  "file_size_zero_error": "फाइल आकार शून्य या फाइल नहीं बनाई गई",
+  "cookie_file_updated": "✅ कुकी फाइल अपडेट हुई!",
+  "cookie_installation_error": "⚠️ कुकी इंस्टॉलेशन त्रुटि: {}",
+  "google_drive_mount_error": "❌ Google Drive लिंक त्रुटि: {}",
+  "file_size_zero_error": "फाइल का आकार शून्य है या फाइल नहीं बनाई गई",
   "google_drive_error": "❌ Google Drive डाउनलोड त्रुटि: {}",
   "wav_conversion_failed": "WAV रूपांतरण विफल",
   "download_error": "❌ डाउनलोड त्रुटि: {}",
-  "download_success": "🎉 सफलतापूर्वक डाउनलोड किया गया!",
+  "file_copied_to_drive": "फाइल Google Drive पर कॉपी की गई: {}",
+  "copy_to_drive_error": "❌ Google Drive पर कॉपी करने में त्रुटि: {}",
+  "download_success": "🎉 सफलतापूर्वक डाउनलोड हुआ!",
   "download_failed": "❌ डाउनलोड विफल",
-  "no_output_files_found": "❌ कोई आउटपुट फाइलें नहीं मिलीं।",
-  "output_refreshed_successfully": "✅ आउटपुट सफलतापूर्वक ताज़ा किया गया!",
-  "error_refreshing_output": "❌ आउटपुट ताज़ा करने में त्रुटि: {}",
+  "no_output_files_found": "❌ कोई आउटपुट फाइल नहीं मिली।",
+  "output_refreshed_successfully": "✅ आउटपुट सफलतापूर्वक रिफ्रेश हुआ!",
+  "error_refreshing_output": "❌ आउटपुट रिफ्रेश करने में त्रुटि: {}",
   "starting_audio_separation": "ऑडियो पृथक्करण शुरू हो रहा है...",
-  "processing_audio": "ऑडियो प्रोसेसिंग",
-  "separating_audio": "ऑडियो अलग कर रहा है... ({:.1f}%)",
-  "separation_complete": "पृथक्करण पूरा!",
+  "processing_audio": "ऑडियो प्रोसेस हो रहा है",
+  "separating_audio": "ऑडियो अलग हो रहा है... ({:.1f}%)",
+  "separation_complete": "पृथक्करण पूर्ण!",
   "progress_parsing_error": "प्रगति पार्सिंग त्रुटि: {}",
   "error_occurred": "एक त्रुटि हुई: {}",
-  "separation_process_completed": "पृथक्करण प्रक्रिया पूरी हुई!",
-  "no_audio_file_error": "❌ कोई ऑडियो फाइल प्रदान नहीं की गई और इनपुट डायरेक्टरी में कोई मौजूदा फाइल नहीं है।",
-  "no_input_progress_label": "त्रुटि: कोई इनपुट प्रदान नहीं किया गया -- 0.0%",
-  "processing_audio_print": "ऑडियो प्रोसेसिंग: {} से मॉडल का उपयोग करके: {}",
+  "separation_process_completed": "पृथक्करण प्रक्रिया पूर्ण!",
+  "no_audio_file_error": "❌ कोई ऑडियो फाइल नहीं दी गई और इनपुट डायरेक्टरी में कोई मौजूदा फाइल नहीं है।",
+  "no_input_progress_label": "त्रुटि: कोई इनपुट नहीं दिया गया -- 0.0%",
+  "processing_audio_print": "ऑडियो प्रोसेस हो रहा है: {} मॉडल: {}",
   "starting_audio_separation_progress_label": "ऑडियो पृथक्करण शुरू हो रहा है... -- 0.0%",
-  "separating_audio_progress_label": "ऑडियो अलग कर रहा है... -- {}.0%",
-  "audio_processing_completed_progress_label": "ऑडियो प्रोसेसिंग पूरी हुई! -- 100.0%",
-  "audio_processing_completed": "✅ ऑडियो प्रोसेसिंग पूरी हुई!",
+  "separating_audio_progress_label": "ऑडियो अलग हो रहा है... -- {}.0%",
+  "audio_processing_completed_progress_label": "ऑडियो प्रक्रिया पूर्ण! -- 100.0%",
+  "audio_processing_completed": "✅ ऑडियो प्रक्रिया पूर्ण!",
   "error_occurred_progress_label": "त्रुटि ह���ई -- 0.0%",
   "minimum_files_required": "⚠️ न्यूनतम 2 फाइलें आवश्यक",
-  "valid_files_not_found": "❌ मान्य फाइलें नहीं मिलीं",
-  "starting_ensemble_process": "संयोजन प्रक्रिया शुरू हो रही है...",
-  "ensembling_progress": "संयोजन कर रहा है... ({:.1f}%)",
-  "finalizing_ensemble_output": "संयोजन आउटपुट को अंतिम रूप दे रहा है...",
-  "success_log": "✅ सफलता!\n{}",
+  "valid_files_not_found": "❌ कोई वैध फाइल नहीं मिली",
+  "starting_ensemble_process": "एन्सेम्बल प्रक्रिया शुरू हो रही है...",
+  "ensembling_progress": "एन्सेम्बल हो रहा है... ({:.1f}%)",
+  "finalizing_ensemble_output": "एन्सेम्बल आउटपुट अंतिम रूप दे रहे हैं...",
+  "success_log": "✅ सफल!\n{}",
   "error_log": "❌ त्रुटि!\n{}",
-  "critical_error": "⛔ गंभीर त्रुटि: {}",
-  "ensemble_process_completed": "संयोजन प्रक्रिया पूरी हुई!",
-  "no_models_selected": "❌ कोई मॉडल चयनित नहीं",
-  "no_input_audio_provided": "❌ कोई इनपुट ऑडियो प्रदान नहीं किया गया",
-  "loading_model": "मॉडल लोड कर रहा है {}/{}: {}...",
-  "loading_model_progress_label": "मॉडल लोड कर रहा है {}/{}: {}... -- {}.0%",
-  "completed_model": "मॉडल पूरा हुआ {}/{}: {}",
-  "completed_model_progress_label": "मॉडल पूरा हुआ {}/{}: {} -- {}.0%",
-  "running_command": "कमांड चला रहा है: {}",
+  "ensemble_process_completed": "एन्सेम्बल प्रक्रिया पूर्ण!",
+  "no_models_selected": "❌ कोई मॉडल नहीं चुना गया",
+  "no_input_audio_provided": "❌ कोई इनपुट ऑडियो नहीं दिया गया",
+  "loading_model": "मॉडल लोड हो रहा है {}/{}: {}...",
+  "loading_model_progress_label": "मॉडल लोड हो रहा है {}/{}: {}... -- {}.0%",
+  "completed_model": "मॉडल पूर्ण {}/{}: {}",
+  "completed_model_progress_label": "मॉडल पूर्ण {}/{}: {} -- {}.0%",
+  "running_command": "कमांड चला रहे हैं: {}",
   "model_failed": "मॉडल {} विफल: {}",
   "critical_error_with_model": "{} के साथ गंभीर त्रुटि: {}",
-  "model_output_failed": "{} आउटपुट उत्पन्न करने में विफल",
-  "waiting_for_files": "सभी फाइलों के तैयार होने की प्रतीक्षा में...",
-  "waiting_for_files_progress_label": "सभी फाइलों के तैयार होने की प्रतीक्षा में... -- 90.0%",
-  "performing_ensemble": "संयोजन कर रहा है...",
-  "performing_ensemble_progress_label": "संयोजन कर रहा है... -- 92.0%",
-  "memory_usage_before_ensemble": "संयोजन से पहले मेमोरी उपयोग: {}%",
-  "memory_usage_after_ensemble": "संयोजन के बाद मेमोरी उपयोग: {}%",
-  "finalizing_ensemble_output_progress_label": "संयोजन आउटपुट को अंतिम रूप दे रहा है... -- 98.0%",
-  "ensemble_file_creation_failed": "संयोजन फाइल बनाई नहीं जा सकी: {}",
-  "ensemble_completed_progress_label": "संयोजन सफलतापूर्वक पूरा हुआ! -- 100.0%",
-  "success_output_created": "✅ सफलता! आउटपुट फाइल बनाई गई।",
-  "drive_mounted_copying_ensemble": "Google Drive माउंट किया गया। संयोजन आउटपुट कॉपी हो रहा है...",
-  "drive_already_mounted_copying_ensemble": "Google Drive पहले से माउंट है। संयोजन आउटपुट कॉपी हो रहा है...",
-  "no_ensemble_output_files_found": "❌ कोई संयोजन आउटपुट फाइलें नहीं मिलीं।",
-  "ensemble_output_copied": "✅ संयोजन आउटपुट {} पर कॉपी किया गया",
-  "error_copying_ensemble_output": "❌ संयोजन आउटपुट कॉपी करने में त्रुटि: {}",
-  "drive_mounted_copying_files": "Google Drive माउंट किया गया। फाइलें कॉपी हो ��ही हैं...",
-  "drive_already_mounted_copying_files": "Google Drive पहले से माउंट है। फाइलें कॉपी हो रही हैं...",
-  "files_copied_to_drive": "✅ फाइलें {} पर कॉपी की गईं",
-  "error_copying_files": "❌ फाइलें कॉपी करने में त्रुटि: {}",
-  "mounting_drive": "Google Drive माउंट कर रहा है...",
-  "settings_saved": "सेटिंग्स config.json में संग्रहीत: भाषा={lang}, विधि={method}, पोर्ट={port}",
-  "selected_language": "चयनित भाषा: {lang}",
-  "available_languages": "उपलब्ध भाषाएँ: {langs}",
-  "select_language_prompt": "एक भाषा चुनें (डिफ़ॉल्ट: {default}): ",
-  "invalid_language": "अमान्य भाषा चयन! डिफ़ॉल्ट भाषा का उपयोग कर रहा हूँ ({default})।",
-  "available_sharing_methods": "उपलब्ध साझाकरण विधियाँ: gradio, localtunnel, ngrok",
-  "select_sharing_method_prompt": "एक साझाकरण विधि चुनें (डिफ़ॉल्ट: {default}): ",
-  "enter_ngrok_token_prompt": "अपना Ngrok टोकन दर्ज करें (डिफ़ॉल्ट: {default}): ",
-  "ngrok_token_required": "त्रुटि: ngrok विधि के लिए Ngrok टोकन आवश्यक है!",
-  "enter_port_prompt": "पोर्ट नंबर दर्ज करें (डिफ़ॉल्ट: {default}): ",
-  "opening_previous_url": "पिछला URL खोल रहा है: {url}",
-  "open_url_manually": "Colab में URL मैन्युअल रूप से खोलें: {url}",
-  "invalid_previous_url": "अमान्य पिछला URL: {url}, खोला नहीं गया।",
-  "starting_method": "{method} को पोर्ट {port} पर शुरू कर रहा है...",
-  "process_stopped": "\n🛑 उपयोगकर्ता द्वारा प्रक्रिया रोक दी गई",
-  "selected_port": "चयनित पोर्ट: {port}",
-  "starting_gradio_with_sharing": "Gradio को अंतर्निहित साझाकरण के साथ शुरू कर रहा है...",
-  "starting_localtunnel": "Localtunnel को पोर्ट {port} पर शुरू कर रहा है...",
-  "share_link": "साझा लिंक: {url}",
-  "password_ip": "पासवर्ड IP: {ip}",
-  "starting_ngrok": "Ngrok को पोर्ट {port} पर शुरू कर रहा है...",
-  "ngrok_url": "Ngrok URL: {url}",
-  "ngrok_error": "Ngrok शुरू करने में त्रुटि: {error}",
-  "apollo_chunk_size": "Apollo खंड आकार",
-  "apollo_chunk_size_info": "Apollo सुधार के लिए खंड आकार (यूनिवर्सल मॉडल के लिए अनुशंसित: 19)",
-  "apollo_overlap": "Apollo ओवरलैप",
-  "enhancing_with_apollo": "Apollo के साथ सुधार कर रहा है ({}/{} फाइलें)...",
-  "apollo_processing_completed": "Apollo प्रसंस्करण पूरा!",
-  "apollo_overlap_info": "Apollo सुधार के लिए ओवरलैप (अनुशंसित: 2)",
-  "invalid_method": "त्रुटि: अमान्य विधि! 'gradio', 'localtunnel', या 'ngrok' का उपयोग करें।",
-  "apollo_enhancement_settings": "Apollo सुधार सेटिंग्स",
-  "enhance_with_apollo": "Apollo के साथ सुधार करें",
-  "enhance_with_apollo_info": "पृथक्करण के बाद ऑडियो सुधार के लिए Apollo सक्षम करें",
-  "apollo_method": "Apollo विधि",
-  "apollo_method_info": "Apollo के लिए प्रसंस्करण विधि चुनें",
-  "normal_method": "सामान्य विधि",
-  "mid_side_method": "मध्य/पार्श्व विधि",
-  "apollo_normal_model": "Apollo सामान्य मॉडल",
-  "apollo_normal_model_info": "सामान्य Apollo प्रसंस्करण के लिए उपयोग करने वाला मॉडल",
-  "apollo_midside_model": "Apollo मध्य/पार्श्व मॉडल",
-  "apollo_enhancement_info": "ध्वनि गुणवत्ता सुधार",
-  "selected_models": "चयनित मॉडल",
-  "save_preset": "प्रीसेट सहेजें",
-  "delete_preset": "प्रीसेट हटाएं",
-  "refresh_presets": "प्रीसेट रिफ्रेश करें",
-  "presets_name": "प्रीसेट का नाम",
-  "select_preset": "प्रीसेट चुनें",
-  "add_favorite": "पसंदीदा में जोड़ें",
-  "apollo_midside_model_info": "मध्य/पार्श्व प्रसंस्करण के लिए उपयोग करने वाला मॉडल (वैकल्पिक)",
-  "language_changed_message": "भाषा बदली गई। कृपया सेल को पुनः शुरू करें।"
-  "selected_models": "選択されたモデル",
-  "save_preset": "プリセットを保存",
-  "delete_preset": "プリセットを削除",
-  "refresh_presets": "プリセットを更新",
-  "preset_name": "プリセット名",
-  "select_preset": "プリセットを選択",
-  "add_favorite": "お気に入りに追加",
-  "apply_matchering": "मैचरिंग लागू करें",
-  "matchering_info": "ऑडियो मास्टरिंग और संतुलन को बढ़ाने के लिए मैचरिंग लागू करें",
-  "matchering_passes": "मैचरिंग पास",
-  "matchering_passes_info": "मैचरिंग की पुनरावृत्तियों की संख्या (1-5, उच्च मान गुणवत्ता को बेहतर बनाते हैं लेकिन अधिक समय लेते हैं)",
-  "matchering_processing": "मैचरिंग लागू किया जा रहा है... ({:.1f}%)",
-  "matchering_completed": "✅ मैचरिंग प्रक्रिया पूरी हुई!",
-  "matchering_error": "❌ मैचरिंग के दौरान त्रुटि: {}"
+  "model_output_failed": "{} आउटपुट नहीं दे सका",
+  "waiting_for_files": "सभी फाइलों के तैयार होने का इंतजार...",
+  "waiting_for_files_progress_label": "सभी फाइलों के तैयार होने का इंतजार... -- 90.0%",
+  "performing_ensemble": "एन्सेम्बल कर रहे हैं...",
+  "performing_ensemble_progress_label": "एन्सेम्बल कर रहे हैं... -- 92.0%",
+  "memory_usage_before_ensemble": "एन्सेम्बल से पहले मेमोरी उपयोग: {}%",
+  "memory_usage_after_ensemble": "एन्सेम्बल के बाद मेमोरी उपयोग: {}%",
+  "finalizing_ensemble_output_progress_label": "एन्सेम्बल आउटपुट अंतिम रूप दे रहे हैं... -- 98.0%",
+  "ensemble_file_creation_failed": "एन्सेम्बल फाइल नहीं बनाई जा सकी: {}",
+  "ensemble_completed_progress_label": "एन्सेम्बल सफलतापूर्वक पूर्ण! -- 100.0%",
+  "success_output_created": "✅ सफल! आउटपुट फाइल बनाई गई।",
+  "custom_models_tab": "⚡ कस्टम मॉडल",
+  "add_custom_model": "➕ कस्टम मॉडल जोड़ें",
+  "custom_model_name": "मॉडल नाम",
+  "custom_model_name_placeholder": "अपने मॉडल के लिए एक नाम दर्ज करें",
+  "checkpoint_url": "Checkpoint URL (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "Config URL (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "कस्टम Python फाइल URL (वैकल्पिक)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "मॉडल प्रकार स्वचालित रूप से पहचानें",
+  "model_type": "मॉडल प्रकार",
+  "add_model_btn": "➕ मॉडल जोड़ें",
+  "custom_models_list": "📋 कस्टम मॉडल सूची",
+  "delete_model": "🗑️ हटाएं",
+  "no_custom_models": "अभी तक कोई कस्टम मॉडल नहीं जोड़ा गया",
+  "model_added_success": "✅ मॉडल सफलतापूर्वक जोड़ा गया!",
+  "model_deleted_success": "✅ मॉडल सफलतापूर्वक हटाया गया!",
+  "model_add_error": "❌ मॉडल जोड़ने में त्रुटि: {}",
+  "model_delete_error": "❌ मॉडल हटाने में त्रुटि: {}",
+  "refresh_models": "🔄 रिफ्रेश",
+  "custom_model_info": "डाउनलोड लिंक प्रदान करके कस्टम मॉडल जोड़ें। मॉडल उपयोग होने पर स्वचालित रूप से डाउनलोड होंगे।",
+  "select_model_to_delete": "हटाने के लिए मॉडल चु��ें"
 }

assets/i18n/languages/it_it.json

@@ -17,8 +17,8 @@
   "overlap": "Sovrapposizione",
   "overlap_info": "Consigliato: 2-10 (valori più alti migliorano la qualità ma richiedono più VRAM)",
   "tta_boost": "Boost TTA",
-  "tta_info": "Migliora la qualità ma rallenta la velocità di elaborazione",
   "phase_fix": "Correzione fase",
+  "use_phase_correction": "Applica Correzione Fase Avanzata",
   "phase_fix_info": "Correzione avanzata della fase per tracce strumentali",
   "instrumental": "Strumentale",
   "instrumental_info": "Di solito produce 2 output, ma a volte solo uno, prova! :)",
@@ -80,111 +80,199 @@
   "downloaded_file": "File scaricato",
   "cookie_management": "🍪 Gestione cookie",
   "upload_cookies_txt": "Carica Cookies.txt",
-  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 Perché è necessario?**<br>- Accesso a contenuti con restrizione d'età<br>- Download di video privati/non elencati<br>- Superamento delle restrizioni regionali<br>- Evitare i limiti di download di YouTube<br><br>**⚠️ Note importanti**<br>- NON CONDIVIDERE MAI i tuoi file cookie!<br>- Aggiorna i cookie quando:<br>  • Ricevi errori \"403 Vietato\"<br>  • I download si interrompono improvvisamente<br>  • Vedi il messaggio \"Sessione scaduta\"<br><br>**🔄 Passaggi per l'aggiornamento**<br>1. Installa questa <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">estensione Chrome</a><br>2. Accedi a YouTube su Chrome<br>3. Clicca sull'icona dell'estensione → \"Esporta\"<br>4. Carica il file scaricato qui<br><br>**⏳ Durata dei cookie**<br>- Sessioni normali: 24 ore<br>- Operazioni sensibili: 1 ora<br>- Cambio password: Invalidità immediata</div>",
+  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 Perché è necessario?**<br>- Accesso a contenuti con restrizione d'età<br>- Download di video privati/non elencati<br>- Superamento delle restrizioni regionali<br>- Evitare i limiti di download di YouTube<br><br>**⚠️ Note Importanti**<br>- NON condividere mai i tuoi file cookie!<br>- Aggiorna i cookie quando:<br>  • Ricevi errori \"403 Vietato\"<br>  • I download si fermano improvvisamente<br>  • Vedi messaggi \"Sessione scaduta\"<br><br>**🔄 Passi per il Rinnovo**<br>1. Installa questa <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">estensione Chrome</a><br>2. Accedi a YouTube su Chrome<br>3. Clicca sull'icona dell'estensione → \"Esporta\"<br>4. Carica il file scaricato qui<br><br>**⏳ Durata Cookie**<br>- Sessioni normali: 24 ore<br>- Operazioni sensibili: 1 ora<br>- Cambio password: Rinnovo immediato</div>",
+  "ensemble_output_copied": "✅ Output ensemble copiato in {}",
+  "error_copying_ensemble_output": "❌ Errore copia output ensemble: {}",
+  "drive_mounted_copying_ensemble": "Google Drive montato. Copia output ensemble...",
+  "drive_already_mounted_copying_ensemble": "Google Drive già montato. Copia output ensemble...",
+  "no_ensemble_output_files_found": "❌ Nessun file output ensemble trovato.",
+  "drive_mounted_copying_files": "Google Drive montato. Copia file...",
+  "drive_already_mounted_copying_files": "Google Drive già montato. Copia file...",
+  "files_copied_to_drive": "✅ File copiati in {}",
+  "error_copying_files": "❌ Errore copia file: {}",
+  "mounting_drive": "Montaggio Google Drive...",
+  "settings_saved": "Impostazioni salvate in config.json: Lingua={lang}, Metodo={method}, Porta={port}",
+  "selected_language": "Lingua selezionata: {lang}",
+  "available_languages": "Lingue disponibili: {langs}",
+  "select_language_prompt": "Seleziona una lingua (default: {default}): ",
+  "invalid_language": "Selezione lingua non valida! Uso lingua default ({default}).",
+  "available_sharing_methods": "Metodi di condivisione disponibili: gradio, localtunnel, ngrok",
+  "select_sharing_method_prompt": "Scegli un metodo di condivisione (default: {default}): ",
+  "enter_ngrok_token_prompt": "Inserisci il tuo token Ngrok (default: {default}): ",
+  "ngrok_token_required": "Errore: Token Ngrok richiesto per il metodo ngrok!",
+  "enter_port_prompt": "Inserisci il numero di porta (default: {default}): ",
+  "opening_previous_url": "Apertura URL precedente: {url}",
+  "open_url_manually": "Apri l'URL manualmente in Colab: {url}",
+  "invalid_previous_url": "URL precedente non valido: {url}, non aperto.",
+  "starting_method": "Avvio di {method} sulla porta {port}...",
+  "process_stopped": "\n🛑 Processo fermato dall'utente",
+  "selected_port": "Porta selezionata: {port}",
+  "starting_gradio_with_sharing": "Avvio Gradio con condivisione integrata...",
+  "starting_localtunnel": "Avvio Localtunnel sulla porta {port}...",
+  "share_link": "Link di condivisione: {url}",
+  "password_ip": "Password IP: {ip}",
+  "starting_ngrok": "Avvio Ngrok sulla porta {port}...",
+  "ngrok_url": "URL Ngrok: {url}",
+  "ngrok_error": "Errore avvio Ngrok: {error}",
+  "auto_apollo_chunk_size": "Dimensione frammento Apollo",
+  "auto_apollo_chunk_size_info": "Dimensione frammento per miglioramento Apollo (consigliato: 19 per modello universale)",
+  "auto_apollo_overlap": "Sovrapposizione Apollo",
+  "auto_enhancing_with_apollo": "Miglioramento con Apollo ({}/{} file)...",
+  "auto_apollo_processing_completed": "Elaborazione Apollo completata!",
+  "auto_apollo_overlap_info": "Sovrapposizione per miglioramento Apollo (consigliato: 2)",
+  "apollo_chunk_size": "Dimensione frammento Apollo",
+  "apollo_chunk_size_info": "Dimensione frammento per miglioramento Apollo (consigliato: 19 per modello universale)",
+  "apollo_overlap": "Sovrapposizione Apollo",
+  "enhance_with_apollo": "Migliora con Apollo",
+  "apollo_enhancement_info": "Miglioramento qualità audio",
+  "enhancing_with_apollo": "Miglioramento con Apollo ({}/{} file)...",
+  "apollo_processing_completed": "Elaborazione Apollo completata!",
+  "apollo_overlap_info": "Sovrapposizione per miglioramento Apollo (consigliato: 2)",
+  "selected_models": "Modelli selezionati",
+  "save_preset": "Salva preimpostazione",
+  "delete_preset": "Elimina preimpostazione",
+  "refresh_presets": "Aggiorna preimpostazioni",
+  "preset_name": "Nome preimpostazione",
+  "select_preset": "Seleziona preimpostazione",
+  "add_favorite": "Aggiungi ai preferiti",
+  "apply_matchering": "Applica Matchering",
+  "matchering_info": "Applica Matchering per migliorare il mastering e il bilanciamento dell'audio",
+  "matchering_passes": "Passi Matchering",
+  "matchering_passes_info": "Numero di iterazioni Matchering (1-5, valori più alti migliorano la qualità ma richiedono più tempo)",
+  "matchering_processing": "Applicazione Matchering in corso... ({:.1f}%)",
+  "matchering_completed": "✅ Elaborazione Matchering completata!",
+  "matchering_error": "❌ Errore durante Matchering: {}",
+  "invalid_method": "Errore: Metodo non valido! Usa 'gradio', 'localtunnel' o 'ngrok'.",
+  "language_changed_message": "Lingua cambiata. Riavvia la cella.",
+  "backend_settings": "🚀 Impostazioni backend",
+  "inference_backend": "Backend inferenza",
+  "optimization_mode": "Modalità ottimizzazione",
+  "optimization_mode_info": "Selezione modalità ottimizzazione PyTorch",
+  "mixed_precision_amp": "Precisione mista (AMP)",
+  "mixed_precision_info": "Inferenza 2x più veloce - consigliato",
+  "tf32_acceleration": "Accelerazione TF32",
+  "tf32_acceleration_info": "Boost di velocità extra per RTX 30xx+",
+  "cudnn_benchmark": "Benchmark cuDNN",
+  "cudnn_benchmark_info": "Più lento al primo run, molto più veloce nei run successivi",
+  "ultra_optimized_pytorch": "Backend PyTorch ULTRA-OTTIMIZZATO",
+  "default_active_max_speed": "Attivo per impostazione predefinita - Ottimizzazione velocità massima",
+  "channels_last_mode": "Il più veloce per GPU RTX",
+  "compile_mode": "Velocità extra per PyTorch 2.0+",
+  "default_mode": "Standard",
+  "tta_info": "Aumenta la qualità ma rallenta l'elaborazione",
   "manual_ensemble_tab": "🎚️ Ensemble manuale",
-  "input_sources": "📂 Sorgenti di input",
+  "input_sources": "📂 Sorgenti input",
   "refresh": "🔄 Aggiorna",
   "ensemble_algorithm": "Algoritmo ensemble",
   "select_audio_files": "Seleziona file audio",
   "available_files": "File disponibili",
-  "custom_weights": "Pesi personalizzati (separati da virgole)",
+  "custom_weights": "Pesi personalizzati (separati da virgola)",
   "custom_weights_placeholder": "Esempio: 0.8, 1.2, 1.0, ...",
   "custom_weights_info": "Lascia vuoto per pesi uguali",
   "result_preview_tab": "🎧 Anteprima risultato",
   "ensembled_output": "Output ensemble",
-  "processing_log_tab": "📋 Log di elaborazione",
+  "processing_log_tab": "📋 Registro elaborazione",
   "processing_details": "Dettagli elaborazione",
   "process_ensemble": "⚡ Elabora ensemble",
   "languages_tab": "🌐 Lingue",
   "Vocal Models": "Modelli vocali",
   "Instrumental Models": "Modelli strumentali",
-  "4-Stem Models": "Modelli a 4 stem",
-  "Denoise Models": "Modelli di denoising",
-  "Dereverb Models": "Modelli di rimozione riverbero",
+  "4-Stem Models": "Modelli 4-stem",
+  "Denoise Models": "Modelli denoising",
+  "Dereverb Models": "Modelli dereverb",
   "Other Models": "Altri modelli",
-  "ensemble_files_description": "Combina più file audio in un unico output utilizzando metodi di ensemble specificati.",
+  "Other Shapes": "Altri modelli",
+  "ensemble_files_description": "Combina più file audio in un singolo output usando metodi ensemble specificati.",
   "ensemble_files_help": "Percorso a tutti i file audio per l'ensemble",
-  "ensemble_type_help": "Uno tra: avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
-  "ensemble_weights_help": "Pesi per creare l'ensemble. Il numero di pesi deve corrispondere al numero di file",
-  "ensemble_output_help": "Percorso al file WAV dove verrà salvato il risultato dell'ensemble",
+  "ensemble_type_help": "Uno di: avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
+  "ensemble_weights_help": "Pesi per creare l'ensemble. Il numero di pesi deve essere uguale al numero di file",
+  "ensemble_output_help": "Percorso al file WAV dove verrà memorizzato il risultato dell'ensemble",
   "ensemble_type_print": "Tipo ensemble: {}",
   "num_input_files_print": "Numero file input: {}",
   "weights_print": "Pesi: {}",
   "output_file_print": "File output: {}",
   "duration_mismatch_error": "Tutti i file devono avere la stessa durata",
-  "file_not_found_error": "Errore. File non trovato: {}. Controlla i percorsi.",
-  "reading_chunk_print": "Lettura frammento dal file: {} (inizio: {}s, durata: {}s)",
-  "chunk_result_shape_print": "Forma risultato frammento: {}",
+  "file_not_found_error": "Errore. Impossibile trovare il file: {}. Controlla i percorsi.",
+  "reading_chunk_print": "Lettura chunk dal file: {} (inizio: {}s, durata: {}s)",
+  "chunk_result_shape_print": "Forma risultato chunk: {}",
   "ensemble_completed_print": "Ensemble completato. Output salvato in: {}",
   "file_deletion_error": "{} non può essere eliminato: {}",
   "directory_not_exist_warning": "⚠️ La directory non esiste: {}",
   "not_a_directory_warning": "⚠️ Il percorso non è una directory: {}",
-  "item_deletion_error": "⚠️ Errore durante l'eliminazione di {}: {}",
+  "item_deletion_error": "⚠️ Errore eliminazione {}: {}",
   "old_output_not_exist": "❌ La cartella output vecchia non esiste",
-  "old_outputs_cleared": "✅ Output vecchi eliminati con successo!",
+  "old_outputs_cleared": "✅ Output vecchi cancellati con successo!",
   "error": "🔥 Errore: {}",
   "ffmpeg_error": "Errore FFmpeg ({}): {}",
   "file_saved_successfully": "File salvato con successo: {}",
-  "total_files_found": "Totale file trovati: {}. Frequenza di campionamento usata: {}",
+  "total_files_found": "File totali trovati: {}. Sample rate usato: {}",
   "total_progress": "Progresso totale",
-  "detailed_pbar_enabled": "Barra di progresso dettagliata abilitata: {}",
+  "detailed_pbar_enabled": "Barra progresso dettagliata abilitata: {}",
   "loaded_audio": "Audio caricato: {}, forma: {}",
-  "cannot_read_track": "Impossibile leggere la traccia: {}",
-  "error_message": "Messaggio di errore: {}",
-  "demudding_track": "Pulizia traccia (remix fase - strumentale): {}",
+  "cannot_read_track": "Impossibile leggere traccia: {}",
+  "error_message": "Messaggio errore: {}",
+  "demudding_track": "Pulizia traccia (phase remix - strumentale): {}",
   "elapsed_time": "Tempo trascorso: {:.2f} secondi.",
-  "proc_folder_description": "Elabora i file audio in una cartella usando un modello specificato.",
-  "model_type_help": "Tipo di modello (bandit, bs_roformer, mdx23c, ecc.)",
-  "config_path_help": "Percorso al file di configurazione",
-  "demud_phaseremix_help": "Abilita remix fase per lo strumentale",
-  "start_checkpoint_help": "Punto di controllo iniziale per pesi validi",
-  "input_folder_help": "Cartella contenente i mix da elaborare",
-  "audio_path_help": "Percorso a un singolo file audio da elaborare",
-  "store_dir_help": "Percorso per salvare i risultati",
-  "device_ids_help": "Lista di ID GPU",
-  "extract_instrumental_help": "Inverti le vocali per ottenere lo strumentale, se fornito",
-  "disable_detailed_pbar_help": "Disabilita barra di progresso dettagliata nel demixing",
-  "force_cpu_help": "Forza l'uso della CPU anche se CUDA è disponibile",
-  "flac_file_help": "Genera un file FLAC invece di WAV",
-  "export_format_help": "Formato di esportazione e tipo PCM",
+  "proc_folder_description": "Elabora file audio in una cartella usando un modello specificato.",
+  "model_type_help": "Tipo modello (bandit, bs_roformer, mdx23c, ecc.)",
+  "config_path_help": "Percorso al file configurazione",
+  "demud_phaseremix_help": "Abilita phase remix per strumentale",
+  "start_checkpoint_help": "Checkpoint iniziale per pesi validi",
+  "input_folder_help": "Cartella con mix da elaborare",
+  "audio_path_help": "Percorso a singolo file audio da elaborare",
+  "store_dir_help": "Percorso per memorizzare risultati",
+  "device_ids_help": "Lista ID GPU",
+  "extract_instrumental_help": "Inverti vocali per ottenere strumentale se fornito",
+  "disable_detailed_pbar_help": "Disabilita barra progresso dettagliata in demix",
+  "force_cpu_help": "Forza uso CPU anche se CUDA è disponibile",
+  "flac_file_help": "Output file FLAC invece di WAV",
+  "export_format_help": "Formato export e tipo PCM",
   "pcm_type_help": "Tipo PCM per file FLAC",
-  "use_tta_help": "Abilita l'aumento del tempo di test",
-  "lora_checkpoint_help": "Punto di controllo iniziale per pesi LoRA",
-  "cuda_available": "CUDA disponibile, usa --force_cpu per disabilitarlo.",
-  "using_device": "Dispositivo in uso: {}",
+  "use_tta_help": "Abilita test time augmentation",
+  "lora_checkpoint_help": "Checkpoint iniziale per pesi LoRA",
+  "cuda_available": "CUDA è disponibile, usa --force_cpu per disabilitarlo.",
+  "using_device": "Uso dispositivo: {}",
   "instruments_print": "Strumenti: {}",
-  "model_load_time": "Tempo di caricamento modello: {:.2f} sec",
+  "model_load_time": "Tempo caricamento modello: {:.2f} sec",
   "invalid_url": "❌ URL non valido",
   "cookie_file_updated": "✅ File cookie aggiornato!",
   "cookie_installation_error": "⚠️ Errore installazione cookie: {}",
+  "google_drive_mount_error": "❌ Errore connessione Google Drive: {}",
   "file_size_zero_error": "Dimensione file zero o file non creato",
   "google_drive_error": "❌ Errore download Google Drive: {}",
   "wav_conversion_failed": "Conversione WAV fallita",
   "download_error": "❌ Errore download: {}",
-  "download_success": "🎉 Download completato con successo!",
+  "file_copied_to_drive": "File copiato su Google Drive: {}",
+  "copy_to_drive_error": "❌ Errore copia su Google Drive: {}",
+  "direct_download_failed": "Download diretto file fallito",
+  "direct_download_error": "❌ Errore download diretto: {}",
+  "invalid_google_drive_url": "❌ URL Google Drive non valido",
+  "continuing_without_google_drive": "Connessione Google Drive fallita, continuo senza...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive non montato, salto copia su Drive...",
+  "download_success": "🎉 Scaricato con successo!",
   "download_failed": "❌ Download fallito",
   "no_output_files_found": "❌ Nessun file output trovato.",
   "output_refreshed_successfully": "✅ Output aggiornato con successo!",
-  "error_refreshing_output": "❌ Errore durante aggiornamento output: {}",
-  "starting_audio_separation": "Inizio separazione audio...",
+  "error_refreshing_output": "❌ Errore aggiornamento output: {}",
+  "starting_audio_separation": "Avvio separazione audio...",
   "processing_audio": "Elaborazione audio",
   "separating_audio": "Separazione audio... ({:.1f}%)",
   "separation_complete": "Separazione completata!",
-  "progress_parsing_error": "Errore analisi progresso: {}",
+  "progress_parsing_error": "Errore parsing progresso: {}",
   "error_occurred": "Si è verificato un errore: {}",
-  "separation_process_completed": "Processo di separazione completato!",
-  "no_audio_file_error": "❌ Nessun file audio fornito e nessun file esistente nella directory di input.",
+  "separation_process_completed": "Processo separazione completato!",
+  "no_audio_file_error": "❌ Nessun file audio fornito e nessun file esistente nella directory input.",
   "no_input_progress_label": "Errore: Nessun input fornito -- 0.0%",
-  "processing_audio_print": "Elaborazione audio da: {} usando il modello: {}",
-  "starting_audio_separation_progress_label": "Inizio separazione audio... -- 0.0%",
+  "processing_audio_print": "Elaborazione audio da: {} usando modello: {}",
+  "starting_audio_separation_progress_label": "Avvio separazione audio... -- 0.0%",
   "separating_audio_progress_label": "Separazione audio... -- {}.0%",
   "audio_processing_completed_progress_label": "Elaborazione audio completata! -- 100.0%",
   "audio_processing_completed": "✅ Elaborazione audio completata!",
   "error_occurred_progress_label": "Errore verificato -- 0.0%",
-  "minimum_files_required": "⚠️ Richiesti minimo 2 file",
+  "minimum_files_required": "⚠️ Minimo 2 file richiesti",
   "valid_files_not_found": "❌ Nessun file valido trovato",
-  "starting_ensemble_process": "Inizio processo ensemble...",
-  "ensembling_progress": "Elaborazione ensemble... ({:.1f}%)",
+  "starting_ensemble_process": "Avvio processo ensemble...",
+  "ensembling_progress": "Ensemble in corso... ({:.1f}%)",
   "finalizing_ensemble_output": "Finalizzazione output ensemble...",
   "success_log": "✅ Successo!\n{}",
   "error_log": "❌ Errore!\n{}",
@@ -200,81 +288,92 @@
   "model_failed": "Modello {} fallito: {}",
   "critical_error_with_model": "Errore critico con {}: {}",
   "model_output_failed": "{} non è riuscito a produrre output",
-  "waiting_for_files": "In attesa che tutti i file siano pronti...",
-  "waiting_for_files_progress_label": "In attesa che tutti i file siano pronti... -- 90.0%",
+  "waiting_for_files": "Attesa che tutti i file siano pronti...",
+  "waiting_for_files_progress_label": "Attesa che tutti i file siano pronti... -- 90.0%",
   "performing_ensemble": "Esecuzione ensemble...",
   "performing_ensemble_progress_label": "Esecuzione ensemble... -- 92.0%",
-  "memory_usage_before_ensemble": "Uso memoria prima dell'ensemble: {}%",
-  "memory_usage_after_ensemble": "Uso memoria dopo l'ensemble: {}%",
+  "memory_usage_before_ensemble": "Uso memoria prima ensemble: {}%",
+  "memory_usage_after_ensemble": "Uso memoria dopo ensemble: {}%",
   "finalizing_ensemble_output_progress_label": "Finalizzazione output ensemble... -- 98.0%",
-  "ensemble_file_creation_failed": "Creazione file ensemble fallita: {}",
+  "ensemble_file_creation_failed": "Impossibile creare file ensemble: {}",
   "ensemble_completed_progress_label": "Ensemble completato con successo! -- 100.0%",
   "success_output_created": "✅ Successo! File output creato.",
-  "drive_mounted_copying_ensemble": "Google Drive montato. Copia output ensemble...",
-  "drive_already_mounted_copying_ensemble": "Google Drive già montato. Copia output ensemble...",
-  "no_ensemble_output_files_found": "❌ Nessun file output ensemble trovato.",
-  "ensemble_output_copied": "✅ Output ensemble copiato in {}",
-  "error_copying_ensemble_output": "❌ Errore durante copia output ensemble: {}",
-  "drive_mounted_copying_files": "Google Drive montato. Copia file...",
-  "drive_already_mounted_copying_files": "Google Drive già montato. Copia file...",
-  "files_copied_to_drive": "✅ File copiati in {}",
-  "error_copying_files": "❌ Errore durante copia file: {}",
-  "mounting_drive": "Montaggio Google Drive...",
-  "settings_saved": "Impostazioni salvate in config.json: Lingua={lang}, Metodo={method}, Porta={port}",
-  "selected_language": "Lingua selezionata: {lang}",
-  "available_languages": "Lingue disponibili: {langs}",
-  "select_language_prompt": "Seleziona una lingua (default: {default}): ",
-  "invalid_language": "Selezione lingua non valida! Uso lingua default ({default}).",
-  "available_sharing_methods": "Metodi di condivisione disponibili: gradio, localtunnel, ngrok",
-  "select_sharing_method_prompt": "Scegli un metodo di condivisione (default: {default}): ",
-  "enter_ngrok_token_prompt": "Inserisci il tuo token Ngrok (default: {default}): ",
-  "ngrok_token_required": "Errore: Token Ngrok richiesto per il metodo ngrok!",
-  "enter_port_prompt": "Inserisci il numero di porta (default: {default}): ",
-  "opening_previous_url": "Apertura URL precedente: {url}",
-  "open_url_manually": "Apri l'URL manualmente in Colab: {url}",
-  "invalid_previous_url": "URL precedente non valido: {url}, non aperto.",
-  "starting_method": "Avvio di {method} sulla porta {port}...",
-  "process_stopped": "\n🛑 Processo fermato dall'utente",
-  "selected_port": "Porta selezionata: {port}",
-  "starting_gradio_with_sharing": "Avvio Gradio con condivisione integrata...",
-  "starting_localtunnel": "Avvio Localtunnel sulla porta {port}...",
-  "share_link": "Link di condivisione: {url}",
-  "password_ip": "Password IP: {ip}",
-  "starting_ngrok": "Avvio Ngrok sulla porta {port}...",
-  "ngrok_url": "URL Ngrok: {url}",
-  "ngrok_error": "Errore avvio Ngrok: {error}",
-  "apollo_chunk_size": "Dimensione frammento Apollo",
-  "apollo_chunk_size_info": "Dimensione frammento per miglioramento Apollo (consigliato: 19 per modello universale)",
-  "apollo_overlap": "Sovrapposizione Apollo",
-  "enhancing_with_apollo": "Miglioramento con Apollo ({}/{} file)...",
-  "apollo_processing_completed": "Elaborazione Apollo completata!",
-  "apollo_overlap_info": "Sovrapposizione per miglioramento Apollo (consigliato: 2)",
-  "invalid_method": "Errore: Metodo non valido! Usa 'gradio', 'localtunnel' o 'ngrok'.",
-  "apollo_enhancement_settings": "Impostazioni miglioramento Apollo",
-  "enhance_with_apollo": "Migliora con Apollo",
-  "enhance_with_apollo_info": "Abilita Apollo per migliorare l'audio dopo la separazione",
-  "apollo_method": "Metodo Apollo",
-  "apollo_method_info": "Seleziona il metodo di elaborazione per Apollo",
+  "phase_fixer_tab": "🔧 Correttore di fase",
+  "batch_processing_tab": "📦 Elaborazione batch",
+  "batch_description": "Elabora più file contemporaneamente con lo stesso modello",
+  "batch_input_folder": "Cartella di input",
+  "batch_input_folder_placeholder": "Inserisci il percorso della cartella contenente i file audio",
+  "batch_output_folder": "Cartella di output",
+  "batch_output_folder_placeholder": "Inserisci il percorso della cartella di output",
+  "batch_file_list": "File da elaborare",
+  "batch_add_files": "Aggiungi file",
+  "batch_clear_list": "Svuota lista",
+  "batch_start": "🚀 Avvia elaborazione batch",
+  "batch_stop": "⏹️ Ferma",
+  "batch_progress": "Progresso batch",
+  "batch_current_file": "In elaborazione",
+  "batch_completed": "✅ Elaborazione batch completata!",
+  "batch_stopped": "⏹️ Elaborazione batch fermata",
+  "batch_no_files": "❌ Nessun file selezionato per l'elaborazione",
+  "source_model": "Modello sorgente (Vocali)",
+  "source_model_info": "Modello per estrarre vocali/dati di fase",
+  "target_model": "Modello destinazione (Strumentale)",
+  "target_model_info": "Modello a cui applicare la correzione di fase",
+  "custom_models_tab": "⚡ Modelli personalizzati",
+  "Custom Models": "Modelli personalizzati",
+  "add_custom_model": "➕ Aggiungi modello personalizzato",
+  "custom_model_name": "Nome modello",
+  "custom_model_name_placeholder": "Inserisci un nome per il tuo modello",
+  "checkpoint_url": "URL Checkpoint (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "URL Configurazione (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "URL file Python personalizzato (opzionale)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "Rileva automaticamente il tipo di modello",
+  "model_type": "Tipo modello",
+  "add_model_btn": "➕ Aggiungi modello",
+  "custom_models_list": "📋 Lista modelli personalizzati",
+  "delete_model": "🗑️ Elimina",
+  "no_custom_models": "Nessun modello personalizzato aggiunto",
+  "model_added_success": "✅ Modello aggiunto con successo!",
+  "model_deleted_success": "✅ Modello eliminato con successo!",
+  "model_add_error": "❌ Errore aggiunta modello: {}",
+  "model_delete_error": "❌ Errore eliminazione modello: {}",
+  "refresh_models": "🔄 Aggiorna",
+  "custom_model_info": "Aggiungi modelli personalizzati fornendo link di download. I modelli verranno scaricati automaticamente quando usati.",
+  "select_model_to_delete": "Seleziona un modello da eliminare",
+  "downloading_model_file": "Download modello: {}",
+  "downloading_file_progress": "Download: {} - {}%",
+  "chunk_size_mode": "Modalità Dimensione Frammento",
+  "chunk_size_mode_info": "base: preset standard | custom: inserisci il tuo valore | yaml: leggi dallo YAML del modello",
+  "chunk_size_custom_label": "Dimensione Frammento Personalizzata",
+  "chunk_size_custom_info": "Inserisci un valore personalizzato per la dimensione del frammento",
+  "chunk_size_yaml_label": "Dimensione Frammento YAML",
+  "chunk_size_yaml_display_info": "Verrà letta dal file di configurazione YAML del modello dopo il download",
+  "chunk_size_yaml_not_downloaded": "⏳ YAML del modello non ancora scaricato — il valore verrà letto durante l'elaborazione",
+  "chunk_size_yaml_detected": "✅ Da YAML: {}",
+  "remove_favorite": "Rimuovi dai Preferiti",
   "normal_method": "Metodo normale",
-  "mid_side_method": "Metodo mid/side",
-  "apollo_normal_model": "Modello Apollo normale",
-  "apollo_normal_model_info": "Modello da usare per l'elaborazione Apollo normale",
-  "apollo_midside_model": "Modello Apollo mid/side",
-  "apollo_enhancement_info": "Miglioramento qualità audio",
-  "selected_models": "Modelli Selezionati",
-  "save_preset": "Salva Preimpostazione",
-  "delete_preset": "Elimina Preimpostazione",
-  "refresh_presets": "Aggiorna Preimpostazioni",
-  "preset_name": "Nome Preimpostazione",
-  "select_preset": "Seleziona Preimpostazione",
-  "add_favorite": "Aggiungi ai Preferiti",
-  "apollo_midside_model_info": "Modello da usare per l'elaborazione mid/side (opzionale)",
-  "language_changed_message": "Lingua cambiata. Riavvia la cella.",
-  "apply_matchering": "Applica Matchering",
-  "matchering_info": "Applica Matchering per migliorare il mastering e il bilanciamento dell'audio",
-  "matchering_passes": "Passaggi di Matchering",
-  "matchering_passes_info": "Numero di iterazioni di Matchering (1-5, valori più alti migliorano la qualità ma richiedono più tempo)",
-  "matchering_processing": "Applicazione di Matchering in corso... ({:.1f}%)",
-  "matchering_completed": "✅ Elaborazione Matchering completata!",
-  "matchering_error": "❌ Errore durante Matchering: {}"
+  "mid_side_method": "Metodo Mid-Side",
+  "apollo_processing_method": "Metodo di Elaborazione Apollo",
+  "apollo_normal_model": "Modello Normale Apollo",
+  "apollo_mid_side_model": "Modello Mid-Side Apollo",
+  "scale_factor": "Fattore di Scala",
+  "scale_factor_info": "Fattore di scala della fase (consigliato: 1.4)",
+  "source_file_label": "File Sorgente (Voce)",
+  "target_file_label": "File di Destinazione (Strumentale)",
+  "run_phase_fixer": "🔧 Avvia Correttore di Fase",
+  "phase_fixed_output": "Output con Fase Corretta",
+  "phase_fixer_settings": "Impostazioni Correttore di Fase",
+  "low_cutoff": "Taglio Basso (Hz)",
+  "high_cutoff": "Taglio Alto (Hz)",
+  "low_cutoff_info": "Frequenza di taglio bassa per la correzione della fase",
+  "high_cutoff_info": "Frequenza di taglio alta per la correzione della fase",
+  "no_preset_name_provided": "⚠️ Nessun nome preset fornito",
+  "no_models_selected_for_preset": "⚠️ Nessun modello selezionato per il preset",
+  "preset_saved": "✅ Preset '{}' salvato!",
+  "please_upload_both_files": "⚠️ Carica sia il file sorgente che quello di destinazione",
+  "processing_log_placeholder": "Il registro di elaborazione apparirà qui...",
+  "enter_preset_name": "Inserisci il nome del preset..."
 }

assets/i18n/languages/ja_jp.json

@@ -17,11 +17,10 @@
   "overlap": "オーバーラップ",
   "overlap_info": "推奨: 2-10 (高い値は品質を向上させますが、より多くのVRAMが必要です)",
   "tta_boost": "TTAブースト",
-  "tta_info": "品質を向上させますが、処理速度が遅くなります",
   "phase_fix": "位相補正",
+  "use_phase_correction": "高度な位相補正を適用",
   "phase_fix_info": "インストゥルメンタルトラックのための高度な位相補正",
   "instrumental": "インストゥルメンタル",
-  "instrumental_info": "通常2つの出力が得られますが、時には1つの出力だけの場合もあります。試してみてください :)",
   "process": "🚀 処理",
   "reset": "🧹 リセット",
   "status": "ステータス",
@@ -80,201 +79,267 @@
   "downloaded_file": "ダウンロードされたファイル",
   "cookie_management": "🍪 クッキー管理",
   "upload_cookies_txt": "Cookies.txtをアップロード",
-  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 なぜ必要か?**<br>- 年齢制限付きコンテンツへのアクセス<br>- プライベート/非公開ビデオのダウンロード<br>- 地域制限の回避<br>- YouTubeのダウンロード制限の回避<br><br>**⚠️ 重要な注意**<br>- クッキーファイルを決して共有しないでください!<br>- クッキーを更新するタイミング:<br>  • \"403 Forbidden\" エラーが出た場合<br>  • ダウンロードが突然停止した場合<br>  • \"セッションが期限切れ\" メッセージが表示された場合<br><br>**🔄 更新手順**<br>1. この <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">Chrome拡張機能</a> をインストール<br>2. ChromeでYouTubeにログイン<br>3. 拡張機能アイコンをクリック → \"エクスポート\"<br>4. ダウンロードしたファイルをここにアップロード<br><br>**⏳ クッキーの寿命**<br>- 通常セッション: 24時間<br>- 機密操作: 1時間<br>- パスワード変更: 即時無効化</div>",
+  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 なぜ必要？**<br>- 年齢制限コンテンツへのアクセス<br>- プライベート/非公開ビデオのダウンロード<br>- 地域制限の回避<br>- YouTubeダウンロード制限の回避<br><br>**⚠️ 重要な注意**<br>- クッキーファイルを共有しないでください！<br>- 以下の時に更新:<br>  • \"403 Forbidden\"エラー<br>  • ダウンロードが突然停止<br>  • \"セッション期限切れ\"メッセージ<br><br>**🔄 更新手順**<br>1. この<a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">Chrome拡張</a>をインストール<br>2. ChromeでYouTubeにログイン<br>3. 拡張アイコンをクリック → \"エクスポート\"<br>4. ここにダウンロードしたファイルをアップロード<br><br>**⏳ クッキー寿命**<br>- 通常セッション: 24時間<br>- 敏感な操作: 1時間<br>- パスワード変更: 即時更新</div>",
+  "ensemble_output_copied": "✅ アンサンブル出力が {} にコピーされました",
+  "error_copying_ensemble_output": "❌ アンサンブル出力のコピーエラー: {}",
+  "drive_mounted_copying_ensemble": "Google Driveがマウントされました。アンサンブル出力をコピー中...",
+  "drive_already_mounted_copying_ensemble": "Google Driveはすでにマウントされています。アンサンブル出力をコピー中...",
+  "no_ensemble_output_files_found": "❌ アンサンブル出力ファイルが見つかりません。",
+  "drive_mounted_copying_files": "Google Driveがマウントされました。ファイルをコピー中...",
+  "drive_already_mounted_copying_files": "Google Driveはすでにマウントされています。ファイルをコピー中...",
+  "files_copied_to_drive": "✅ ファイルが {} にコピーされました",
+  "error_copying_files": "❌ ファイルのコピーエラー: {}",
+  "mounting_drive": "Google Driveをマウント中...",
+  "settings_saved": "設定が config.json に保存されました: 言語={lang}, メソッド={method}, ポート={port}",
+  "selected_language": "選択された言語: {lang}",
+  "available_languages": "利用可能な言語: {langs}",
+  "select_language_prompt": "言語を選択 (デフォルト: {default}): ",
+  "invalid_language": "無効な言語選択! デフォルト言語 ({default}) を使用します。",
+  "available_sharing_methods": "利用可能な共有方法: gradio, localtunnel, ngrok",
+  "select_sharing_method_prompt": "共有方法を選択 (デフォルト: {default}): ",
+  "enter_ngrok_token_prompt": "Ngrokトークンを入力 (デフォルト: {default}): ",
+  "ngrok_token_required": "エラー: ngrokメソッドにはNgrokトークンが必要です!",
+  "enter_port_prompt": "ポート���号を入力 (デフォルト: {default}): ",
+  "opening_previous_url": "前のURLを開く: {url}",
+  "open_url_manually": "ColabでURLを手動で開く: {url}",
+  "invalid_previous_url": "無効な前のURL: {url}, 開かれませんでした。",
+  "starting_method": "{method} をポート {port} で開始...",
+  "process_stopped": "\n🛑 ユーザーによってプロセスが停止されました",
+  "selected_port": "選択されたポート: {port}",
+  "starting_gradio_with_sharing": "Gradioを共有機能付きで開始...",
+  "starting_localtunnel": "Localtunnelをポート {port} で開始...",
+  "share_link": "共有リンク: {url}",
+  "password_ip": "パスワードIP: {ip}",
+  "starting_ngrok": "Ngrokをポート {port} で開始...",
+  "ngrok_url": "Ngrok URL: {url}",
+  "ngrok_error": "Ngrok開始エラー: {error}",
+  "auto_apollo_chunk_size": "Apolloチャンクサイズ",
+  "auto_apollo_chunk_size_info": "Apollo改善のためのチャンクサイズ (ユニバーサルモデル推奨: 19)",
+  "auto_apollo_overlap": "Apolloオーバーラップ",
+  "auto_enhancing_with_apollo": "Apolloで改善中 ({}/{} ファイル)...",
+  "auto_apollo_processing_completed": "Apollo処理完了!",
+  "auto_apollo_overlap_info": "Apollo改善のためのオーバーラップ (推奨: 2)",
+  "apollo_chunk_size": "Apolloチャンクサイズ",
+  "apollo_chunk_size_info": "Apollo改善のためのチャンクサイズ (ユニバーサルモデル推奨: 19)",
+  "apollo_overlap": "Apolloオーバーラップ",
+  "enhance_with_apollo": "Apolloで改善",
+  "apollo_enhancement_info": "音質の改善",
+  "enhancing_with_apollo": "Apolloで改善中 ({}/{} ファイル)...",
+  "apollo_processing_completed": "Apollo処理完了!",
+  "apollo_overlap_info": "Apollo改善のためのオーバーラップ (推奨: 2)",
+  "selected_models": "選択されたモデル",
+  "save_preset": "プリセットを保存",
+  "delete_preset": "プリセットを削除",
+  "refresh_presets": "プリセットを更新",
+  "preset_name": "プリセット名",
+  "select_preset": "プリセットを選択",
+  "add_favorite": "お気に入りに追加",
+  "apply_matchering": "Matcheringを適用",
+  "matchering_info": "オーディオのマスタリングとバランスを向上させるためにMatcheringを適用",
+  "matchering_passes": "Matcheringの反復回数",
+  "matchering_passes_info": "Matcheringの反復回数（1-5、高い値は品質を向上させますが時間がかかります）",
+  "matchering_processing": "Matcheringを適用中... ({:.1f}%)",
+  "matchering_completed": "✅ Matchering処理が完了しました！",
+  "matchering_error": "❌ Matchering中にエラーが発生しました：{}",
+  "invalid_method": "エラー: 無効なメソッド! 'gradio', 'localtunnel', または 'ngrok' を使用してください。",
+  "language_changed_message": "言語が変更されました。セルを再起動してください。",
+  "backend_settings": "🚀 バックエンド設定",
+  "inference_backend": "推論バックエンド",
+  "optimization_mode": "最適化モード",
+  "optimization_mode_info": "PyTorch最適化モードの選択",
+  "mixed_precision_amp": "混合精度 (AMP)",
+  "mixed_precision_info": "2x速い推論 - 推奨",
+  "tf32_acceleration": "TF32アクセラレーション",
+  "tf32_acceleration_info": "RTX 30xx+のための追加速度ブースト",
+  "cudnn_benchmark": "cuDNNベンチマーク",
+  "cudnn_benchmark_info": "最初の実行で遅い、後続の実行で非常に速い",
+  "ultra_optimized_pytorch": "超最適化PyTorchバックエンド",
+  "default_active_max_speed": "デフォルトでアクティブ - 最大速度最適化",
+  "channels_last_mode": "RTX GPUで最速",
+  "compile_mode": "PyTorch 2.0+のための追加速度",
+  "default_mode": "標準",
+  "tta_info": "品質を向上させますが、処理速度が遅くなります",
+  "instrumental_info": "通常2つのステムを出力しますが、時には1つだけ - 試してみてください！ :)",
+  "direct_download_error": "❌ 直接ダウンロードエラー：{}",
+  "invalid_google_drive_url": "❌ 無効な Google Drive URL",
+  "continuing_without_google_drive": "Google Drive接続に失敗しました。続行します...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Driveがマウントされていません、Driveへのコピーをスキップします...",
+  "direct_download_failed": "ファイルの直接ダウンロードに失敗しました",
   "manual_ensemble_tab": "🎚️ 手動アンサンブル",
+  "phase_fixer_tab": "🔧 位相補正ツール",
+  "batch_processing_tab": "📦 バッチ処理",
+  "batch_description": "同じモデルで複数のファイルを一度に処理",
+  "batch_input_folder": "入力フォルダ",
+  "batch_input_folder_placeholder": "オーディオファイルを含むフォルダのパスを入力",
+  "batch_output_folder": "出力フォルダ",
+  "batch_output_folder_placeholder": "出力ファイル��フォルダのパスを入力",
+  "batch_file_list": "処理するファイル",
+  "batch_add_files": "ファイルを追加",
+  "batch_clear_list": "リストをクリア",
+  "batch_start": "🚀 バッチ処理を開始",
+  "batch_stop": "⏹️ 停止",
+  "batch_progress": "バッチ処理の進捗",
+  "batch_current_file": "現在処理中",
+  "batch_completed": "✅ バッチ処理が完了しました！",
+  "batch_stopped": "⏹️ バッチ処理が停止されました",
+  "batch_no_files": "❌ 処理するファイルが選択されていません",
+  "source_model": "ソースモデル（ボーカル）",
+  "source_model_info": "ボーカル/位相データを抽出するモデル",
+  "target_model": "ターゲットモデル（インストゥルメンタル）",
+  "target_model_info": "位相補正を適用するモデル",
   "input_sources": "📂 入力ソース",
   "refresh": "🔄 更新",
   "ensemble_algorithm": "アンサンブルアルゴリズム",
   "select_audio_files": "オーディオファイルを選択",
   "available_files": "利用可能なファイル",
-  "custom_weights": "カスタムウェイト (カンマで区切り)",
+  "custom_weights": "カスタムウェイト（カンマ区切り）",
   "custom_weights_placeholder": "例: 0.8, 1.2, 1.0, ...",
-  "custom_weights_info": "均等なウェイトの場合は空白のまま",
+  "custom_weights_info": "等しいウェイトにするには空白のままにしてください",
   "result_preview_tab": "🎧 結果プレビュー",
   "ensembled_output": "アンサンブル出力",
   "processing_log_tab": "📋 処理ログ",
-  "processing_details": "処理詳細",
+  "processing_details": "処理の詳細",
   "process_ensemble": "⚡ アンサンブルを処理",
   "languages_tab": "🌐 言語",
   "Vocal Models": "ボーカルモデル",
   "Instrumental Models": "インストゥルメンタルモデル",
   "4-Stem Models": "4ステムモデル",
   "Denoise Models": "ノイズ除去モデル",
-  "Dereverb Models": "リバーブ除去モデル",
-  "Other Models": "その他のモデル",
-  "ensemble_files_description": "指定されたアンサンブル手法を使用して複数のオーディオファイルを1つの出力に統合します。",
+  "Dereverb Models": "残響除去モデル",
+  "Other Shapes": "その他のモデル",
+  "Custom Models": "カスタムモデル",
+  "ensemble_files_description": "指定されたアンサンブル方法を使用して複数のオーディオファイルを1つの出力に結合します。",
   "ensemble_files_help": "アンサンブル用のすべてのオーディオファイルへのパス",
-  "ensemble_type_help": "avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft のいずれか",
-  "ensemble_weights_help": "アンサンブルを作成するためのウェイト。ウェイト数はファイル数と一致する必要があります",
-  "ensemble_output_help": "アンサンブル結果を保存するWAVファイルへのパス",
+  "ensemble_type_help": "次のいずれか: avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
+  "ensemble_weights_help": "アンサンブル作成のウェイト。ウェイトの数はファイルの数と同じである必要があります",
+  "ensemble_output_help": "アンサンブル結果を保存するWAVファイルのパス",
   "ensemble_type_print": "アンサンブルタイプ: {}",
   "num_input_files_print": "入力ファイル数: {}",
   "weights_print": "ウェイト: {}",
   "output_file_print": "出力ファイル: {}",
-  "duration_mismatch_error": "すべてのファイルは同じ長さでなければなりません",
+  "duration_mismatch_error": "すべてのファイルの長さが同じである必要があります",
   "file_not_found_error": "エラー。ファイルが見つかりません: {}。パスを確認してください。",
-  "reading_chunk_print": "ファイルからチャンクを読み込み: {} (開始: {}秒, 長さ: {}秒)",
+  "reading_chunk_print": "ファイルからチャンクを読み込み中: {} (開始: {}s, 長さ: {}s)",
   "chunk_result_shape_print": "チャンク結果の形状: {}",
-  "ensemble_completed_print": "アンサンブル完了。出力は以下に保存: {}",
+  "ensemble_completed_print": "アンサンブル完了。出力を保存: {}",
   "file_deletion_error": "{} を削除できませんでした: {}",
   "directory_not_exist_warning": "⚠️ ディレクトリが存在しません: {}",
-  "not_a_directory_warning": "⚠️ パスはディレクトリではありません: {}",
+  "not_a_directory_warning": "⚠️ パスがディレクトリではありません: {}",
   "item_deletion_error": "⚠️ {} の削除エラー: {}",
+  "critical_error": "⛔ 致命的なエラー: {}",
   "old_output_not_exist": "❌ 古い出力フォルダが存在しません",
-  "old_outputs_cleared": "✅ 古い出力が正常に削除されました!",
+  "old_outputs_cleared": "✅ 古い出��が正常に削除されました！",
   "error": "🔥 エラー: {}",
   "ffmpeg_error": "FFmpegエラー ({}): {}",
   "file_saved_successfully": "ファイルが正常に保存されました: {}",
   "total_files_found": "見つかったファイルの総数: {}。使用サンプルレート: {}",
-  "total_progress": "総進捗",
-  "detailed_pbar_enabled": "詳細プログレスバー有効: {}",
-  "loaded_audio": "ロードされたオーディオ: {}, 形状: {}",
+  "total_progress": "全体の進捗",
+  "detailed_pbar_enabled": "詳細プログレスバーが有効: {}",
+  "loaded_audio": "オーディオを読み込みました: {}、形状: {}",
   "cannot_read_track": "トラックを読み込めません: {}",
   "error_message": "エラーメッセージ: {}",
-  "demudding_track": "トラックのデマッディング (位相リミックス - インストゥルメンタル): {}",
-  "elapsed_time": "経過時間: {:.2f} 秒。",
+  "demudding_track": "トラックをクリーニング中（位相リミックス - インストゥルメンタル）: {}",
+  "elapsed_time": "経過時間: {:.2f}秒。",
   "proc_folder_description": "指定されたモデルを使用してフォルダ内のオーディオファイルを処理します。",
-  "model_type_help": "モデルタイプ (bandit, bs_roformer, mdx23c など)",
+  "model_type_help": "モデルタイプ (bandit, bs_roformer, mdx23cなど)",
   "config_path_help": "設定ファイルへのパス",
-  "demud_phaseremix_help": "インストゥルメンタルの位相リミックスを有効化",
+  "demud_phaseremix_help": "インストゥルメンタルの位相リミックスを有効にする",
   "start_checkpoint_help": "有効なウェイトの初期チェックポイント",
-  "input_folder_help": "処理するミックスを含むフォルダ",
-  "audio_path_help": "処理する単一のオーディオファイルへのパス",
+  "input_folder_help": "処理するミックスが入ったフォルダ",
+  "audio_path_help": "処理する単一オーディオファイルのパス",
   "store_dir_help": "結果を保存するパス",
   "device_ids_help": "GPU IDのリスト",
   "extract_instrumental_help": "提供されている場合、ボーカルを反転してインストゥルメンタルを取得",
-  "disable_detailed_pbar_help": "デミックスで詳細プログレスバーを無効化",
-  "force_cpu_help": "CUDAが利用可能でもCPUを強制使用",
+  "disable_detailed_pbar_help": "デミックスの詳細プログレスバーを無効にする",
+  "force_cpu_help": "CUDAが利用可能でもCPUを強制使用する",
   "flac_file_help": "WAVの代わりにFLACファイルを出力",
-  "export_format_help": "エクスポートフォーマットとPCMタイプ",
+  "export_format_help": "エクスポート形式とPCMタイプ",
   "pcm_type_help": "FLACファイルのPCMタイプ",
-  "use_tta_help": "テスト時間増強を有効化",
+  "use_tta_help": "テスト時間拡張を有効にする",
   "lora_checkpoint_help": "LoRAウェイトの初期チェックポイント",
-  "cuda_available": "CUDAが利用可能です。無効化するには --force_cpu を使用してください。",
+  "cuda_available": "CUDAが利用可能です。無効にするには --force_cpu を使用してください。",
   "using_device": "使用デバイス: {}",
   "instruments_print": "楽器: {}",
-  "model_load_time": "モデルロード時間: {:.2f} 秒",
+  "model_load_time": "モデル読み込み時間: {:.2f}秒",
   "invalid_url": "❌ 無効なURL",
-  "cookie_file_updated": "✅ クッキーファイルが更新されました!",
-  "cookie_installation_error": "⚠️ クッキーインストールエラー: {}",
+  "cookie_file_updated": "✅ Cookieファイルが更新されました！",
+  "cookie_installation_error": "⚠️ Cookieインストールエラー: {}",
+  "google_drive_mount_error": "❌ Google Drive接続エラー: {}",
   "file_size_zero_error": "ファイルサイズがゼロまたはファイルが作成されていません",
   "google_drive_error": "❌ Google Driveダウンロードエラー: {}",
-  "wav_conversion_failed": "WAV変換に失敗",
+  "wav_conversion_failed": "WAV変換に失敗しました",
   "download_error": "❌ ダウンロードエラー: {}",
-  "download_success": "🎉 ダウンロード成功!",
-  "download_failed": "❌ ダウンロード失敗",
+  "file_copied_to_drive": "ファイルがGoogle Driveにコピーされました: {}",
+  "copy_to_drive_error": "❌ Google Driveへのコピーエラー: {}",
+  "download_success": "🎉 ダウンロードに成功しました！",
+  "download_failed": "❌ ダウンロードに失敗しました",
   "no_output_files_found": "❌ 出力ファイルが見つかりません。",
-  "output_refreshed_successfully": "✅ 出力が正常に更新されました!",
+  "output_refreshed_successfully": "✅ 出力が正常に更新されました！",
   "error_refreshing_output": "❌ 出力更新エラー: {}",
-  "starting_audio_separation": "オーディオ分離を開始...",
-  "processing_audio": "オーディオ処理中",
-  "separating_audio": "オーディオ分離中... ({:.1f}%)",
-  "separation_complete": "分離完了!",
+  "starting_audio_separation": "オーディオ分離を開始しています...",
+  "processing_audio": "オーディオを処理中",
+  "separating_audio": "オーディオを分離中... ({:.1f}%)",
+  "separation_complete": "分離完了！",
   "progress_parsing_error": "進捗解析エラー: {}",
   "error_occurred": "エラーが発生しました: {}",
-  "separation_process_completed": "分離プロセス完了!",
+  "separation_process_completed": "分離プロセスが完了しました！",
   "no_audio_file_error": "❌ オーディオファイルが提供されておらず、入力ディレクトリに既存のファイルもありません。",
   "no_input_progress_label": "エラー: 入力が提供されていません -- 0.0%",
-  "processing_audio_print": "{} からオーディオを処理中、モデル使用: {}",
-  "starting_audio_separation_progress_label": "オーディオ分離を開始... -- 0.0%",
-  "separating_audio_progress_label": "オーディオ分離中... -- {}.0%",
-  "audio_processing_completed_progress_label": "オーディオ処理完了! -- 100.0%",
-  "audio_processing_completed": "✅ オーディオ処理完了!",
+  "processing_audio_print": "オーディオを処理中: {}、モデル使用: {}",
+  "starting_audio_separation_progress_label": "オーディオ分離を開始しています... -- 0.0%",
+  "separating_audio_progress_label": "オーディオを分離中... -- {}.0%",
+  "audio_processing_completed_progress_label": "オーディオ処理が完了しました！ -- 100.0%",
+  "audio_processing_completed": "✅ オーディオ処理が完了しました！",
   "error_occurred_progress_label": "エラーが発生しました -- 0.0%",
-  "minimum_files_required": "⚠️ 最低2ファイル必要",
+  "minimum_files_required": "⚠️ 最低2つのファイルが必要です",
   "valid_files_not_found": "❌ 有効なファイルが見つかりません",
-  "starting_ensemble_process": "アンサンブルプロセスを開始...",
+  "starting_ensemble_process": "アンサンブルプロセスを開始しています...",
   "ensembling_progress": "アンサンブル中... ({:.1f}%)",
-  "finalizing_ensemble_output": "アンサンブル出力を最終処理中...",
-  "success_log": "✅ 成功!\n{}",
-  "error_log": "❌ エラー!\n{}",
-  "critical_error": "⛔ 重大なエラー: {}",
-  "ensemble_process_completed": "アンサンブルプロセス完了!",
+  "finalizing_ensemble_output": "アンサンブル出力を最終化中...",
+  "success_log": "✅ 成功！\n{}",
+  "error_log": "❌ エラー！\n{}",
+  "ensemble_process_completed": "アンサンブルプロセスが完了しました！",
   "no_models_selected": "❌ モデルが選択されていません",
   "no_input_audio_provided": "❌ 入力オーディオが提供されていません",
-  "loading_model": "モデル {}/{} をロード中: {}...",
-  "loading_model_progress_label": "モデル {}/{} をロード中: {}... -- {}.0%",
-  "completed_model": "モデル {}/{} 完了: {}",
-  "completed_model_progress_label": "モデル {}/{} 完了: {} -- {}.0%",
-  "running_command": "コマンド実行中: {}",
-  "model_failed": "モデル {} が失敗: {}",
-  "critical_error_with_model": "{} で重大なエラー: {}",
-  "model_output_failed": "{} が出力を生成できませんでした",
-  "waiting_for_files": "すべてのファイルの準備を待機中...",
-  "waiting_for_files_progress_label": "すべてのファイルの準備を待機中... -- 90.0%",
-  "performing_ensemble": "アンサンブル実行中...",
-  "performing_ensemble_progress_label": "アンサンブル実行中... -- 92.0%",
+  "loading_model": "モデルを読み込み中 {}/{}: {}...",
+  "loading_model_progress_label": "モデルを読み込み中 {}/{}: {}... -- {}.0%",
+  "completed_model": "モデル完了 {}/{}: {}",
+  "completed_model_progress_label": "モデル完了 {}/{}: {} -- {}.0%",
+  "running_command": "コマンドを実行中: {}",
+  "model_failed": "モデル {} が失敗しました: {}",
+  "critical_error_with_model": "{} で致命的なエラー: {}",
+  "model_output_failed": "{} は出力を生成できませんでした",
+  "waiting_for_files": "すべてのファイルが準備できるまで待機中...",
+  "waiting_for_files_progress_label": "すべてのファイルが準備できるまで待機中... -- 90.0%",
+  "performing_ensemble": "アンサンブルを実行中...",
+  "performing_ensemble_progress_label": "アンサンブルを実行中... -- 92.0%",
   "memory_usage_before_ensemble": "アンサンブル前のメモリ使用量: {}%",
   "memory_usage_after_ensemble": "アンサンブル後のメモリ使用量: {}%",
-  "finalizing_ensemble_output_progress_label": "アンサンブル出力を最終処理中... -- 98.0%",
-  "ensemble_file_creation_failed": "アンサンブルファイルの作成に失敗: {}",
-  "ensemble_completed_progress_label": "アンサンブルが正常に完了しました! -- 100.0%",
-  "success_output_created": "✅ 成功! 出力ファイルが作成されました。",
-  "drive_mounted_copying_ensemble": "Google Driveがマウントされました。アンサンブル出力をコピー中...",
-  "drive_already_mounted_copying_ensemble": "Google Driveはすでにマウントされています。アンサンブル出力をコピー中...",
-  "no_ensemble_output_files_found": "❌ アンサンブル出力ファイルが見つかりません。",
-  "ensemble_output_copied": "✅ アンサンブル出力が {} にコピーされました",
-  "error_copying_ensemble_output": "❌ アンサンブル出力のコピーにエラー: {}",
-  "drive_mounted_copying_files": "Google Driveがマウントされました。ファイルをコピー中...",
-  "drive_already_mounted_copying_files": "Google Driveはすでにマウントされています。ファイルをコピー中...",
-  "files_copied_to_drive": "✅ ファイルが {} にコピーされました",
-  "error_copying_files": "❌ ファイルのコピーにエラー: {}",
-  "mounting_drive": "Google Driveをマウント中...",
-  "settings_saved": "設定が config.json に保存されました: 言語={lang}, メソッド={method}, ポート={port}",
-  "selected_language": "選択された言語: {lang}",
-  "available_languages": "利用可能な言語: {langs}",
-  "select_language_prompt": "言語を選択 (デフォルト: {default}): ",
-  "invalid_language": "無効な言語選択! デフォルト言語 ({default}) を使用します。",
-  "available_sharing_methods": "利用可能な共有方法: gradio, localtunnel, ngrok",
-  "select_sharing_method_prompt": "共有方法を選択 (デフォルト: {default}): ",
-  "enter_ngrok_token_prompt": "Ngrokトークンを入力 (デフォルト: {default}): ",
-  "ngrok_token_required": "エラー: ngrokメソッドにはNgrokトークンが必要です!",
-  "enter_port_prompt": "ポート番号を入力 (デフォルト: {default}): ",
-  "opening_previous_url": "前のURLを開く: {url}",
-  "open_url_manually": "ColabでURLを手動で開く: {url}",
-  "invalid_previous_url": "無効な前のURL: {url}, 開かれませんでした。",
-  "starting_method": "{method} をポート {port} で開始...",
-  "process_stopped": "\n🛑 ユーザーによってプロセスが停止されました",
-  "selected_port": "選択されたポート: {port}",
-  "starting_gradio_with_sharing": "Gradioを共有機能付きで開始...",
-  "starting_localtunnel": "Localtunnelをポート {port} で開始...",
-  "share_link": "共有リンク: {url}",
-  "password_ip": "パスワードIP: {ip}",
-  "starting_ngrok": "Ngrokをポート {port} で開始...",
-  "ngrok_url": "Ngrok URL: {url}",
-  "ngrok_error": "Ngrok開始エラー: {error}",
-  "apollo_chunk_size": "Apolloチャンクサイズ",
-  "apollo_chunk_size_info": "Apollo改善のためのチャンクサイズ (ユニバーサルモデル推奨: 19)",
-  "apollo_overlap": "Apolloオーバーラップ",
-  "enhancing_with_apollo": "Apolloで改善中 ({}/{} ファイル)...",
-  "apollo_processing_completed": "Apollo処理完了!",
-  "apollo_overlap_info": "Apollo改善のためのオーバーラップ (推奨: 2)",
-  "invalid_method": "エラー: 無効なメソッド! 'gradio', 'localtunnel', または 'ngrok' を使用してください。",
-  "apollo_enhancement_settings": "Apollo改善設定",
-  "enhance_with_apollo": "Apolloで改善",
-  "enhance_with_apollo_info": "分離後のオーディオ改善のためにApolloを有効化",
-  "apollo_method": "Apolloメソッド",
-  "apollo_method_info": "Apolloの処理メソッドを選択",
-  "normal_method": "通常メソッド",
-  "mid_side_method": "ミッド/サイドメソッド",
-  "apollo_normal_model": "Apollo通常モデル",
-  "apollo_normal_model_info": "通常Apollo処理に使用するモデル",
-  "apollo_midside_model": "Apolloミッド/サイドモデル",
-  "apollo_enhancement_info": "音質の改善",
-  "selected_models": "選択されたモデル",
-  "save_preset": "プリセットを保存",
-  "delete_preset": "プリセットを削除",
-  "refresh_presets": "プリセットを更新",
-  "preset_name": "プリセット名",
-  "select_preset": "プリセットを選択",
-  "add_favorite": "お気に入りに追加",
-  "apollo_midside_model_info": "ミッド/サイド処理に使用するモデル (オプション)",
-  "language_changed_message": "言語が変更されました。セルを再起動してください。",
-  "apply_matchering": "Matcheringを適用",
-  "matchering_info": "オーディオのマスタリングとバランスを向上させるためにMatcheringを適用",
-  "matchering_passes": "Matcheringの反復回数",
-  "matchering_passes_info": "Matcheringの反復回数（1-5、高い値は品質を向上させますが時間がかかります）",
-  "matchering_processing": "Matcheringを適用中... ({:.1f}%)",
-  "matchering_completed": "✅ Matchering処理が完了しました！",
-  "matchering_error": "❌ Matchering中にエラーが発生しました：{}"
+  "finalizing_ensemble_output_progress_label": "アンサンブル出力を最終化中... -- 98.0%",
+  "ensemble_file_creation_failed": "アンサンブルファイルを作成できませんでした: {}",
+  "ensemble_completed_progress_label": "アンサンブルが正常に完了しました！ -- 100.0%",
+  "success_output_created": "✅ 成功！出力ファイルが作成されました。",
+  "custom_models_tab": "⚡ カスタムモデル",
+  "add_custom_model": "➕ カスタムモデルを追加",
+  "custom_model_name": "モデル名",
+  "custom_model_name_placeholder": "モデルの名前を入力してください",
+  "checkpoint_url": "チェックポイントURL (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "設定URL (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "カスタムPythonファイルURL (オプション)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "モデルタイプを自動検出",
+  "model_type": "モデルタイプ",
+  "add_model_btn": "➕ モデルを追加",
+  "custom_models_list": "📋 カスタムモデルリスト",
+  "delete_model": "🗑️ 削除",
+  "no_custom_models": "カスタムモデルはまだ追加されていません",
+  "model_added_success": "✅ モデルが正常に追加されました！",
+  "model_deleted_success": "✅ モデルが正常に削除されました！",
+  "model_add_error": "❌ モデル追加エラー: {}",
+  "model_delete_error": "❌ モデル削除エラー: {}",
+  "refresh_models": "🔄 更新",
+  "custom_model_info": "ダウンロードリンクを提供してカスタムモデルを追加します。モデルは使用時に自動的にダウンロードされます。",
+  "select_model_to_delete": "削除するモデルを選択してください"
 }

assets/i18n/languages/ru_ru.json

@@ -19,6 +19,7 @@
   "tta_boost": "Усиление TTA",
   "tta_info": "Улучшает качество, но замедляет скорость обработки",
   "phase_fix": "Коррекция фазы",
+  "use_phase_correction": "Применить расширенную коррекцию фазы",
   "phase_fix_info": "Расширенная коррекция фазы для инструментальных треков",
   "instrumental": "Инструментал",
   "instrumental_info": "Обычно дает 2 выхода, но иногда только один, нужно экспериментировать :)",
@@ -46,7 +47,7 @@
   "copy_to_drive": "📂 Копировать на Диск",
   "copy_status": "Статус копирования",
   "files_will_be_copied": "Файлы будут скопированы сюда...",
-  "processing_tip": "<b>🔈 Совет по обработке:</b> Для шумных результатов используйте модели <code>bleed_suppressor_v1</code> или <code>denoisedebleed</code> в категории <i>\"Удаление шума и эффектов\"</i>, чтобы очистить выходные данные",
+  "processing_tip": "<b>🔈 Совет по обработке:</b> Для шумных результатов используйте модели <code>bleed_suppressor_v1</code> или <code>denoisedebleed</code> в категории <i>Удаление шума и эффектов</i>, чтобы очистить выходные данные",
   "waiting_for_processing": "Ожидание обработки...",
   "auto_ensemble_tab": "🤖 Автоматический ансамбль",
   "upload_file": "Загрузить файл",
@@ -71,68 +72,147 @@
   "output_preview": "Предпросмотр вывода",
   "refresh_output": "🔄 Обновить вывод",
   "ensemble_copy_status": "Вывод ансамбля будет скопирован сюда...",
-  "guidelines": "<h4>Руководство по выбору моделей</h4><ul><li><strong>Избегайте смешивания категорий:</strong> Сочетание вокальных и инструментальных моделей может привести к нежелательным смешиваниям</li><li><strong>Заметки о специальных моделях:</strong><ul><li>Модели дуальности (v1/v2) - выдают оба стема</li><li>Сепаратор MDX23C - гибридные результаты</li></ul></li><li><strong>Лучшая практика:</strong> Используйте 3-5 схожих моделей из одной категории</li></ul><div>💡 Профессиональный совет: Начните с комбинации \"VOCALS-MelBand-Roformer BigBeta5e\" + \"VOCALS-BS-Roformer_1297\"</div>",
+  "guidelines": "<h4>Руководство по выбору модели</h4><ul><li><strong>Избегайте смешивания категорий:</strong> Сочетание вокальных и инструментальных моделей может привести к нежелательным смесям</li><li><strong>Примечания к специальным моделям:</strong><ul><li>Модели двойственности (v1/v2) - Выводят оба стема</li><li>Разделитель MDX23C - Гибридные результаты</li></ul></li><li><strong>Лучшая практика:</strong> Используйте 3-5 похожих моделей из одной категории</li></ul><div>💡 Совет профессионала: Начните с комбинации \"VOCALS-MelBand-Roformer BigBeta5e\" + \"VOCALS-BS-Roformer_1297\"</div>",
   "download_sources_tab": "⬇️ Скачать источники",
   "direct_links": "🌐 Прямые ссылки",
   "audio_file_url": "URL аудиофайла",
-  "download_from_url": "⬇️ Скачать по URL",
+  "download_from_url": "⬇️ Скачать с URL",
   "download_status": "Статус загрузки",
-  "downloaded_file": "Скачанный файл",
-  "cookie_management": "🍪 Управление cookies",
+  "downloaded_file": "Загруженный файл",
+  "cookie_management": "🍪 Управление куки",
   "upload_cookies_txt": "Загрузить Cookies.txt",
-  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 Зачем это нужно?**<br>- Доступ к контенту с возрастными ограничениями<br>- Скачивание приватных/неопубликованных видео<br>- Обход региональных ограничений<br>- Избежание лимитов скачивания YouTube<br><br>**⚠️ Важные заметки**<br>- НИКОГДА не делитесь своими файлами cookies!<br>- Обновляйте cookies, если:<br>  • П��являются ошибки \"403 Запрещено\"<br>  • Загрузки внезапно останавливаются<br>  • Появляется сообщение \"Сессия истекла\"<br><br>**🔄 Шаги обновления**<br>1. Установите это <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">расширение Chrome</a><br>2. Войдите в YouTube в Chrome<br>3. Нажмите на иконку расширения → \"Экспорт\"<br>4. Загрузите скачанный файл сюда<br><br>**⏳ Срок действия cookies**<br>- Обычные сессии: 24 часа<br>- Чувствительные операции: 1 час<br>- Смена пароля: мгновенная недействительность</div>",
+  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 Почему необходимо?**<br>- Доступ к контенту с ограничением по возрасту<br>- Скачивание приватных/неопубликованных видео<br>- Обход региональных ограничений<br>- Избежание лимитов скачивания YouTube<br><br>**⚠️ Важные замечания**<br>- НИКОГДА не делитесь своими файлами куки!<br>- Обновляйте куки, если:<br>  • Получаете ошибки \"403 Запрещено\"<br>  • Загрузки внезапно останавливаются<br>  • Видите сообщения \"Сессия истекла\"<br><br>**🔄 Шаги обновления**<br>1. Установите это <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">расширение Chrome</a><br>2. Войдите в YouTube в Chrome<br>3. Нажмите на иконку расширения → \"Экспорт\"<br>4. Загрузите скачанный файл сюда<br><br>**⏳ Срок действия куки**<br>- Обычные сессии: 24 часа<br>- Чувствительные операции: 1 час<br>- Смена пароля: Немедленное обновление</div>",
+  "ensemble_output_copied": "✅ Вывод ансамбля скопирован в {}",
+  "error_copying_ensemble_output": "❌ Ошибка копирования вывода ансамбля: {}",
+  "drive_mounted_copying_ensemble": "Google Drive подключен. Копирование вывода ансамбля...",
+  "drive_already_mounted_copying_ensemble": "Google Drive уже подключен. Копирование вывода ансамбля...",
+  "no_ensemble_output_files_found": "❌ Файлы вывода ансамбля не найдены.",
+  "drive_mounted_copying_files": "Google Drive подключен. Копирование файлов...",
+  "drive_already_mounted_copying_files": "Google Drive уже подключен. Копирование файлов...",
+  "files_copied_to_drive": "✅ Файлы скопированы в {}",
+  "error_copying_files": "❌ Ошибка копирования файлов: {}",
+  "mounting_drive": "Подключение Google Drive...",
+  "settings_saved": "Настройки сохранены в config.json: Язык={lang}, Метод={method}, Порт={port}",
+  "selected_language": "Выбранный язык: {lang}",
+  "available_languages": "Доступные языки: {langs}",
+  "select_language_prompt": "Выберите язык (по умолчанию: {default}): ",
+  "invalid_language": "Неверный выбор языка! Используется язык по умолчанию ({default}).",
+  "available_sharing_methods": "Доступные методы совместного доступа: gradio, localtunnel, ngrok",
+  "select_sharing_method_prompt": "Выберите метод совместного доступа (по умолчанию: {default}): ",
+  "enter_ngrok_token_prompt": "Введите ваш токен Ngrok (по умолчанию: {default}): ",
+  "ngrok_token_required": "Ошибка: Для метода ngrok требуется токен Ngrok!",
+  "enter_port_prompt": "Введите номер порта (по умолчанию: {default}): ",
+  "opening_previous_url": "Открытие предыдущего URL: {url}",
+  "open_url_manually": "Вручную откройте URL в Colab: {url}",
+  "invalid_previous_url": "Неверный предыдущий URL: {url}, не открыт.",
+  "starting_method": "Запуск {method} на порту {port}...",
+  "process_stopped": "\n🛑 Процесс остановлен пользователем",
+  "selected_port": "Выбранный порт: {port}",
+  "starting_gradio_with_sharing": "Запуск Gradio с встроенным совместным до��тупом...",
+  "starting_localtunnel": "Запуск Localtunnel на порту {port}...",
+  "share_link": "Ссылка для совместного доступа: {url}",
+  "password_ip": "Пароль IP: {ip}",
+  "starting_ngrok": "Запуск Ngrok на порту {port}...",
+  "ngrok_url": "URL Ngrok: {url}",
+  "ngrok_error": "Ошибка запуска Ngrok: {error}",
+  "auto_apollo_chunk_size": "Размер фрагмента Apollo",
+  "auto_apollo_chunk_size_info": "Размер фрагмента для улучшения Apollo (рекомендуется: 19 для универсальной модели)",
+  "auto_apollo_overlap": "Перекрытие Apollo",
+  "auto_enhancing_with_apollo": "Улучшение с Apollo ({}/{} файлов)...",
+  "auto_apollo_processing_completed": "Обработка Apollo завершена!",
+  "auto_apollo_overlap_info": "Перекрытие для улучшения Apollo (рекомендуется: 2)",
+  "apollo_chunk_size": "Размер фрагмента Apollo",
+  "apollo_chunk_size_info": "Размер фрагмента для улучшения Apollo (рекомендуется: 19 для универсальной модели)",
+  "apollo_overlap": "Перекрытие Apollo",
+  "enhance_with_apollo": "Улучшить с Apollo",
+  "apollo_enhancement_info": "Улучшение качества звука",
+  "enhancing_with_apollo": "Улучшение с Apollo ({}/{} файлов)...",
+  "apollo_processing_completed": "Обработка Apollo завершена!",
+  "apollo_overlap_info": "Перекрытие для улучшения Apollo (рекомендуется: 2)",
+  "selected_models": "Выбранные модели",
+  "save_preset": "Сохранить пресет",
+  "delete_preset": "Удалить пресет",
+  "refresh_presets": "Обновить пресеты",
+  "preset_name": "Имя пресета",
+  "select_preset": "Выбрать пресет",
+  "add_favorite": "Добавить в избранное",
+  "apply_matchering": "Применить Matchering",
+  "matchering_info": "Применить Matchering для улучшения мастеринга и баланса звука",
+  "matchering_passes": "Проходы Matchering",
+  "matchering_passes_info": "Количество итераций Matchering (1-5, более высокие значения улучшают качество, но занимают больше времени)",
+  "matchering_processing": "Применение Matchering... ({:.1f}%)",
+  "matchering_completed": "✅ Обработка Matchering завершена!",
+  "matchering_error": "❌ Ошибка во время Matchering: {}",
+  "invalid_method": "Ошибка: Неверный метод! Используйте 'gradio', 'localtunnel' или 'ngrok'.",
+  "language_changed_message": "Язык изменен. Пожалуйста, перезапустите ячейку.",
+  "backend_settings": "🚀 Настройки Backend",
+  "inference_backend": "Backend инференса",
+  "optimization_mode": "Режим оптимизации",
+  "optimization_mode_info": "Выбор режима оптимизации PyTorch",
+  "mixed_precision_amp": "Смешанная точность (AMP)",
+  "mixed_precision_info": "Inference в 2 раза быстрее - рекомендуется",
+  "tf32_acceleration": "Ускорение TF32",
+  "tf32_acceleration_info": "Дополнительный прирост скорости для RTX 30xx+",
+  "cudnn_benchmark": "Бенчмарк cuDNN",
+  "cudnn_benchmark_info": "Медленнее на первом запуске, гораздо быстрее на последующих",
+  "ultra_optimized_pytorch": "УЛЬТРА-ОПТИМИЗИРОВАННЫЙ PyTorch Backend",
+  "default_active_max_speed": "Активен по умолчанию - Максимальная оптимизация скорости",
+  "channels_last_mode": "Самый быстрый для RTX GPU",
+  "compile_mode": "Дополнительная скорость для PyTorch 2.0+",
+    "default_mode": "Стандартный",
   "manual_ensemble_tab": "🎚️ Ручной ансамбль",
   "input_sources": "📂 Источники ввода",
   "refresh": "🔄 Обновить",
   "ensemble_algorithm": "Алгоритм ансамбля",
-  "select_audio_files": "Выберите аудиофайлы",
+  "select_audio_files": "Выбрать аудиофайлы",
   "available_files": "Доступные файлы",
-  "custom_weights": "Пользовательские веса (разделены запятыми)",
+  "custom_weights": "Пользовательские веса (через запятую)",
   "custom_weights_placeholder": "Пример: 0.8, 1.2, 1.0, ...",
   "custom_weights_info": "Оставьте пустым для равных весов",
   "result_preview_tab": "🎧 Предпросмотр результата",
-  "ensembled_output": "Результат ансамбля",
+  "ensembled_output": "Вывод ансамбля",
   "processing_log_tab": "📋 Журнал обработки",
   "processing_details": "Детали обработки",
   "process_ensemble": "⚡ Обработать ансамбль",
   "languages_tab": "🌐 Языки",
   "Vocal Models": "Вокальные модели",
   "Instrumental Models": "Инструментальные модели",
-  "4-Stem Models": "Модели с 4 стемами",
+  "4-Stem Models": "4-стемовые модели",
   "Denoise Models": "Модели шумоподавления",
   "Dereverb Models": "Модели удаления реверберации",
   "Other Models": "Другие модели",
-  "ensemble_files_description": "Объединяет несколько аудиофайлов в один выход с использованием указанных методов ансамбля.",
+  "Other Shapes": "Другие модели",
+  "ensemble_files_description": "Объединяет несколько аудиофайлов в один выходной файл с использованием указанных методов ансамбля.",
   "ensemble_files_help": "Путь ко всем аудиофайлам для ансамбля",
   "ensemble_type_help": "Один из: avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
   "ensemble_weights_help": "Веса для создания ансамбля. Количество весов должно совпадать с количеством файлов",
-  "ensemble_output_help": "Путь к WAV-файлу, куда будет сохранен результат ансамбля",
+  "ensemble_output_help": "Путь к WAV-файлу, где будет сохранен результат ансамбля",
   "ensemble_type_print": "Тип ансамбля: {}",
   "num_input_files_print": "Количество входных файлов: {}",
   "weights_print": "Веса: {}",
   "output_file_print": "Выходной файл: {}",
-  "duration_mismatch_error": "Все файлы должны иметь одинаковую длительность",
-  "file_not_found_error": "Ошибка. Файл не найден: {}. Проверьте пути.",
-  "reading_chunk_print": "Чтение фрагмента из файла: {} (начало: {} сек, длительность: {} сек)",
+  "duration_mismatch_error": "Все файлы должны иметь одинаковую продолжительность",
+  "file_not_found_error": "Ошибка. Не удается найти файл: {}. Проверьте пути.",
+  "reading_chunk_print": "Чтение фрагмента из файла: {} (начало: {}с, продолжительность: {}с)",
   "chunk_result_shape_print": "Форма результата фрагмента: {}",
-  "ensemble_completed_print": "Ансамбль завершен. Выход сохранен в: {}",
-  "file_deletion_error": "Не удалось удалить {}: {}",
-  "directory_not_exist_warning": "⚠️ Директория не существует: {}",
-  "not_a_directory_warning": "⚠️ Путь не является директорией: {}",
+  "ensemble_completed_print": "Ансамбль завершен. Вывод сохранен в: {}",
+  "file_deletion_error": "{} не удалось удалить: {}",
+  "directory_not_exist_warning": "⚠️ Каталог не существует: {}",
+  "not_a_directory_warning": "⚠️ Путь не является каталогом: {}",
   "item_deletion_error": "⚠️ Ошибка удаления {}: {}",
   "old_output_not_exist": "❌ Старая папка вывода не существует",
-  "old_outputs_cleared": "✅ Старые выходные данные успешно удалены!",
+  "old_outputs_cleared": "✅ Старые выходные данные успешно очищены!",
   "error": "🔥 Ошибка: {}",
   "ffmpeg_error": "Ошибка FFmpeg ({}): {}",
   "file_saved_successfully": "Файл успешно сохранен: {}",
   "total_files_found": "Всего найдено файлов: {}. Используемая частота дискретизации: {}",
   "total_progress": "Общий прогресс",
-  "detailed_pbar_enabled": "Подробная шкала прогресса включена: {}",
-  "loaded_audio": "Загруженное аудио: {}, форма: {}",
+  "detailed_pbar_enabled": "Подробная полоса прогресса включена: {}",
+  "loaded_audio": "Загружено аудио: {}, форма: {}",
   "cannot_read_track": "Не удается прочитать трек: {}",
   "error_message": "Сообщение об ошибке: {}",
   "demudding_track": "Очистка трека (ремикс фазы - инструментал): {}",
-  "elapsed_time": "Прошедшее время: {:.2f} сек.",
+  "elapsed_time": "Прошедшее время: {:.2f} секунд.",
   "proc_folder_description": "Обрабатывает аудиофайлы в папке с использованием указанной модели.",
   "model_type_help": "Тип модели (bandit, bs_roformer, mdx23c и т.д.)",
   "config_path_help": "Путь к файлу конфигурации",
@@ -140,32 +220,40 @@
   "start_checkpoint_help": "Начальная контрольная точка для действительных весов",
   "input_folder_help": "Папка с миксами для обработки",
   "audio_path_help": "Путь к одному аудиофайлу для обработки",
-  "store_dir_help": "Путь для хранения результатов",
+  "store_dir_help": "Путь для сохранения результатов",
   "device_ids_help": "Список идентификаторов GPU",
-  "extract_instrumental_help": "Инвертировать вокал для получения инструментала, если предоставлено",
-  "disable_detailed_pbar_help": "Отключить подробную шкалу прогресса при демиксе",
-  "force_cpu_help": "Принудительно использовать CPU, даже если CUDA доступен",
-  "flac_file_help": "Выводить файл FLAC вместо WAV",
+  "extract_instrumental_help": "Инвертировать вокал для получения инструментала, если предоставлен",
+  "disable_detailed_pbar_help": "Отключить подробную полосу прогресса в демиксе",
+  "force_cpu_help": "Принудительно использовать CPU, даже если доступна CUDA",
+  "flac_file_help": "Вывод FLAC-файла вместо WAV",
   "export_format_help": "Формат экспорта и тип PCM",
-  "pcm_type_help": "Тип PCM для файлов FLAC",
-  "use_tta_help": "Включить увеличение времени теста",
+  "pcm_type_help": "Тип PCM для FLAC-файлов",
+  "use_tta_help": "Включить тестовое усиление времени",
   "lora_checkpoint_help": "Начальная контрольная точка для весов LoRA",
-  "cuda_available": "CUDA доступен, используйте --force_cpu для отключения.",
+  "cuda_available": "CUDA доступна, используйте --force_cpu для отключения.",
   "using_device": "Используемое устройство: {}",
   "instruments_print": "Инструменты: {}",
   "model_load_time": "Время загрузки модели: {:.2f} сек",
-  "invalid_url": "❌ Неверный URL",
-  "cookie_file_updated": "✅ Файл cookies обновлен!",
-  "cookie_installation_error": "⚠️ Ошибка установки cookies: {}",
-  "file_size_zero_error": "Размер файла нулевой или файл не создан",
-  "google_drive_error": "❌ Ошибка загрузки с Google Drive: {}",
-  "wav_conversion_failed": "Не удалось конвертировать в WAV",
+  "invalid_url": "❌ Недействительный URL",
+  "cookie_file_updated": "✅ Файл куки обновлен!",
+  "cookie_installation_error": "⚠️ Ошибка установки куки: {}",
+  "google_drive_mount_error": "❌ Ошибка подключения Google Drive: {}",
+  "file_size_zero_error": "Размер файла ноль или файл не создан",
+  "google_drive_error": "❌ Ошибка загрузки Google Drive: {}",
+  "wav_conversion_failed": "Конвертация WAV не удалась",
   "download_error": "❌ Ошибка загрузки: {}",
-  "download_success": "🎉 Загрузка успешно завершена!",
+  "file_copied_to_drive": "Файл скопирован в Google Drive: {}",
+  "copy_to_drive_error": "❌ Ошибка копирования в Google Drive: {}",
+  "direct_download_failed": "Прямая загрузка файла не удалась",
+  "direct_download_error": "❌ Ошибка прямой загрузки: {}",
+  "invalid_google_drive_url": "❌ Недействительный URL Google Drive",
+  "continuing_without_google_drive": "Подключение Google Drive не удалось, продолжаем без него...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive не подключен, пропуск копирования на Drive...",
+  "download_success": "🎉 Успешно загружено!",
   "download_failed": "❌ Загрузка не удалась",
   "no_output_files_found": "❌ Выходные файлы не найдены.",
   "output_refreshed_successfully": "✅ Вывод успешно обновлен!",
-  "error_refreshing_output": "❌ Ошибка при обновлении вывода: {}",
+  "error_refreshing_output": "❌ Ошибка обновления вывода: {}",
   "starting_audio_separation": "Запуск разделения аудио...",
   "processing_audio": "Обработка аудио",
   "separating_audio": "Разделение аудио... ({:.1f}%)",
@@ -173,7 +261,7 @@
   "progress_parsing_error": "Ошибка разбора прогресса: {}",
   "error_occurred": "Произошла ошибка: {}",
   "separation_process_completed": "Процесс разделения завершен!",
-  "no_audio_file_error": "❌ Аудиофайл не предоставлен, и в директории ввода нет существующего файла.",
+  "no_audio_file_error": "❌ Аудиофайл не предоставлен и нет существующего файла в каталоге ввода.",
   "no_input_progress_label": "Ошибка: Ввод не предоставлен -- 0.0%",
   "processing_audio_print": "Обработка аудио из: {} с использованием модели: {}",
   "starting_audio_separation_progress_label": "Запуск разделения аудио... -- 0.0%",
@@ -182,11 +270,11 @@
   "audio_processing_completed": "✅ Обработка аудио завершена!",
   "error_occurred_progress_label": "Произошла ошибка -- 0.0%",
   "minimum_files_required": "⚠️ Требуется минимум 2 файла",
-  "valid_files_not_found": "❌ Валидные файлы не найдены",
+  "valid_files_not_found": "❌ Действительные файлы не найдены",
   "starting_ensemble_process": "Запуск процесса ансамбля...",
-  "ensembling_progress": "Создание ансамбля... ({:.1f}%)",
+  "ensembling_progress": "Ансамблирование... ({:.1f}%)",
   "finalizing_ensemble_output": "Завершение вывода ансамбля...",
-  "success_log": "✅ Успех!\n{}",
+  "success_log": "✅ Успешно!\n{}",
   "error_log": "❌ Ошибка!\n{}",
   "critical_error": "⛔ Критическая ошибка: {}",
   "ensemble_process_completed": "Процесс ансамбля завершен!",
@@ -204,77 +292,55 @@
   "waiting_for_files_progress_label": "Ожидание готовности всех файлов... -- 90.0%",
   "performing_ensemble": "Выполнение ансамбля...",
   "performing_ensemble_progress_label": "Выполнение ансамбля... -- 92.0%",
-  "memory_usage_before_ensemble": "Использование памяти перед ансамблем: {}%",
+  "memory_usage_before_ensemble": "Использование памяти до ансамбля: {}%",
   "memory_usage_after_ensemble": "Использование памяти после ансамбля: {}%",
   "finalizing_ensemble_output_progress_label": "Завершение вывода ансамбля... -- 98.0%",
   "ensemble_file_creation_failed": "Не удалось создать файл ансамбля: {}",
   "ensemble_completed_progress_label": "Ансамбль успешно завершен! -- 100.0%",
-  "success_output_created": "✅ Успех! Выходной файл создан.",
-  "drive_mounted_copying_ensemble": "Google Drive подключен. Копирование вывода ансамбля...",
-  "drive_already_mounted_copying_ensemble": "Google Drive уже подключен. Копирование вывода ансамбля...",
-  "no_ensemble_output_files_found": "❌ Файлы вывода ансамбля не найдены.",
-  "ensemble_output_copied": "✅ Вывод ансамбля скопирован в {}",
-  "error_copying_ensemble_output": "❌ Ошибка копирования вывода ансамбля: {}",
-  "drive_mounted_copying_files": "Google Drive подключен. Копирование файлов...",
-  "drive_already_mounted_copying_files": "Google Drive уже подключен. Копирование файлов...",
-  "files_copied_to_drive": "✅ Файлы скопированы в {}",
-  "error_copying_files": "❌ Ошибка копирования файлов: {}",
-  "mounting_drive": "Подключение Google Drive...",
-  "settings_saved": "Настройки сохранены в config.json: Язык={lang}, Метод={method}, Порт={port}",
-  "selected_language": "Выбранный язык: {lang}",
-  "available_languages": "Доступные языки: {langs}",
-  "select_language_prompt": "Выберите язык (по умолчанию: {default}): ",
-  "invalid_language": "Неверный выбор языка! Используется язык по умолчанию ({default}).",
-  "available_sharing_methods": "Доступные методы совместного доступа: gradio, localtunnel, ngrok",
-  "select_sharing_method_prompt": "Выберите метод совместного доступа (по умолчанию: {default}): ",
-  "enter_ngrok_token_prompt": "Введите ваш токен Ngrok (по умолчанию: {default}): ",
-  "ngrok_token_required": "Ошибка: Для метода ngrok требуется токен Ngrok!",
-  "enter_port_prompt": "Введите номер порта (по умолчанию: {default}): ",
-  "opening_previous_url": "Открытие предыдущего URL: {url}",
-  "open_url_manually": "Откройте URL вручную в Colab: {url}",
-  "invalid_previous_url": "Неверный предыдущий URL: {url}, не открыт.",
-  "starting_method": "Запуск {method} на порту {port}...",
-  "process_stopped": "\n🛑 Процесс остановлен пользователем",
-  "selected_port": "Выбранный порт: {port}",
-  "starting_gradio_with_sharing": "Запуск Gradio с встроенным совместным доступом...",
-  "starting_localtunnel": "Запуск Localtunnel на порту {port}...",
-  "share_link": "Ссылка для совместного доступа: {url}",
-  "password_ip": "Пароль IP: {ip}",
-  "starting_ngrok": "Запуск Ngrok на порту {port}...",
-  "ngrok_url": "URL Ngrok: {url}",
-  "ngrok_error": "Ошибка запуска Ngrok: {error}",
-  "apollo_chunk_size": "Размер фрагмента Apollo",
-  "apollo_chunk_size_info": "Размер фрагмента для улучшения Apollo (рекомендуется: 19 для универсальной модели)",
-  "apollo_overlap": "Перекрытие Apollo",
-  "enhancing_with_apollo": "Улучшение с Apollo ({}/{} файлов)...",
-  "apollo_processing_completed": "Обработка Apollo завершена!",
-  "apollo_overlap_info": "Перекрытие для улучшения Apollo (рекомендуется: 2)",
-  "invalid_method": "Ошибка: Неверный метод! Используйте 'gradio', 'localtunnel' или 'ngrok'.",
-  "apollo_enhancement_settings": "Настройки улучшения Apollo",
-  "enhance_with_apollo": "Улучшить с Apollo",
-  "enhance_with_apollo_info": "Включить Apollo для улучшения аудио после разделения",
-  "apollo_method": "Метод Apollo",
-  "apollo_method_info": "Выберите метод обработки для Apollo",
-  "normal_method": "Обычный метод",
-  "mid_side_method": "Метод середины/сторон",
-  "apollo_normal_model": "Обычная модель Apollo",
-  "apollo_normal_model_info": "Модель, используемая для обычной обработки Apollo",
-  "apollo_midside_model": "Модель Apollo середины/сторон",
-  "apollo_enhancement_info": "Улучшение качества звука",
-  "selected_models": "Выбранные модели",
-  "save_preset": "Сохранить пресет",
-  "delete_preset": "Удалить пресет",
-  "refresh_presets": "Обновить пресеты",
-  "preset_name": "Имя пресета",
-  "select_preset": "Выбрать пресет",
-  "add_favorite": "Добавить в избранное",
-  "apollo_midside_model_info": "Модель, используемая для обработки середины/сторон (опционально)",
-  "language_changed_message": "Язык изменен. Пожалуйста, перезапустите ячейку.",
-  "apply_matchering": "Применить Matchering",
-  "matchering_info": "Применить Matchering для улучшения мастеринга и баланса звука",
-  "matchering_passes": "Проходы Matchering",
-  "matchering_passes_info": "Количество итераций Matchering (1-5, более высокие значения улучшают качество, но занимают больше времени)",
-  "matchering_processing": "Применение Matchering... ({:.1f}%)",
-  "matchering_completed": "✅ Обработка Matchering завершена!",
-  "matchering_error": "❌ Ошибка во время Matchering: {}"
+  "success_output_created": "✅ Успешно! Создан выходной файл.",
+  "phase_fixer_tab": "🔧 Корректор фазы",
+  "batch_processing_tab": "📦 Пакетная обработка",
+  "batch_description": "Обработка нескольких файлов одновременно с одной моделью",
+  "batch_input_folder": "Входная папка",
+  "batch_input_folder_placeholder": "Введите путь к папке с аудиофайлами",
+  "batch_output_folder": "Выходная папка",
+  "batch_output_folder_placeholder": "Введите путь к папке для выходных файлов",
+  "batch_file_list": "Файлы для обработки",
+  "batch_add_files": "Добавить файлы",
+  "batch_clear_list": "Очистить список",
+  "batch_start": "🚀 Начать пакетную обработку",
+  "batch_stop": "⏹️ Стоп",
+  "batch_progress": "Прогресс пакетной обработки",
+  "batch_current_file": "Сейчас обрабатывается",
+  "batch_completed": "✅ Пакетная обработка завершена!",
+  "batch_stopped": "⏹️ Пакетная обработка остановлена",
+  "batch_no_files": "❌ Файлы для обработки не выбраны",
+  "source_model": "Исходная модель (Вокал)",
+  "source_model_info": "Модель для извлечения вокала/фазовых данных",
+  "target_model": "Целевая модель (Инструментал)",
+  "target_model_info": "Модель для применения коррекции фазы",
+  "custom_models_tab": "⚡ Пользовательские модели",
+  "Custom Models": "Пользовательские модели",
+  "add_custom_model": "➕ Добавить пользовательскую модель",
+  "custom_model_name": "Название модели",
+  "custom_model_name_placeholder": "Введите название для вашей модели",
+  "checkpoint_url": "URL чекпоинта (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "URL конфигурации (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "URL пользовательского Python файла (необязательно)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "Автоматически определить тип модели",
+  "model_type": "Тип модели",
+  "add_model_btn": "➕ Добавить модель",
+  "custom_models_list": "📋 Список пользовательских моделей",
+  "delete_model": "🗑️ Удалить",
+  "no_custom_models": "Пользовательские модели ещё не добавлены",
+  "model_added_success": "✅ Модель успешно добавлена!",
+  "model_deleted_success": "✅ Модель успешно удалена!",
+  "model_add_error": "❌ Ошибка добавления модели: {}",
+  "model_delete_error": "❌ Ошибка удаления модели: {}",
+  "refresh_models": "🔄 Обновить",
+  "custom_model_info": "Добавляйте пользовательские модели, предоставив ссылки для скачивания. Модели будут скачаны автоматически при использовании.",
+  "select_model_to_delete": "Выберите модель для удаления"
 }

assets/i18n/languages/tr_tr.json

@@ -19,6 +19,7 @@
   "tta_boost": "TTA Güçlendirme",
   "tta_info": "Kaliteyi artırır ancak işlem hızını yavaşlatır",
   "phase_fix": "Faz Düzeltme",
+  "use_phase_correction": "Gelişmiş Faz Düzeltme Uygula",
   "phase_fix_info": "Enstrümantal parçalar için gelişmiş faz düzeltme",
   "instrumental": "Enstrümantal",
   "instrumental_info": "Genellikle 2 çıktı verir, ama bazen tek çıktı olabilir, denemen lazım! :)",
@@ -48,7 +49,7 @@
   "files_will_be_copied": "Dosyalar buraya kopyalanacak...",
   "processing_tip": "<b>🔈 İşleme İpucu:</b> Gürültülü sonuçlar için <i>\"Gürültü ve Efekt Kaldırma\"</i> kategorisindeki <code>bleed_suppressor_v1</code> veya <code>denoisedebleed</code> modellerini kullanarak çıkışı temizleyin",
   "waiting_for_processing": "İşleme Bekleniyor...",
-  "auto_ensemble_tab": "🤖 Otomatik birleştirme",
+  "auto_ensemble_tab": "🤖 Otomatik Birleştirme",
   "upload_file": "Dosya Yükle",
   "enter_file_path": "Veya Dosya Yolunu Gir",
   "file_path_placeholder": "Ses dosyasının tam yolunu gir",
@@ -61,16 +62,16 @@
   "model_selection": "🧠 Model Seçimi",
   "model_category": "Model Kategorisi",
   "select_models": "Kategoriden Model Seç",
-  "ensemble_settings": "⚡ birleştirme Ayarları",
+  "ensemble_settings": "⚡ Birleştirme Ayarları",
   "method": "Yöntem",
   "recommendation": "**Öneri:** avg_wave ve max_fft en iyi sonuçları verir",
   "start_processing": "🚀 İşlemi Başlat",
   "original_audio_tab": "🔊 Orijinal Ses",
   "original_audio": "Orijinal Ses",
-  "ensemble_result_tab": "🎚️ birleştirme Sonucu",
+  "ensemble_result_tab": "🎚️ Birleştirme Sonucu",
   "output_preview": "Çıktı Önizlemesi",
   "refresh_output": "🔄 Çıktıyı Yenile",
-  "ensemble_copy_status": "birleştirme çıktısı buraya kopyalanacak...",
+  "ensemble_copy_status": "Birleştirme çıktısı buraya kopyalanacak...",
   "guidelines": "<h4>Model Seçim Kılavuzu</h4><ul><li><strong>Kategorileri Karıştırmaktan Kaçının:</strong> Vokal ve enstrümantal modellerin birleşimi istenmeyen karışımlara yol açabilir</li><li><strong>Özel Modeller Hakkında Notlar:</strong><ul><li>İkili modeller (v1/v2) - Her iki stemi de üretir</li><li>MDX23C Ayırıcı - Hibrit sonuçlar</li></ul></li><li><strong>En İyi Uygulama:</strong> Aynı kategoriden 3-5 benzer model kullanın</li></ul><div>💡 Profesyonel İpucu: \"VOCALS-MelBand-Roformer BigBeta5e\" + \"VOCALS-BS-Roformer_1297\" kombinasyonuyla başlayın</div>",
   "download_sources_tab": "⬇️ Kaynakları İndir",
   "direct_links": "🌐 Doğrudan Bağlantılar",
@@ -79,22 +80,121 @@
   "download_status": "İndirme Durumu",
   "downloaded_file": "İndirilen Dosya",
   "cookie_management": "🍪 Çerez Yönetimi",
-  "upload_cookies_txt": "Cookies.txt Yükle",
-  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 Neden Gerekli?**<br>- Yaş sınırlı içeriklere erişim<br>- Özel/listelenmemiş videoları indirme<br>- Bölgesel kısıtlamaları aşma<br>- YouTube indirme limitlerini geçme<br><br>**⚠️ Önemli Notlar**<br>- Çerez dosyalarınızı ASLA paylaşmayın!<br>- Çerezleri şu durumlarda yenileyin:<br>  • \"403 Yasak\" hataları alırsanız<br>  • İndirmeler aniden durursa<br>  • \"Oturum Süresi Doldu\" mesajı görürseniz<br><br>**🔄 Yenileme Adımları**<br>1. Bu <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">Chrome uzantısını</a> yükleyin<br>2. Chrome'da YouTube'a giriş yapın<br>3. Uzantı simgesine tıklayın → \"Dışa Aktar\"<br>4. İndirilen dosyayı buraya yükleyin<br><br>**⏳ Çerez Ömrü**<br>- Normal oturumlar: 24 saat<br>- Hassas işlemler: 1 saat<br>- Şifre değişikliği: Anında geçersizlik</div>",
-  "manual_ensemble_tab": "🎚️ Manuel birleştirme",
+  "upload_cookies_txt": "Çerezler.txt Yükle",
+  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 Neden Gerekli?**<br>- Yaş kısıtlamalı içeriklere erişim<br>- Özel/listelenmemiş videoları indirme<br>- Bölgesel kısıtlamaları aşma<br>- YouTube indirme limitlerini aşma<br><br>**⚠️ Önemli Notlar**<br>- Çerez dosyalarınızı ASLA paylaşmayın!<br>- Çerezleri yenileyin:<br>  • \"403 Yasak\" hataları alıyorsanız<br>  • İndirmeler aniden duruyorsa<br>  • \"Oturum sona erdi\" mesajı görüyorsanız<br><br>**🔄 Yenileme Adımları**<br>1. Bu <a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">Chrome eklentisini</a> yükleyin<br>2. Chrome'da YouTube'a giriş yapın<br>3. Eklenti simgesine tıklayın → \"Dışa Aktar\"<br>4. İndirilen dosyayı buraya yükleyin<br><br>**⏳ Çerez Ömrü**<br>- Normal oturumlar: 24 saat<br>- Hassas işlemler: 1 saat<br>- Şifre değişikliği: Anında yenileme</div>",
+  "ensemble_output_copied": "✅ Birleştirme çıktısı {}'ya kopyalandı",
+  "error_copying_ensemble_output": "��� Birleştirme çıktısı kopyalama hatası: {}",
+  "drive_mounted_copying_ensemble": "Google Drive bağlandı. Birleştirme çıktısını kopyalıyor...",
+  "drive_already_mounted_copying_ensemble": "Google Drive zaten bağlı. Birleştirme çıktısını kopyalıyor...",
+  "no_ensemble_output_files_found": "❌ Birleştirme çıktı dosyası bulunamadı.",
+  "drive_mounted_copying_files": "Google Drive bağlandı. Dosyaları kopyalıyor...",
+  "drive_already_mounted_copying_files": "Google Drive zaten bağlı. Dosyaları kopyalıyor...",
+  "files_copied_to_drive": "✅ Dosyalar {}'ya kopyalandı",
+  "error_copying_files": "❌ Dosya kopyalama hatası: {}",
+  "mounting_drive": "Google Drive bağlanıyor...",
+  "settings_saved": "Ayarlar config.json'a kaydedildi: Dil={lang}, Yöntem={method}, Port={port}",
+  "selected_language": "Seçilen dil: {lang}",
+  "available_languages": "Kullanılabilir diller: {langs}",
+  "select_language_prompt": "Bir dil seçin (varsayılan: {default}): ",
+  "invalid_language": "Geçersiz dil seçimi! Varsayılan dil ({default}) kullanılıyor.",
+  "available_sharing_methods": "Kullanılabilir paylaşım yöntemleri: gradio, localtunnel, ngrok",
+  "select_sharing_method_prompt": "Bir paylaşım yöntemi seçin (varsayılan: {default}): ",
+  "enter_ngrok_token_prompt": "Ngrok token'ınızı girin (varsayılan: {default}): ",
+  "ngrok_token_required": "Hata: ngrok yöntemi için Ngrok token'ı gerekli!",
+  "enter_port_prompt": "Port numarasını girin (varsayılan: {default}): ",
+  "opening_previous_url": "Önceki URL açılıyor: {url}",
+  "open_url_manually": "Colab'da URL'yi manuel açın: {url}",
+  "invalid_previous_url": "Geçersiz önceki URL: {url}, açılmadı.",
+  "starting_method": "{method} port {port}'ta başlatılıyor...",
+  "process_stopped": "\n🛑 Kullanıcı tarafından işlem durduruldu",
+  "selected_port": "Seçilen port: {port}",
+  "starting_gradio_with_sharing": "Dahili paylaşım ile Gradio başlatılıyor...",
+  "starting_localtunnel": "Localtunnel port {port}'ta başlatılıyor...",
+  "share_link": "Paylaşım bağlantısı: {url}",
+  "password_ip": "Şifre IP: {ip}",
+  "starting_ngrok": "Ngrok port {port}'ta başlatılıyor...",
+  "ngrok_url": "Ngrok URL: {url}",
+  "ngrok_error": "Ngrok başlatma hatası: {error}",
+  "auto_apollo_chunk_size": "Apollo Parça Boyutu",
+  "auto_apollo_chunk_size_info": "Apollo iyileştirme için parça boyutu (evrensel model için önerilen: 19)",
+  "auto_apollo_overlap": "Apollo Çakışması",
+  "auto_enhancing_with_apollo": "Apollo ile iyileştiriliyor ({}/{} dosya)...",
+  "auto_apollo_processing_completed": "Apollo işleme tamamlandı!",
+  "auto_apollo_overlap_info": "Apollo iyileştirme için çakışma (önerilen: 2)",
+  "apollo_chunk_size": "Apollo Parça Boyutu",
+  "apollo_chunk_size_info": "Apollo iyileştirme için parça boyutu (evrensel model için önerilen: 19)",
+  "apollo_overlap": "Apollo Çakışması",
+  "enhance_with_apollo": "Apollo ile İyileştir",
+  "apollo_enhancement_info": "Ses kalitesi iyileştirme",
+  "enhancing_with_apollo": "Apollo ile iyileştiriliyor ({}/{} dosya)...",
+  "apollo_processing_completed": "Apollo işleme tamamlandı!",
+  "apollo_overlap_info": "Apollo iyileştirme için çakışma (önerilen: 2)",
+  "selected_models": "Seçilen Modeller",
+  "save_preset": "Ön Ayar Kaydet",
+  "delete_preset": "Ön Ayar Sil",
+  "refresh_presets": "Ön Ayarları Güncelle",
+  "preset_name": "Ön Ayar Adı",
+  "select_preset": "Ön Ayar Seç",
+  "add_favorite": "Favorilere Ekle",
+  "apply_matchering": "Matchering Uygula",
+  "matchering_info": "Ses masteringini ve dengesini iyileştirmek için Matchering uygula",
+  "matchering_passes": "Matchering Geçişleri",
+  "matchering_passes_info": "Matchering yineleme sayısı (1-5, yüksek değerler kaliteyi artırır ancak daha uzun sürer)",
+  "matchering_processing": "Matchering uygulanıyor... ({:.1f}%)",
+  "matchering_completed": "✅ Matchering işlemi tamamlandı!",
+  "matchering_error": "❌ Matchering sırasında hata: {}",
+  "invalid_method": "Hata: Geçersiz yöntem! 'gradio', 'localtunnel' veya 'ngrok' kullan.",
+  "language_changed_message": "Dil değiştirildi. Lütfen hücreyi yeniden başlat.",
+  "backend_settings": "🚀 Backend Ayarları",
+  "inference_backend": "Inference Backend",
+  "optimization_mode": "Optimizasyon Modu",
+  "optimization_mode_info": "PyTorch optimizasyon modu seçimi",
+  "mixed_precision_amp": "Karışık Hassasiyet (AMP)",
+  "mixed_precision_info": "2x daha hızlı inference - önerilir",
+  "tf32_acceleration": "TF32 Hızlandırma",
+  "tf32_acceleration_info": "RTX 30xx+ için ekstra hız artışı",
+  "cudnn_benchmark": "cuDNN Benchmark",
+  "cudnn_benchmark_info": "İlk çalışmada yavaş, sonraki çalışmalarda çok hızlı",
+  "ultra_optimized_pytorch": "ULTRA-OPTİMİZE PyTorch Backend",
+  "default_active_max_speed": "Varsayılan olarak aktif - Maximum hız optimizasyonu",
+  "channels_last_mode": "RTX GPUs için en hızlı",
+  "compile_mode": "PyTorch 2.0+ için ekstra hız",
+  "default_mode": "Standart",
+  "manual_ensemble_tab": "🎚️ Manuel Birleştirme",
+  "phase_fixer_tab": "🔧 Faz Düzeltici",
+  "batch_processing_tab": "📦 Toplu İşleme",
+  "batch_description": "Aynı model ile birden fazla dosyayı aynı anda işle",
+  "batch_input_folder": "Giriş Klasörü",
+  "batch_input_folder_placeholder": "Ses dosyalarını içeren klasör yolunu girin",
+  "batch_output_folder": "Çıkış Klasörü",
+  "batch_output_folder_placeholder": "Çıkış dosyaları için klasör yolunu girin",
+  "batch_file_list": "İşlenecek Dosyalar",
+  "batch_add_files": "Dosya Ekle",
+  "batch_clear_list": "Listeyi Temizle",
+  "batch_start": "🚀 Toplu İşlemeyi Başlat",
+  "batch_stop": "⏹️ Durdur",
+  "batch_progress": "Toplu İşleme İlerlemesi",
+  "batch_current_file": "Şu An İşlenen",
+  "batch_completed": "✅ Toplu işleme tamamlandı!",
+  "batch_stopped": "⏹️ Toplu işleme durduruldu",
+  "batch_no_files": "❌ İşlenecek dosya seçilmedi",
+  "source_model": "Kaynak Model (Vokal)",
+  "source_model_info": "Vokal/faz verisi çıkarmak için model",
+  "target_model": "Hedef Model (Enstrümantal)",
+  "target_model_info": "Faz düzeltmesi uygulanacak model",
   "input_sources": "📂 Giriş Kaynakları",
   "refresh": "🔄 Yenile",
-  "ensemble_algorithm": "birleştirme Algoritması",
+  "ensemble_algorithm": "Birleştirme Algoritması",
   "select_audio_files": "Ses Dosyalarını Seç",
   "available_files": "Mevcut Dosyalar",
   "custom_weights": "Özel Ağırlıklar (virgülle ayrılmış)",
   "custom_weights_placeholder": "Örnek: 0.8, 1.2, 1.0, ...",
   "custom_weights_info": "Eşit ağırlıklar için boş bırakın",
   "result_preview_tab": "🎧 Sonuç Önizlemesi",
-  "ensembled_output": "birleştirme Çıktısı",
+  "ensembled_output": "Birleştirme Çıktısı",
   "processing_log_tab": "📋 İşlem Günlüğü",
   "processing_details": "İşlem Ayrıntıları",
-  "process_ensemble": "⚡ birleştirmeyi İşle",
+  "process_ensemble": "⚡ Birleştirmeyi İşle",
   "languages_tab": "🌐 Diller",
   "Vocal Models": "Vokal Modeller",
   "Instrumental Models": "Enstrümantal Modeller",
@@ -102,12 +202,13 @@
   "Denoise Models": "Gürültü Giderme Modelleri",
   "Dereverb Models": "Yankı Giderme Modelleri",
   "Other Models": "Diğer Modeller",
+  "Other Shapes": "Diğer Modeller",
   "ensemble_files_description": "Belirtilen birleştirme yöntemlerini kullanarak birden fazla ses dosyasını tek bir çıktıda birleştirir.",
-  "ensemble_files_help": "birleştirme için tüm ses dosyalarına giden yol",
+  "ensemble_files_help": "Birleştirme için tüm ses dosyalarına giden yol",
   "ensemble_type_help": "Şunlardan biri: avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
-  "ensemble_weights_help": "birleştirme oluşturmak için ağırlıklar. Ağırlık sayısı dosya sayısına eşit olmalı",
-  "ensemble_output_help": "birleştirme sonucunun kaydedileceği WAV dosyasının yolu",
-  "ensemble_type_print": "birleştirme türü: {}",
+  "ensemble_weights_help": "Birleştirme oluşturmak için ağırlıklar. Ağırlık sayısı dosya sayısına eşit olmalı",
+  "ensemble_output_help": "Birleştirme sonucunun kaydedileceği WAV dosyasının yolu",
+  "ensemble_type_print": "Birleştirme türü: {}",
   "num_input_files_print": "Giriş dosyası sayısı: {}",
   "weights_print": "Ağırlıklar: {}",
   "output_file_print": "Çıktı dosyası: {}",
@@ -115,7 +216,7 @@
   "file_not_found_error": "Hata. Dosya bulunamadı: {}. Yolları kontrol et.",
   "reading_chunk_print": "Dosyadan parça okunuyor: {} (başlangıç: {}s, süre: {}s)",
   "chunk_result_shape_print": "Parça sonucu şekli: {}",
-  "ensemble_completed_print": "birleştirme tamamlandı. Çıktı şuraya kaydedildi: {}",
+  "ensemble_completed_print": "Birleştirme tamamlandı. Çıktı şuraya kaydedildi: {}",
   "file_deletion_error": "{} silinemedi: {}",
   "directory_not_exist_warning": "⚠️ Dizin mevcut değil: {}",
   "not_a_directory_warning": "⚠️ Yol bir dizin değil: {}",
@@ -157,10 +258,18 @@
   "invalid_url": "❌ Geçersiz URL",
   "cookie_file_updated": "✅ Çerez dosyası güncellendi!",
   "cookie_installation_error": "⚠️ Çerez yükleme hatası: {}",
+  "google_drive_mount_error": "❌ Google Drive bağlantı hatası: {}",
   "file_size_zero_error": "Dosya boyutu sıfır veya dosya oluşturulmadı",
   "google_drive_error": "❌ Google Drive indirme hatası: {}",
   "wav_conversion_failed": "WAV dönüşümü başarısız",
   "download_error": "❌ İndirme hatası: {}",
+  "file_copied_to_drive": "Dosya Google Drive'a kopyalandı: {}",
+  "copy_to_drive_error": "❌ Google Drive'a kopyalama hatası: {}",
+  "direct_download_failed": "Dosyanın doğrudan indirilmesi başarısız",
+  "direct_download_error": "❌ Doğrudan indirme hatası: {}",
+  "invalid_google_drive_url": "❌ Geçersiz Google Drive URL'si",
+  "continuing_without_google_drive": "Google Drive bağlantısı yapılamadı, işlem devam ediyor...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive bağlı değil, Drive'a kopyalama atlanıyor...",
   "download_success": "🎉 Başarıyla indirildi!",
   "download_failed": "❌ İndirme başarısız",
   "no_output_files_found": "❌ Çıktı dosyaları bulunamadı.",
@@ -183,13 +292,13 @@
   "error_occurred_progress_label": "Hata oluştu -- 0.0%",
   "minimum_files_required": "⚠️ En az 2 dosya gerekli",
   "valid_files_not_found": "❌ Geçerli dosya bulunamadı",
-  "starting_ensemble_process": "birleştirme işlemi başlıyor...",
-  "ensembling_progress": "birleştirme yapılıyor... ({:.1f}%)",
-  "finalizing_ensemble_output": "birleştirme çıktısı tamamlanıyor...",
+  "starting_ensemble_process": "Birleştirme işlemi başlıyor...",
+  "ensembling_progress": "Birleştirme yapılıyor... ({:.1f}%)",
+  "finalizing_ensemble_output": "Birleştirme çıktısı tamamlanıyor...",
   "success_log": "✅ Başarılı!\n{}",
   "error_log": "❌ Hata!\n{}",
   "critical_error": "⛔ Kritik hata: {}",
-  "ensemble_process_completed": "birleştirme işlemi tamamlandı!",
+  "ensemble_process_completed": "Birleştirme işlemi tamamlandı!",
   "no_models_selected": "❌ Model seçilmedi",
   "no_input_audio_provided": "❌ Giriş sesi sağlanmadı",
   "loading_model": "Model yükleniyor {}/{}: {}...",
@@ -202,100 +311,69 @@
   "model_output_failed": "{} çıktı üretemedi",
   "waiting_for_files": "Tüm dosyaların hazır olması bekleniyor...",
   "waiting_for_files_progress_label": "Tüm dosyaların hazır olması bekleniyor... -- 90.0%",
-  "performing_ensemble": "birleştirme gerçekleştiriliyor...",
-  "performing_ensemble_progress_label": "birleştirme gerçekleştiriliyor... -- 92.0%",
-  "memory_usage_before_ensemble": "birleştirme öncesi bellek kullanımı: {}%",
-  "memory_usage_after_ensemble": "birleştirme sonrası bellek kullanımı: {}%",
-  "finalizing_ensemble_output_progress_label": "birleştirme çıktısı tamamlanıyor... -- 98.0%",
-  "ensemble_file_creation_failed": "birleştirme dosyası oluşturulamadı: {}",
-  "ensemble_completed_progress_label": "birleştirme başarıyla tamamlandı! -- 100.0%",
+  "performing_ensemble": "Birleştirme gerçekleştiriliyor...",
+  "performing_ensemble_progress_label": "Birleştirme gerçekleştiriliyor... -- 92.0%",
+  "memory_usage_before_ensemble": "Birleştirme öncesi bellek kullanımı: {}%",
+  "memory_usage_after_ensemble": "Birleştirme sonrası bellek kullanımı: {}%",
+  "finalizing_ensemble_output_progress_label": "Birleştirme çıktısı tamamlanıyor... -- 98.0%",
+  "ensemble_file_creation_failed": "Birleştirme dosyası oluşturulamadı: {}",
+  "ensemble_completed_progress_label": "Birleştirme başarıyla tamamlandı! -- 100.0%",
   "success_output_created": "✅ Başarılı! Çıktı dosyası oluşturuldu.",
-  "drive_mounted_copying_ensemble": "Google Drive bağlandı. birleştirme çıktısı kopyalanıyor...",
-  "drive_already_mounted_copying_ensemble": "Google Drive zaten bağlı. birleştirme çıktısı kopyalanıyor...",
-  "no_ensemble_output_files_found": "❌ birleştirme çıktı dosyaları bulunamadı.",
-  "ensemble_output_copied": "✅ birleştirme çıktısı {}'e kopyalandı",
-  "error_copying_ensemble_output": "❌ birleştirme çıktısı kopyalanırken hata: {}",
-  "drive_mounted_copying_files": "Google Drive bağlandı. Dosyalar kopyalanıyor...",
-  "drive_already_mounted_copying_files": "Google Drive zaten bağlı. Dosyalar kopyalanıyor...",
-  "files_copied_to_drive": "✅ Dosyalar {}'e kopyalandı",
-  "error_copying_files": "❌ Dosyalar kopyalanırken hata: {}",
-  "mounting_drive": "Google Drive bağlanıyor...",
-  "settings_saved": "Ayarlar config.json'a kaydedildi: Dil={lang}, Yöntem={method}, Port={port}",
-  "selected_language": "Seçilen dil: {lang}",
-  "availa ble_languages": "Kullanılabilir diller: {langs}",
-  "select_language_prompt": "Bir dil seç (varsayılan: {default}): ",
-  "invalid_language": "Geçersiz dil seçimi! Varsayılan dil kullanılıyor ({default}).",
-  "available_sharing_methods": "Kullanılabilir paylaşım yöntemleri: gradio, localtunnel, ngrok",
-  "select_sharing_method_prompt": "Bir paylaşım yöntemi seç (varsayılan: {default}): ",
-  "enter_ngrok_token_prompt": "Ngrok token'ını gir (varsayılan: {default}): ",
-  "ngrok_token_required": "Hata: ngrok yöntemi için Ngrok token'ı gerekli!",
-  "enter_port_prompt": "Port numarasını gir (varsayılan: {default}): ",
-  "opening_previous_url": "Önceki URL açılıyor: {url}",
-  "open_url_manually": "URL'yi Colab'da manuel olarak aç: {url}",
-  "invalid_previous_url": "Geçersiz önceki URL: {url}, açılmadı.",
-  "starting_method": "{method} port {port} üzerinde başlatılıyor...",
-  "process_stopped": "\n🛑 Kullanıcı tarafından işlem durduruldu",
-  "selected_port": "Seçilen port: {port}",
-  "starting_gradio_with_sharing": "Gradio dahili paylaşım ile başlatılıyor...",
-  "starting_localtunnel": "Localtunnel port {port} üzerinde başlatılıyor...",
-  "share_link": "Paylaşım bağlantısı: {url}",
-  "password_ip": "Parola IP: {ip}",
-  "starting_ngrok": "Ngrok port {port} üzerinde başlatılıyor...",
-  "ngrok_url": "Ngrok URL: {url}",
-  "ngrok_error": "Ngrok başlatma hatası: {error}",
-  "auto_apollo_chunk_size": "Apollo Parça Boyutu",
-  "auto_apollo_chunk_size_info": "Apollo iyileştirme için parça boyutu (evrensel model için önerilen: 19)",
-  "auto_apollo_overlap": "Apollo Çakışması",
-  "auto_enhancing_with_apollo": "Apollo ile iyileştiriliyor ({}/{} dosya)...",
-  "auto_apollo_processing_completed": "Apollo işleme tamamlandı!",
-  "autoapollo_overlap_info": "Apollo iyileştirme için çakışma (önerilen: 2)",
-  "auto_invalid_method": "Hata: Geçersiz yöntem! 'gradio', 'localtunnel' veya 'ngrok' kullan.",
-  "auto_apollo_enhancement_settings": "Apollo İyileştirme Ayarları",
-  "auto_enhance_with_apollo": "Apollo ile İyileştir",
-  "auto_enhance_with_apollo_info": "Ayrıştırmadan sonra ses iyileştirme için Apollo'yu etkinleştir",
-  "auto_apollo_method": "Apollo Yöntemi",
-  "auto_apollo_method_info": "Apollo için işlem yöntemini seç",
-  "auto_normal_method": "Normal Yöntem",
-  "auto_mid_side_method": "Orta/Yan Yöntem",
-  "auto_apollo_normal_model": "Apollo Normal Model",
-  "auto_apollo_normal_model_info": "Normal Apollo işleme için kullanılacak model",
-  "auto_apollo_midside_model": "Apollo Orta/Yan Model",
-  "auto_apollo_enhancement_info": "Ses kalitesi iyileştirme",
-  "auto_apollo_midside_model_info": "Orta/yan işleme için kullanılacak model (isteğe bağlı)",
-  "apollo_chunk_size": "Apollo Parça Boyutu",
-  "apollo_chunk_size_info": "Apollo iyileştirme için parça boyutu (evrensel model için önerilen: 19)",
-  "apollo_overlap": "Apollo Çakışması",
-  "enhancing_with_apollo": "Apollo ile iyileştiriliyor ({}/{} dosya)...",
-  "apollo_processing_completed": "Apollo işleme tamamlandı!",
-  "apollo_overlap_info": "Apollo iyileştirme için çakışma (önerilen: 2)",
-  "invalid_method": "Hata: Geçersiz yöntem! 'gradio', 'localtunnel' veya 'ngrok' kullan.",
-  "apollo_enhancement_settings": "Apollo İyileştirme Ayarları",
-  "enhance_with_apollo": "Apollo ile İyileştir",
-  "enhance_with_apollo_info": "Ayrıştırmadan sonra ses iyileştirme için Apollo'yu etkinleştir",
-  "apollo_method": "Apollo Yöntemi",
-  "apollo_method_info": "Apollo için işlem yöntemini seç",
-  "normal_method": "Normal Yöntem",
-  "mid_side_method": "Orta/Yan Yöntem",
-  "apollo_normal_model": "Apollo Normal için Model",
-  "apollo_normal_model_info": "Normal Apollo işleme için kullanılacak model",
-  "apollo_midside_model": "Apollo Orta/Yan Model",
-  "apollo_enhancement_info": "Ses kalitesi iyileştirme",
-  "apollo_processing_method": "Apollo işlem metodu",
-  "apollo_mid_side_model": "Apollo orta/yan için modeller",
-  "selected_models": "Seçilecek modeller",
-  "save_preset": "Ön ayarı kaydet",
-  "delete_preset": "Ön ayyarı sil",
-  "refresh_presets": "Ön ayarları güncelle",
-  "preset_name": "ön ayar ismi",
-  "select_preset": "Ön ayarlar",
-  "add_favorite": "favorilere ekle",
-  "apollo_midside_model_info": "Orta/yan işleme için kullanılacak model (isteğe bağlı)",
-  "language_changed_message": "Dil değiştirildi. Lütfen hücreyi yeniden başlat.",
-  "apply_matchering": "Matchering Uygula",
-  "matchering_info": "Ses masteringini ve dengesini iyileştirmek için Matchering uygula",
-  "matchering_passes": "Matchering Geçişleri",
-  "matchering_passes_info": "Matchering yineleme sayısı (1-5, yüksek değerler kaliteyi artırır ancak daha uzun sürer)",
-  "matchering_processing": "Matchering uygulanıyor... ({:.1f}%)",
-  "matchering_completed": "✅ Matchering işlemi tamamlandı!",
-  "matchering_error": "❌ Matchering sırasında hata: {}"
+  "custom_models_tab": "⚡ Özel Modeller",
+  "Custom Models": "Özel Modeller",
+  "add_custom_model": "➕ Özel Model Ekle",
+  "custom_model_name": "Model Adı",
+  "custom_model_name_placeholder": "Modeliniz için bir isim girin",
+  "checkpoint_url": "Checkpoint URL (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "Config URL (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "Özel Python Dosyası URL (isteğe bağlı)",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "Model tipini otomatik algıla",
+  "model_type": "Model Tipi",
+  "add_model_btn": "➕ Model Ekle",
+  "custom_models_list": "📋 Özel Modeller Listesi",
+  "delete_model": "🗑️ Sil",
+  "no_custom_models": "Henüz özel model eklenmedi",
+  "model_added_success": "✅ Model başarıyla eklendi!",
+  "model_deleted_success": "��� Model başarıyla silindi!",
+  "model_add_error": "❌ Model ekleme hatası: {}",
+  "model_delete_error": "❌ Model silme hatası: {}",
+  "refresh_models": "🔄 Yenile",
+  "custom_model_info": "İndirme linkleri sağlayarak özel model ekleyin. Modeller kullanıldığında otomatik olarak indirilecektir.",
+  "select_model_to_delete": "Silmek için bir model seçin",
+  "downloading_model_file": "Model indiriliyor: {}",
+  "downloading_file_progress": "İndiriliyor: {} - {}%",
+  "chunk_size_mode": "Parça Boyutu Modu",
+  "chunk_size_mode_info": "base: standart ön ayarlar | custom: kendi değerinizi girin | yaml: model YAML'ından oku",
+  "chunk_size_custom_label": "Özel Parça Boyutu",
+  "chunk_size_custom_info": "Özel bir parça boyutu değeri girin",
+  "chunk_size_yaml_label": "YAML Parça Boyutu",
+  "chunk_size_yaml_display_info": "İndirmeden sonra modelin YAML yapılandırmasından okunacak",
+  "chunk_size_yaml_not_downloaded": "⏳ Model YAML'ı henüz indirilmedi — işlem sırasında değer okunacak",
+  "chunk_size_yaml_detected": "✅ YAML'dan: {}",
+  "remove_favorite": "Favorilerden Çıkar",
+  "normal_method": "Normal yöntem",
+  "mid_side_method": "Orta-Yan yöntemi",
+  "apollo_processing_method": "Apollo İşlem Yöntemi",
+  "apollo_normal_model": "Apollo Normal Model",
+  "apollo_mid_side_model": "Apollo Orta-Yan Model",
+  "scale_factor": "Ölçek Faktörü",
+  "scale_factor_info": "Faz ölçeklendirme faktörü (önerilen: 1.4)",
+  "source_file_label": "Kaynak Dosya (Vokal)",
+  "target_file_label": "Hedef Dosya (Enstrümantal)",
+  "run_phase_fixer": "🔧 Faz Düzelticiyi Çalıştır",
+  "phase_fixed_output": "Faz Düzeltilmiş Çıktı",
+  "phase_fixer_settings": "Faz Düzeltici Ayarları",
+  "low_cutoff": "Düşük Kesme Frekansı (Hz)",
+  "high_cutoff": "Yüksek Kesme Frekansı (Hz)",
+  "low_cutoff_info": "Faz düzeltmesi için düşük frekans kesme",
+  "high_cutoff_info": "Faz düzeltmesi için yüksek frekans kesme",
+  "no_preset_name_provided": "⚠️ Ön ayar adı girilmedi",
+  "no_models_selected_for_preset": "⚠️ Ön ayar için model seçilmedi",
+  "preset_saved": "✅ '{}' ön ayarı kaydedildi!",
+  "please_upload_both_files": "⚠️ Lütfen hem kaynak hem de hedef dosyaları yükleyin",
+  "processing_log_placeholder": "İşlem günlüğü burada görünecek...",
+  "enter_preset_name": "Ön ayar adını girin..."
 }

assets/i18n/languages/zn_cn.json

@@ -19,6 +19,7 @@
   "tta_boost": "TTA增强",
   "tta_info": "提高质量，但会减慢处理速度",
   "phase_fix": "相位修正",
+  "use_phase_correction": "应用高级相位修正",
   "phase_fix_info": "用于乐器轨的高级相位修正",
   "instrumental": "乐器",
   "instrumental_info": "通常输出2个结果，但有时只有一个，试试看吧！:)",
@@ -46,7 +47,7 @@
   "copy_to_drive": "📂 复制到云盘",
   "copy_status": "复制状态",
   "files_will_be_copied": "文件将复制到此处...",
-  "processing_tip": "<b>🔈 处理提示：</b> 对于有噪声的结果，使用<i>“去除噪声和效果”</i>类别中的<code>bleed_suppressor_v1</code>或<code>denoisedebleed</code>模型来清理输出",
+  "processing_tip": "<b>🔈 处理提示：</b> 对于有噪声的结果，使用<i>去除噪声和效果</i>类别中的<code>bleed_suppressor_v1</code>或<code>denoisedebleed</code>模型来清理输出",
   "waiting_for_processing": "等待处理...",
   "auto_ensemble_tab": "🤖 自动集成",
   "upload_file": "上传文件",
@@ -80,16 +81,94 @@
   "downloaded_file": "已下载文件",
   "cookie_management": "🍪 Cookie管理",
   "upload_cookies_txt": "上传Cookies.txt",
-  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 为什么需要？**<br>- 访问年龄限制内容<br>- 下载私人/未列出的视频<br>- 绕过地区限制<br>- 避免YouTube下载限制<br><br>**⚠️ 重要提示**<br>- 切勿分享您的cookie文件！<br>- 在以下情况更新cookie：<br>  • 出现“403禁止”错误<br>  • 下载突然停止<br>  • 看到“会话过期”消息<br><br>**🔄 更新步骤**<br>1. 安装此<a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">Chrome扩展</a><br>2. 在Chrome中登录YouTube<br>3. 点击扩展图标 → “导出”<br>4. 将下载的文件上传到此处<br><br>**⏳ Cookie寿命**<br>- 普通会话：24小时<br>- 敏感操作：1小时<br>- 密码更改：立即失效</div>",
+  "cookie_info": "<div style=\"margin-left:15px; font-size:0.95em\">**📌 为什么必要？**<br>- 访问年龄限制内容<br>- 下载私人/未列出视频<br>- 绕过区域限制<br>- 避免YouTube下载限制<br><br>**⚠️ 重要说明**<br>- 切勿分享您的Cookie文件！<br>- 更新Cookie时：<br>  • 收到“403禁止”错误<br>  • 下载突然停止<br>  • 看到“会话过期”消息<br><br>**🔄 更新步骤**<br>1. 安装此<a href=\"https://chromewebstore.google.com/detail/get-cookiestxt-clean/ahmnmhfbokciafffnknlekllgcnafnie\" target=\"_blank\">Chrome扩展</a><br>2. 在Chrome中登录YouTube<br>3. 点击扩展图标 → “导出”<br>4. 在此处上传下载的文件<br><br>**⏳ Cookie寿命**<br>- 正常会话：24小时<br>- 敏感操作：1小时<br>- 密码更改：立即更新</div>",
+  "ensemble_output_copied": "✅ 集成输出已复制到 {}",
+  "error_copying_ensemble_output": "❌ 复制集成输出时出错：{}",
+  "drive_mounted_copying_ensemble": "Google Drive已挂载。正在复制集成输出...",
+  "drive_already_mounted_copying_ensemble": "Google Drive已挂载。正在复制集成输出...",
+  "no_ensemble_output_files_found": "❌ 未找到集成输出文件。",
+  "drive_mounted_copying_files": "Google Drive已挂载。正在复制文件...",
+  "drive_already_mounted_copying_files": "Google Drive已挂载。正在复制文件...",
+  "files_copied_to_drive": "✅ 文件已复制到 {}",
+  "error_copying_files": "❌ 复制文件时出错：{}",
+  "mounting_drive": "正在挂载Google Drive...",
+  "settings_saved": "设置已保存到config.json：语言={lang}，方法={method}，端口={port}",
+  "selected_language": "已选择语言：{lang}",
+  "available_languages": "可用语言：{langs}",
+  "select_language_prompt": "选择语言（默认：{default}）：",
+  "invalid_language": "无效语言选择！使用默认语言（{default}）。",
+  "available_sharing_methods": "可用共享方法：gradio, localtunnel, ngrok",
+  "select_sharing_method_prompt": "选择共享方法（默认：{default}）：",
+  "enter_ngrok_token_prompt": "输入您的Ngrok令牌（默认：{default}）：",
+  "ngrok_token_required": "错误：ngrok方法需要Ngrok令牌！",
+  "enter_port_prompt": "输入端口号（默认：{default}）：",
+  "opening_previous_url": "打开上一个URL：{url}",
+  "open_url_manually": "在Colab中手动打开URL：{url}",
+  "invalid_previous_url": "无效的上一个URL：{url}，未打开。",
+  "starting_method": "在端口 {port} 上启动 {method}...",
+  "process_stopped": "\n🛑 用户停止了进程",
+  "selected_port": "所选端口：{port}",
+  "starting_gradio_with_sharing": "启动带有内置共享的Gradio...",
+  "starting_localtunnel": "在端口 {port} 上启动Localtunnel...",
+  "share_link": "共享链接：{url}",
+  "password_ip": "密码IP：{ip}",
+  "starting_ngrok": "在端口 {port} 上启动Ngrok...",
+  "ngrok_url": "Ngrok URL：{url}",
+  "ngrok_error": "启动Ngrok时出错：{error}",
+  "auto_apollo_chunk_size": "Apollo分块大小",
+  "auto_apollo_chunk_size_info": "Apollo增强的分块大小（通用模型推荐：19）",
+  "auto_apollo_overlap": "Apollo重叠",
+  "auto_enhancing_with_apollo": "使用Apollo增强（{}/{} 文件）...",
+  "auto_apollo_processing_completed": "Apollo处理完成！",
+  "auto_apollo_overlap_info": "Apollo增强的重叠（推荐：2）",
+  "apollo_chunk_size": "Apollo分块大小",
+  "apollo_chunk_size_info": "Apollo增强的分块大小（通用模型推荐：19）",
+  "apollo_overlap": "Apollo重叠",
+  "enhance_with_apollo": "使用Apollo增强",
+  "apollo_enhancement_info": "音质增强",
+  "enhancing_with_apollo": "使用Apollo增强（{}/{} 文件）...",
+  "apollo_processing_completed": "Apollo处理完成！",
+  "apollo_overlap_info": "Apollo增强的重叠（推荐：2）",
+  "selected_models": "选定模型",
+  "save_preset": "保存预设",
+  "delete_preset": "删除预设",
+  "refresh_presets": "刷新预设",
+  "preset_name": "预设名称",
+  "select_preset": "选择预设",
+  "add_favorite": "添加到收藏",
+  "apply_matchering": "应用 Matchering",
+  "matchering_info": "应用 Matchering 以提升音频母带处理和平衡",
+  "matchering_passes": "Matchering 迭代次数",
+  "matchering_passes_info": "Matchering 迭代次数（1-5，值越高质量越好，但耗时更长）",
+  "matchering_processing": "正在应用 Matchering... ({:.1f}%)",
+  "matchering_completed": "✅ Matchering 处理完成！",
+  "matchering_error": "❌ Matchering 过程中发生错误：{}",
+  "invalid_method": "错误：无效方法！使用 'gradio'、'localtunnel' 或 'ngrok'。",
+  "language_changed_message": "语言已更改。请重新启动单元。",
+  "backend_settings": "🚀 后端设置",
+  "inference_backend": "推理后端",
+  "optimization_mode": "优化模式",
+  "optimization_mode_info": "PyTorch 优化模式选择",
+  "mixed_precision_amp": "混合精度 (AMP)",
+  "mixed_precision_info": "2x 更快推理 - 推荐",
+  "tf32_acceleration": "TF32 加速",
+  "tf32_acceleration_info": "RTX 30xx+ 的额外速度提升",
+  "cudnn_benchmark": "cuDNN 基准测试",
+  "cudnn_benchmark_info": "首次运行较慢，后续运行更快",
+  "ultra_optimized_pytorch": "超优化 PyTorch 后端",
+  "default_active_max_speed": "默认激活 - 最大速度优化",
+  "channels_last_mode": "RTX GPU 最快",
+  "compile_mode": "PyTorch 2.0+ 额外速度",
+  "default_mode": "标准",
   "manual_ensemble_tab": "🎚️ 手动集成",
-  "input_sources": "📂 输入来源",
+  "input_sources": "📂 输入源",
   "refresh": "🔄 刷新",
   "ensemble_algorithm": "集成算法",
   "select_audio_files": "选择音频文件",
   "available_files": "可用文件",
-  "custom_weights": "自定义权重（用逗号分隔）",
-  "custom_weights_placeholder": "示例：0.8, 1.2, 1.0, ...",
-  "custom_weights_info": "留空表示相等权重",
+  "custom_weights": "自定义权重（逗号分隔）",
+  "custom_weights_placeholder": "例如：0.8, 1.2, 1.0, ...",
+  "custom_weights_info": "留空以使用相等权重",
   "result_preview_tab": "🎧 结果预览",
   "ensembled_output": "集成输出",
   "processing_log_tab": "📋 处理日志",
@@ -99,32 +178,33 @@
   "Vocal Models": "人声模型",
   "Instrumental Models": "乐器模型",
   "4-Stem Models": "4声道模型",
-  "Denoise Models": "去噪模型",
+  "Denoise Models": "降噪模型",
   "Dereverb Models": "去混响模型",
   "Other Models": "其他模型",
-  "ensemble_files_description": "使用指定的集成方法将多个音频文件合并为一个输出。",
-  "ensemble_files_help": "所有用于集成的音频文件的路径",
-  "ensemble_type_help": "以下之一：avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
-  "ensemble_weights_help": "创建集成所需的权重。权重数量必须与文件数量匹配",
-  "ensemble_output_help": "保存集成结果的WAV文件路径",
+  "Other Shapes": "其他模型",
+  "ensemble_files_description": "使用指定的集成方法将多个音频文件合并为单个输出。",
+  "ensemble_files_help": "所有音频文件的路径用于集成",
+  "ensemble_type_help": "其中之一：avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft",
+  "ensemble_weights_help": "创建集成的权重。权重数量必须等于文件数量",
+  "ensemble_output_help": "存储集成结果的WAV文件路径",
   "ensemble_type_print": "集成类型：{}",
-  "num_input_files_print": "输入文件数量：{}",
+  "num_input_files_print": "输入文件数：{}",
   "weights_print": "权重：{}",
   "output_file_print": "输出文件：{}",
-  "duration_mismatch_error": "所有文件必须具有相同的时长",
-  "file_not_found_error": "错误。找不到文件：{}。请检查路径。",
-  "reading_chunk_print": "读取文件分块：{}（开始：{}秒，时长：{}秒）",
-  "chunk_result_shape_print": "分块结果形状：{}",
-  "ensemble_completed_print": "集成完成。输出保存至：{}",
-  "file_deletion_error": "{}无法删除：{}",
+  "duration_mismatch_error": "所有文件必须具有相同的持续时间",
+  "file_not_found_error": "错误。找不到文件：{}。检查路径。",
+  "reading_chunk_print": "从文件读取块：{}（开始：{}秒，持续时间：{}秒）",
+  "chunk_result_shape_print": "块结果形状：{}",
+  "ensemble_completed_print": "集成完成。输出保存到：{}",
+  "file_deletion_error": "无法删除 {}：{}",
   "directory_not_exist_warning": "⚠️ 目录不存在：{}",
   "not_a_directory_warning": "⚠️ 路径不是目录：{}",
-  "item_deletion_error": "⚠️ 删除{}时出错：{}",
+  "item_deletion_error": "⚠️ 删除 {} 时出错：{}",
   "old_output_not_exist": "❌ 旧输出文件夹不存在",
   "old_outputs_cleared": "✅ 旧输出已成功清除！",
   "error": "🔥 错误：{}",
-  "ffmpeg_error": "FFmpeg错误（{}）：{}",
-  "file_saved_successfully": "文件保存成功：{}",
+  "ffmpeg_error": "FFmpeg错误 ({})：{}",
+  "file_saved_successfully": "文件成功保存：{}",
   "total_files_found": "找到的文件总数：{}。使用的采样率：{}",
   "total_progress": "总进度",
   "detailed_pbar_enabled": "详细进度条已启用：{}",
@@ -132,149 +212,135 @@
   "cannot_read_track": "无法读取轨道：{}",
   "error_message": "错误消息：{}",
   "demudding_track": "清理轨道（相位混音 - 乐器）：{}",
-  "elapsed_time": "已用时间：{:.2f}秒。",
+  "elapsed_time": "耗时：{:.2f} 秒。",
   "proc_folder_description": "使用指定模型处理文件夹中的音频文件。",
-  "model_type_help": "模型类型（bandit, bs_roformer, mdx23c 等）",
+  "model_type_help": "模型类型（bandit、bs_roformer、mdx23c等）",
   "config_path_help": "配置文件路径",
   "demud_phaseremix_help": "为乐器启用相位混音",
   "start_checkpoint_help": "有效权重的初始检查点",
-  "input_folder_help": "包含待处理混合音频的文件夹",
-  "audio_path_help": "单个待处理音频文件的路径",
+  "input_folder_help": "包含要处理的混音的文件夹",
+  "audio_path_help": "要处理的单个音频文件的路径",
   "store_dir_help": "存储结果的路径",
   "device_ids_help": "GPU ID列表",
-  "extract_instrumental_help": "如提供，则反转人声以获取乐器",
-  "disable_detailed_pbar_help": "在分离中禁用详细进度条",
-  "force_cpu_help": "即使CUDA可用，也强制使用CPU",
-  "flac_file_help": "生成FLAC文件而非WAV",
+  "extract_instrumental_help": "如果提供，反转人声以获得乐器",
+  "disable_detailed_pbar_help": "在demix中禁用详细进度条",
+  "force_cpu_help": "即使CUDA可用也强制使用CPU",
+  "flac_file_help": "输出FLAC文件而不是WAV",
   "export_format_help": "导出格式和PCM类型",
   "pcm_type_help": "FLAC文件的PCM类型",
   "use_tta_help": "启用测试时间增强",
   "lora_checkpoint_help": "LoRA权重的初始检查点",
-  "cuda_available": "CUDA可用，使用 --force_cpu 禁用。",
-  "using_device": "使用的设备：{}",
+  "cuda_available": "CUDA可用，使用 --force_cpu 禁用它。",
+  "using_device": "使用设备：{}",
   "instruments_print": "乐器：{}",
-  "model_load_time": "模型加载时间：{:.2f}秒",
-  "invalid_url": "❌ 无效URL",
+  "model_load_time": "模型加载时间：{:.2f} 秒",
+  "invalid_url": "❌ 无效的URL",
   "cookie_file_updated": "✅ Cookie文件已更新！",
   "cookie_installation_error": "⚠️ Cookie安装错误：{}",
-  "file_size_zero_error": "文件大小为零或文件未创建",
+  "google_drive_mount_error": "❌ Google Drive连接错误：{}",
+  "file_size_zero_error": "文件大小为零或未创建文件",
   "google_drive_error": "❌ Google Drive下载错误：{}",
   "wav_conversion_failed": "WAV转换失败",
   "download_error": "❌ 下载错误：{}",
+  "file_copied_to_drive": "文件已复制到Google Drive：{}",
+  "copy_to_drive_error": "❌ 复制到Google Drive时出错：{}",
+  "direct_download_failed": "直接下载文件失败",
+  "direct_download_error": "❌ 直接下载错误：{}",
+  "invalid_google_drive_url": "❌ 无效的 Google Drive URL",
+  "continuing_without_google_drive": "Google Drive 连接失败，继续不使用它...",
+  "skipping_drive_copy_no_mount": "⚠️ Google Drive 未挂载，跳过复制到 Drive...",
   "download_success": "🎉 下载成功！",
   "download_failed": "❌ 下载失败",
   "no_output_files_found": "❌ 未找到输出文件。",
-  "output_refreshed_successfully": "✅ 输出刷新成功！",
+  "output_refreshed_successfully": "✅ 输出已成功刷新！",
   "error_refreshing_output": "❌ 刷新输出时出错：{}",
   "starting_audio_separation": "开始音频分离...",
-  "processing_audio": "处理音频",
-  "separating_audio": "分离音频... ({:.1f}%)",
+  "processing_audio": "正在处理音频",
+  "separating_audio": "正在分离音频... ({:.1f}%)",
   "separation_complete": "分离完成！",
   "progress_parsing_error": "进度解析错误：{}",
   "error_occurred": "发生错误：{}",
   "separation_process_completed": "分离过程完成！",
-  "no_audio_file_error": "❌ 未提供音频文件，且输入目录中没有现有文件。",
+  "no_audio_file_error": "❌ 未提供音频文件且输入目录中没有现有文件。",
   "no_input_progress_label": "错误：未提供输入 -- 0.0%",
-  "processing_audio_print": "正在处理音频：{}，使用模型：{}",
+  "processing_audio_print": "正在处理音频来自：{} 使用模型：{}",
   "starting_audio_separation_progress_label": "开始音频分离... -- 0.0%",
-  "separating_audio_progress_label": "分离音频... -- {}.0%",
-  "audio_processing_completed_progress_label": "音频处理完成！-- 100.0%",
+  "separating_audio_progress_label": "正在分离音频... -- {}.0%",
+  "audio_processing_completed_progress_label": "音频处理完成！ -- 100.0%",
   "audio_processing_completed": "✅ 音频处理完成！",
   "error_occurred_progress_label": "发生错误 -- 0.0%",
   "minimum_files_required": "⚠️ 至少需要2个文件",
   "valid_files_not_found": "❌ 未找到有效文件",
   "starting_ensemble_process": "开始集成过程...",
-  "ensembling_progress": "集成中... ({:.1f}%)",
-  "finalizing_ensemble_output": "完成集成输出...",
+  "ensembling_progress": "正在集成... ({:.1f}%)",
+  "finalizing_ensemble_output": "正在完成集成输出...",
   "success_log": "✅ 成功！\n{}",
   "error_log": "❌ 错误！\n{}",
   "critical_error": "⛔ 严重错误：{}",
   "ensemble_process_completed": "集成过程完成！",
   "no_models_selected": "❌ 未选择模型",
   "no_input_audio_provided": "❌ 未提供输入音频",
-  "loading_model": "加载模型 {}/{}：{}...",
-  "loading_model_progress_label": "加载模型 {}/{}：{}... -- {}.0%",
+  "loading_model": "正在加载模型 {}/{}：{}...",
+  "loading_model_progress_label": "正在加载模型 {}/{}：{}... -- {}.0%",
   "completed_model": "模型完成 {}/{}：{}",
   "completed_model_progress_label": "模型完成 {}/{}：{} -- {}.0%",
-  "running_command": "执行命令：{}",
+  "running_command": "正在运行命令：{}",
   "model_failed": "模型 {} 失败：{}",
-  "critical_error_with_model": "{} 出现严重错误：{}",
-  "model_output_failed": "{} 未能生成输出",
+  "critical_error_with_model": "{} 的严重错误：{}",
+  "model_output_failed": "{} 无法产生输出",
   "waiting_for_files": "等待所有文件准备就绪...",
   "waiting_for_files_progress_label": "等待所有文件准备就绪... -- 90.0%",
-  "performing_ensemble": "执行集成...",
-  "performing_ensemble_progress_label": "执行集成... -- 92.0%",
+  "performing_ensemble": "正在执行集成...",
+  "performing_ensemble_progress_label": "正在执行集成... -- 92.0%",
   "memory_usage_before_ensemble": "集成前内存使用：{}%",
   "memory_usage_after_ensemble": "集成后内存使用：{}%",
-  "finalizing_ensemble_output_progress_label": "完成集成输出... -- 98.0%",
-  "ensemble_file_creation_failed": "集成文件创建失败：{}",
-  "ensemble_completed_progress_label": "集成成功完成！-- 100.0%",
+  "finalizing_ensemble_output_progress_label": "正在完成集成输出... -- 98.0%",
+  "ensemble_file_creation_failed": "无法创建集成文件：{}",
+  "ensemble_completed_progress_label": "集成成功完成！ -- 100.0%",
   "success_output_created": "✅ 成功！输出文件已创建。",
-  "drive_mounted_copying_ensemble": "Google Drive已挂载。正在复制集成输出...",
-  "drive_already_mounted_copying_ensemble": "Google Drive已挂载。正在复制集成输出...",
-  "no_ensemble_output_files_found": "❌ 未找到集成输出文件。",
-  "ensemble_output_copied": "✅ 集成输出已复制到 {}",
-  "error_copying_ensemble_output": "❌ 复制集成输出时出错：{}",
-  "drive_mounted_copying_files": "Google Drive已挂载。正在复制文件...",
-  "drive_already_mounted_copying_files": "Google Drive已挂载。正在复制文件...",
-  "files_copied_to_drive": "✅ 文件已复制到 {}",
-  "error_copying_files": "❌ 复制文件时出错：{}",
-  "mounting_drive": "挂载Google Drive...",
-  "settings_saved": "设置已保存至 config.json：语言={lang}，方法={method}，端口={port}",
-  "selected_language": "所选语言：{lang}",
-  "available_languages": "可用语言：{langs}",
-  "select_language_prompt": "选择语言（默认：{default}）：",
-  "invalid_language": "无效的语言选择！使用默认语言（{default}）。",
-  "available_sharing_methods": "可用共享方法：gradio, localtunnel, ngrok",
-  "select_sharing_method_prompt": "选���共享方法（默认：{default}）：",
-  "enter_ngrok_token_prompt": "输入您的Ngrok令牌（默认：{default}）：",
-  "ngrok_token_required": "错误：ngrok方法需要Ngrok令牌！",
-  "enter_port_prompt": "输入端口号（默认：{default}）：",
-  "opening_previous_url": "打开上一个URL：{url}",
-  "open_url_manually": "在Colab中手动打开URL：{url}",
-  "invalid_previous_url": "无效的上一个URL：{url}，未打开。",
-  "starting_method": "在端口 {port} 上启动 {method}...",
-  "process_stopped": "\n🛑 用户停止了进程",
-  "selected_port": "所选端口：{port}",
-  "starting_gradio_with_sharing": "启动带有内置共享的Gradio...",
-  "starting_localtunnel": "在端口 {port} 上启动Localtunnel...",
-  "share_link": "共享链接：{url}",
-  "password_ip": "密码IP：{ip}",
-  "starting_ngrok": "在端口 {port} 上启动Ngrok...",
-  "ngrok_url": "Ngrok URL：{url}",
-  "ngrok_error": "启动Ngrok时出错：{error}",
-  "apollo_chunk_size": "Apollo分块大小",
-  "apollo_chunk_size_info": "Apollo改进的分块大小（通用模型推荐：19）",
-  "apollo_overlap": "Apollo重叠",
-  "enhancing_with_apollo": "使用Apollo增强（{}/{} 文件）...",
-  "apollo_processing_completed": "Apollo处理完成！",
-  "apollo_overlap_info": "Apollo改进的重叠（推荐：2）",
-  "invalid_method": "错误：无效方法！使用 'gradio'、'localtunnel' 或 'ngrok'。",
-  "apollo_enhancement_settings": "Apollo增强设置",
-  "enhance_with_apollo": "使用Apollo增强",
-  "enhance_with_apollo_info": "分离后启用Apollo以增强音频",
-  "apollo_method": "Apollo方法",
-  "apollo_method_info": "选择Apollo的处理方法",
-  "normal_method": "普通方法",
-  "mid_side_method": "中/侧方法",
-  "apollo_normal_model": "Apollo普通模型",
-  "apollo_normal_model_info": "用于普通Apollo处理的模型",
-  "apollo_midside_model": "Apollo中/侧模型",
-  "apollo_enhancement_info": "音质增强",
-  "selected_models": "选定模型",
-  "save_preset": "保存预设",
-  "delete_preset": "删除预设",
-  "refresh_presets": "刷新预设",
-  "preset_name": "预设名称",
-  "select_preset": "选择预设",
-  "add_favorite": "添加到收藏",
-  "apollo_midside_model_info": "用于中/侧处理的模型（可选）",
-  "language_changed_message": "语言已更改。请重新启动单元。",
-  "apply_matchering": "应用 Matchering",
-  "matchering_info": "应用 Matchering 以提升音频母带处理和平衡",
-  "matchering_passes": "Matchering 迭代次数",
-  "matchering_passes_info": "Matchering 迭代次数（1-5，值越高质量越好，但耗时更长）",
-  "matchering_processing": "正在应用 Matchering... ({:.1f}%)",
-  "matchering_completed": "✅ Matchering 处理完成！",
-  "matchering_error": "❌ Matchering 过程中发生错误：{}"
+  "phase_fixer_tab": "🔧 相位修复器",
+  "batch_processing_tab": "📦 批量处理",
+  "batch_description": "使用相同模型同时处理多个文件",
+  "batch_input_folder": "输入文件夹",
+  "batch_input_folder_placeholder": "输入包含音频文件的文件夹路径",
+  "batch_output_folder": "输出文件夹",
+  "batch_output_folder_placeholder": "输入输出文件的文件夹路径",
+  "batch_file_list": "待处理文件",
+  "batch_add_files": "添加文件",
+  "batch_clear_list": "清空列表",
+  "batch_start": "🚀 开始批量处理",
+  "batch_stop": "⏹️ 停止",
+  "batch_progress": "批量处理进度",
+  "batch_current_file": "当前处理中",
+  "batch_completed": "✅ 批量处理完成！",
+  "batch_stopped": "⏹️ 批量处理已停止",
+  "batch_no_files": "❌ 未选择要处理的文件",
+  "source_model": "源模型（人声）",
+  "source_model_info": "用于提取人声/相位数据的模型",
+  "target_model": "目标模型（乐器）",
+  "target_model_info": "应用相位校正的模型",
+  "custom_models_tab": "⚡ 自定义模型",
+  "Custom Models": "自定义模型",
+  "add_custom_model": "➕ 添加自定义模型",
+  "custom_model_name": "模型名称",
+  "custom_model_name_placeholder": "为您的模型输入名称",
+  "checkpoint_url": "检查点 URL (.ckpt/.pth)",
+  "checkpoint_url_placeholder": "https://huggingface.co/.../model.ckpt",
+  "config_url": "配置 URL (.yaml)",
+  "config_url_placeholder": "https://huggingface.co/.../config.yaml",
+  "custom_py_url": "自定义 Python 文件 URL（可选）",
+  "custom_py_url_placeholder": "https://huggingface.co/.../bs_roformer.py",
+  "auto_detect_type": "自动检测模型类型",
+  "model_type": "模型类型",
+  "add_model_btn": "➕ 添加模型",
+  "custom_models_list": "📋 自定义模型列表",
+  "delete_model": "🗑️ 删除",
+  "no_custom_models": "尚未添加自定义模型",
+  "model_added_success": "✅ 模型添加成功！",
+  "model_deleted_success": "✅ 模型删除成功！",
+  "model_add_error": "❌ 添加模型错误：{}",
+  "model_delete_error": "❌ 删除模型错误：{}",
+  "refresh_models": "🔄 刷新",
+  "custom_model_info": "通过提供下载链接添加自定义模型。模型在使用时将自动下载。",
+  "select_model_to_delete": "选择要删除的模型"
 }

benchmark_pytorch.py

@@ -0,0 +1,252 @@
+# coding: utf-8
+__author__ = 'PyTorch Optimization Benchmark Tool'
+
+import argparse
+import time
+import torch
+import numpy as np
+from utils import get_model_from_config
+from pytorch_backend import (
+    PyTorchBackend, 
+    PyTorchOptimizer, 
+    benchmark_pytorch_optimizations,
+    get_model_info
+)
+import sys
+
+
+def load_checkpoint(checkpoint_path: str, model, device: str):
+    """Load model from checkpoint."""
+    print(f"Loading checkpoint from: {checkpoint_path}")
+    
+    checkpoint = torch.load(checkpoint_path, map_location=device, weights_only=False)
+    
+    # Handle different checkpoint formats
+    if isinstance(checkpoint, dict):
+        if 'state_dict' in checkpoint:
+            state_dict = checkpoint['state_dict']
+        elif 'model' in checkpoint:
+            state_dict = checkpoint['model']
+        elif 'state' in checkpoint:
+            state_dict = checkpoint['state']
+        else:
+            state_dict = checkpoint
+    else:
+        state_dict = checkpoint
+    
+    model.load_state_dict(state_dict, strict=False)
+    model = model.eval().to(device)
+    
+    print("✓ Checkpoint loaded successfully")
+    return model
+
+
+def benchmark_optimization_modes(args):
+    """
+    Benchmark different PyTorch optimization modes.
+    """
+    parser = argparse.ArgumentParser(description="Benchmark PyTorch Optimization Modes")
+    parser.add_argument("--model_type", type=str, required=True, help="Model type")
+    parser.add_argument("--config_path", type=str, required=True, help="Config path")
+    parser.add_argument("--start_check_point", type=str, required=True, help="Checkpoint path (.ckpt)")
+    parser.add_argument("--device", type=str, default='cuda:0', help="Device")
+    parser.add_argument("--num_iterations", type=int, default=100, help="Number of benchmark iterations")
+    parser.add_argument("--warmup_iterations", type=int, default=10, help="Number of warmup iterations")
+    parser.add_argument("--chunk_size", type=int, default=None, help="Override chunk size (optional)")
+    parser.add_argument("--batch_size", type=int, default=1, help="Batch size")
+    
+    if args is None:
+        args = parser.parse_args()
+    else:
+        args = parser.parse_args(args)
+    
+    # Check device
+    if args.device.startswith('cuda') and not torch.cuda.is_available():
+        print("❌ CUDA is not available!")
+        return
+    
+    print("="*60)
+    print("PyTorch Optimization Benchmark Tool")
+    print("="*60)
+    print(f"Model Type: {args.model_type}")
+    print(f"Checkpoint: {args.start_check_point}")
+    print(f"Device: {args.device}")
+    print(f"Iterations: {args.num_iterations}")
+    print("="*60)
+    
+    # Load model
+    print("\n📦 Loading model...")
+    model, config = get_model_from_config(args.model_type, args.config_path)
+    model = load_checkpoint(args.start_check_point, model, args.device)
+    
+    # Get model info
+    model_info = get_model_info(model)
+    print(f"\n📊 Model Information:")
+    print(f"  Total Parameters: {model_info['total_parameters']:,}")
+    print(f"  Trainable Parameters: {model_info['trainable_parameters']:,}")
+    print(f"  Model Size: {model_info['model_size_mb']:.2f} MB")
+    print(f"  Device: {model_info['device']}")
+    print(f"  Dtype: {model_info['dtype']}")
+    
+    # Get chunk size
+    if args.chunk_size:
+        chunk_size = args.chunk_size
+    else:
+        chunk_size = config.audio.chunk_size
+    
+    num_channels = 2
+    input_shape = (args.batch_size, num_channels, chunk_size)
+    
+    print(f"\n📊 Test Configuration:")
+    print(f"  Batch Size: {args.batch_size}")
+    print(f"  Channels: {num_channels}")
+    print(f"  Chunk Size: {chunk_size}")
+    print(f"  Input Shape: {input_shape}")
+    
+    # Benchmark different optimization modes
+    print("\n" + "="*60)
+    print("Benchmarking Optimization Modes")
+    print("="*60)
+    
+    results = benchmark_pytorch_optimizations(
+        model=model,
+        input_shape=input_shape,
+        device=args.device,
+        num_iterations=args.num_iterations,
+        warmup_iterations=args.warmup_iterations
+    )
+    
+    # Display results
+    print("\n" + "="*60)
+    print("📈 Benchmark Results")
+    print("="*60)
+    
+    baseline = None
+    for mode, time_ms in results.items():
+        if time_ms is not None:
+            if baseline is None:
+                baseline = time_ms
+            speedup = baseline / time_ms if time_ms > 0 else 0
+            improvement = ((baseline - time_ms) / baseline) * 100 if baseline > 0 else 0
+            
+            print(f"\n{mode.upper()}:")
+            print(f"  Average Time: {time_ms:.2f} ms")
+            print(f"  Speedup: {speedup:.2f}x")
+            print(f"  Improvement: {improvement:.1f}%")
+    
+    print("\n" + "="*60)
+    
+    # Recommendations
+    print("\n💡 Recommendations:")
+    
+    if results.get('compile') and results['compile'] < results['default']:
+        print("  ✓ Use 'compile' mode for best performance (PyTorch 2.0+)")
+    elif results.get('channels_last') and results['channels_last'] < results['default']:
+        print("  ✓ Use 'channels_last' mode for better performance")
+    else:
+        print("  ✓ Default mode is optimal for your configuration")
+    
+    if args.device.startswith('cuda'):
+        print("  ✓ Enable TF32 for Ampere GPUs (RTX 30xx+)")
+        print("  ✓ Enable cuDNN benchmark for consistent input sizes")
+    
+    print("\n✅ Benchmark completed!")
+
+
+def test_optimization_modes(args):
+    """
+    Test different optimization modes with verification.
+    """
+    parser = argparse.ArgumentParser(description="Test PyTorch Optimization Modes")
+    parser.add_argument("--model_type", type=str, required=True, help="Model type")
+    parser.add_argument("--config_path", type=str, required=True, help="Config path")
+    parser.add_argument("--start_check_point", type=str, required=True, help="Checkpoint path (.ckpt)")
+    parser.add_argument("--device", type=str, default='cuda:0', help="Device")
+    
+    if args is None:
+        args = parser.parse_args()
+    else:
+        args = parser.parse_args(args)
+    
+    print("="*60)
+    print("PyTorch Optimization Mode Test")
+    print("="*60)
+    
+    # Load model
+    print("\n📦 Loading model...")
+    model, config = get_model_from_config(args.model_type, args.config_path)
+    model = load_checkpoint(args.start_check_point, model, args.device)
+    
+    chunk_size = config.audio.chunk_size
+    input_shape = (1, 2, chunk_size)
+    dummy_input = torch.randn(*input_shape).to(args.device)
+    
+    # Test each optimization mode
+    modes = ['default', 'compile', 'channels_last']
+    outputs = {}
+    
+    for mode in modes:
+        print(f"\n{'='*60}")
+        print(f"Testing: {mode}")
+        print('='*60)
+        
+        try:
+            backend = PyTorchBackend(device=args.device, optimize_mode=mode)
+            
+            if mode == 'jit':
+                backend.optimize_model(model, example_input=dummy_input, use_amp=True)
+            else:
+                backend.optimize_model(
+                    model, 
+                    use_amp=True,
+                    use_channels_last=(mode == 'channels_last')
+                )
+            
+            # Run inference
+            with torch.no_grad():
+                output = backend(dummy_input)
+            
+            outputs[mode] = output
+            print(f"✓ {mode} successful")
+            print(f"  Output shape: {output.shape}")
+            print(f"  Output range: [{output.min().item():.6f}, {output.max().item():.6f}]")
+            
+        except Exception as e:
+            print(f"✗ {mode} failed: {e}")
+            outputs[mode] = None
+    
+    # Verify outputs match
+    print("\n" + "="*60)
+    print("🔍 Output Verification")
+    print("="*60)
+    
+    baseline_key = 'default'
+    if baseline_key in outputs and outputs[baseline_key] is not None:
+        baseline_output = outputs[baseline_key]
+        
+        for mode, output in outputs.items():
+            if mode != baseline_key and output is not None:
+                diff = torch.abs(baseline_output - output)
+                max_diff = torch.max(diff).item()
+                mean_diff = torch.mean(diff).item()
+                
+                print(f"\n{mode} vs {baseline_key}:")
+                print(f"  Max difference: {max_diff:.6f}")
+                print(f"  Mean difference: {mean_diff:.6f}")
+                
+                if max_diff < 1e-3:
+                    print(f"  ✓ Outputs match within tolerance")
+                else:
+                    print(f"  ⚠ Warning: Large difference detected!")
+    
+    print("\n✅ Test completed!")
+
+
+if __name__ == "__main__":
+    import sys
+    
+    if len(sys.argv) > 1 and sys.argv[1] == 'test':
+        sys.argv.pop(1)
+        test_optimization_modes(None)
+    else:
+        benchmark_optimization_modes(None)

config_manager.py

@@ -3,7 +3,7 @@ import json
 from pathlib import Path
 
 # Define config directory in Google Drive
-CONFIG_DIR = "/home/user/app/SESA-Config"
+CONFIG_DIR = "/content/drive/MyDrive/SESA-Config"
 CONFIG_FILE = os.path.join(CONFIG_DIR, "config.json")
 
 def load_config():
@@ -14,6 +14,10 @@ def load_config():
             "chunk_size": 352800,
             "overlap": 2,
             "export_format": "wav FLOAT",
+            "optimize_mode": "channels_last",
+            "enable_amp": True,
+            "enable_tf32": True,
+            "enable_cudnn_benchmark": True,
             "auto_use_tta": False,
             "use_tta": False,
             "use_demud_phaseremix_inst": False,

download.py

@@ -53,8 +53,16 @@ def download_callback(url, download_type='direct', cookie_file=None):
     # Mount Google Drive (optional)
     if drive is not None:
         try:
-            drive.mount('/content/drive', force_remount=True)
-            drive_mounted = True
+            # Check if already mounted first
+            if os.path.exists('/content/drive/MyDrive'):
+                drive_mounted = True
+            else:
+                drive.mount('/content/drive', force_remount=True)
+                drive_mounted = True
+        except AttributeError as ae:
+            # Handle 'NoneType' object has no attribute 'kernel' error
+            print(f"Warning: Google Drive mount skipped (Colab kernel issue): {str(ae)}")
+            print(i18n("continuing_without_google_drive"))
         except Exception as e:
             print(i18n("google_drive_mount_error").format(str(e)))
             print(i18n("continuing_without_google_drive"))
@@ -113,34 +121,95 @@ def download_callback(url, download_type='direct', cookie_file=None):
 
     # 3. YouTube and other media links
     else:
-        ydl_opts = {
-            'format': 'bestaudio/best',
+        # First try: iOS/Android without cookies (best for bot protection bypass)
+        ydl_opts_nocookie = {
+            'format': 'ba[ext=m4a]/ba[ext=webm]/ba/b',
+            'outtmpl': os.path.join(INPUT_DIR, '%(title)s.%(ext)s'),
+            'postprocessors': [{
+                'key': 'FFmpegExtractAudio',
+                'preferredcodec': 'wav',
+                'preferredquality': '0'
+            }],
+            'nocheckcertificate': True,
+            'ignoreerrors': False,
+            'retries': 3,
+            'extractor_retries': 3,
+            'extractor_args': {
+                'youtube': {
+                    'player_client': ['ios', 'android'],
+                    'player_skip': ['webpage', 'configs']
+                }
+            },
+            'http_headers': {
+                'User-Agent': 'com.google.ios.youtube/19.09.3 (iPhone14,3; U; CPU iOS 15_6 like Mac OS X)',
+                'Accept-Language': 'en-US,en;q=0.9'
+            }
+        }
+        
+        # Second try: web client with cookies if available
+        ydl_opts_cookie = {
+            'format': 'ba[ext=m4a]/ba[ext=webm]/ba/b',
             'outtmpl': os.path.join(INPUT_DIR, '%(title)s.%(ext)s'),
             'postprocessors': [{
                 'key': 'FFmpegExtractAudio',
                 'preferredcodec': 'wav',
                 'preferredquality': '0'
             }],
-            'cookiefile': COOKIE_PATH if os.path.exists(COOKIE_PATH) else None,
+            'cookiefile': COOKIE_PATH,
             'nocheckcertificate': True,
-            'ignoreerrors': True,
-            'retries': 3
+            'ignoreerrors': False,
+            'retries': 3,
+            'extractor_retries': 3,
+            'extractor_args': {
+                'youtube': {
+                    'player_client': ['web', 'tv_embedded'],
+                    'player_skip': ['configs']
+                }
+            }
         }
+        
+        # Try without cookies first
+        info_dict = None
+        temp_path = None
+        
         try:
-            with yt_dlp.YoutubeDL(ydl_opts) as ydl:
+            with yt_dlp.YoutubeDL(ydl_opts_nocookie) as ydl:
                 info_dict = ydl.extract_info(url, download=True)
-                temp_path = ydl.prepare_filename(info_dict)
-                base_name = os.path.splitext(os.path.basename(temp_path))[0]
-                sanitized_base_name = sanitize_filename(base_name)
-                wav_path = os.path.join(INPUT_DIR, f"{sanitized_base_name}.wav")
-                temp_wav = os.path.splitext(temp_path)[0] + '.wav'
-                if os.path.exists(temp_wav):
-                    os.rename(temp_wav, wav_path)
-                    download_success = True
-                else:
-                    raise Exception(i18n("wav_conversion_failed"))
+                if info_dict:
+                    temp_path = ydl.prepare_filename(info_dict)
+        except Exception as e:
+            # If no cookies available or first method failed, try with cookies
+            if os.path.exists(COOKIE_PATH):
+                print(f"First attempt failed, trying with cookies...")
+                try:
+                    with yt_dlp.YoutubeDL(ydl_opts_cookie) as ydl:
+                        info_dict = ydl.extract_info(url, download=True)
+                        if info_dict:
+                            temp_path = ydl.prepare_filename(info_dict)
+                except Exception as e2:
+                    raise e2
+            else:
+                raise e
+        
+        try:
+            # Check if extraction was successful
+            if info_dict is None:
+                raise Exception(i18n("youtube_extraction_failed") if "youtube_extraction_failed" in dir(i18n) else "YouTube extraction failed. Please try updating yt-dlp: pip install -U yt-dlp")
+            
+            base_name = os.path.splitext(os.path.basename(temp_path))[0]
+            sanitized_base_name = sanitize_filename(base_name)
+            wav_path = os.path.join(INPUT_DIR, f"{sanitized_base_name}.wav")
+            temp_wav = os.path.splitext(temp_path)[0] + '.wav'
+            if os.path.exists(temp_wav):
+                os.rename(temp_wav, wav_path)
+                download_success = True
+            else:
+                raise Exception(i18n("wav_conversion_failed"))
         except Exception as e:
             error_msg = i18n("download_error").format(str(e))
+            # Add hint for yt-dlp update if it's a YouTube issue
+            if 'youtube' in url.lower() or 'youtu.be' in url.lower():
+                error_msg += "\n\nTry: pip install -U yt-dlp"
             print(error_msg)
             return None, error_msg, None, None, None, None
 

ensemble.py

@@ -1,281 +1,442 @@
+#!/usr/bin/env python3
 # coding: utf-8
-__author__ = 'Roman Solovyev (ZFTurbo): https://github.com/ZFTurbo/' # forked by jarredou
+"""
+Ultimate Audio Ensemble Processor v4.0
+- Tüm ensemble yöntemlerini destekler (avg_wave, median_wave, max_wave, min_wave, max_fft, min_fft, median_fft)
+- Özel karakterli ve uzun dosya yollarını destekler
+- Büyük dosyaları verimli şekilde işler
+- Detaylı hata yönetimi ve loglama
+"""
 
 import os
-import torch
-import torchaudio
-import soundfile as sf
+import sys
 import argparse
+import numpy as np
+import soundfile as sf
+import librosa
+import psutil
+import gc
+import traceback
+from scipy.signal import stft, istft
 from pathlib import Path
+import tempfile
+import shutil
+import json
+from tqdm import tqdm
+import time
+import torch
 
+# PyTorch optimizations
+if torch.cuda.is_available():
+    torch.backends.cudnn.benchmark = True
+    print("✓ Using CUDA acceleration for ensemble")
+else:
+    print("Using CPU for ensemble")
 
-def stft(wave, n_fft, hop_length):
-    """
-    Compute STFT for stereo audio using PyTorch
-    :param wave: tensor of shape (channels, length)
-    :param n_fft: FFT size
-    :param hop_length: hop length
-    :return: complex spectrogram of shape (channels, freq_bins, time_frames)
-    """
-    # Use torchaudio's STFT which handles multi-channel audio efficiently
-    spec = torch.stft(
-        wave,
-        n_fft=n_fft,
-        hop_length=hop_length,
-        window=torch.hann_window(n_fft, device=wave.device),
-        return_complex=True
-    )
-    return spec
-
-
-def istft(spec, hop_length, length):
-    """
-    Compute inverse STFT for stereo audio using PyTorch
-    :param spec: complex spectrogram of shape (channels, freq_bins, time_frames)
-    :param hop_length: hop length
-    :param length: target length
-    :return: waveform of shape (channels, length)
-    """
-    # Use torchaudio's ISTFT which handles multi-channel audio efficiently
-    wave = torch.istft(
-        spec,
-        n_fft=(spec.shape[1] - 1) * 2,
-        hop_length=hop_length,
-        window=torch.hann_window((spec.shape[1] - 1) * 2, device=spec.device),
-        length=length
-    )
-    return wave
-
-
-def absmax(a, *, dim):
-    """Find values with maximum absolute value along dimension"""
-    abs_a = torch.abs(a)
-    indices = torch.argmax(abs_a, dim=dim, keepdim=True)
-    return torch.gather(a, dim, indices).squeeze(dim)
-
-
-def absmin(a, *, dim):
-    """Find values with minimum absolute value along dimension"""
-    abs_a = torch.abs(a)
-    indices = torch.argmin(abs_a, dim=dim, keepdim=True)
-    return torch.gather(a, dim, indices).squeeze(dim)
-
-
-def lambda_max(arr, dim=None, key=None):
-    """Find elements with maximum key value along dimension"""
-    if key is None:
-        key = lambda x: x
-    key_values = key(arr)
-    indices = torch.argmax(key_values, dim=dim, keepdim=True)
-    return torch.gather(arr, dim, indices).squeeze(dim)
-
-
-def lambda_min(arr, dim=None, key=None):
-    """Find elements with minimum key value along dimension"""
-    if key is None:
-        key = lambda x: x
-    key_values = key(arr)
-    indices = torch.argmin(key_values, dim=dim, keepdim=True)
-    return torch.gather(arr, dim, indices).squeeze(dim)
-
-
-def match_tensor_shapes(tensor_1, tensor_2):
-    """Match the time dimension of two tensors by padding or trimming"""
-    target_length = tensor_2.shape[-1]
-    current_length = tensor_1.shape[-1]
-    
-    if current_length > target_length:
-        tensor_1 = tensor_1[..., :target_length]
-    elif current_length < target_length:
-        padding = target_length - current_length
-        tensor_1 = torch.nn.functional.pad(tensor_1, (0, padding), 'constant', 0)
-    
-    return tensor_1
-
-
-def average_waveforms(pred_track, weights, algorithm, n_fft, hop_length):
-    """
-    :param pred_track: tensor of shape (num, channels, length)
-    :param weights: tensor of shape (num,)
-    :param algorithm: One of avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft
-    :param n_fft: FFT size for STFT operations
-    :param hop_length: hop length for STFT operations
-    :return: averaged waveform in shape (channels, length)
-    """
-    pred_track = torch.stack(pred_track)
-    final_length = pred_track.shape[-1]
-    weights = torch.tensor(weights, dtype=pred_track.dtype, device=pred_track.device)
-
-    if algorithm in ['avg_wave', 'median_wave', 'min_wave', 'max_wave']:
-        # Waveform domain operations
-        if algorithm == 'avg_wave':
-            # Weighted average
-            weighted_tracks = pred_track * weights.view(-1, 1, 1)
-            result = torch.sum(weighted_tracks, dim=0) / torch.sum(weights)
-        elif algorithm == 'median_wave':
-            result = torch.median(pred_track, dim=0)[0]
-        elif algorithm == 'min_wave':
-            result = lambda_min(pred_track, dim=0, key=torch.abs)
-        elif algorithm == 'max_wave':
-            result = lambda_max(pred_track, dim=0, key=torch.abs)
-    
-    elif algorithm in ['avg_fft', 'median_fft', 'min_fft', 'max_fft']:
-        # Frequency domain operations
-        # Convert all tracks to spectrograms
-        spec_tracks = []
-        for i in range(pred_track.shape[0]):
-            spec = stft(pred_track[i], n_fft, hop_length)
-            spec_tracks.append(spec)
+class AudioEnsembleEngine:
+    def __init__(self):
+        self.temp_dir = None
+        self.log_file = "ensemble_processor.log"
         
-        spec_tracks = torch.stack(spec_tracks)
+    def __enter__(self):
+        self.temp_dir = tempfile.mkdtemp(prefix='audio_ensemble_')
+        self.setup_logging()
+        return self
         
-        if algorithm == 'avg_fft':
-            # Weighted average in frequency domain
-            weighted_specs = spec_tracks * weights.view(-1, 1, 1, 1)
-            avg_spec = torch.sum(weighted_specs, dim=0) / torch.sum(weights)
-            result = istft(avg_spec, hop_length, final_length)
-        elif algorithm == 'median_fft':
-            # Median in frequency domain (using magnitude and phase separately)
-            median_spec = torch.median(spec_tracks, dim=0)[0]
-            result = istft(median_spec, hop_length, final_length)
-        elif algorithm == 'min_fft':
-            min_spec = lambda_min(spec_tracks, dim=0, key=torch.abs)
-            result = istft(min_spec, hop_length, final_length)
-        elif algorithm == 'max_fft':
-            max_spec = absmax(spec_tracks, dim=0)
-            result = istft(max_spec, hop_length, final_length)
-    
-    return result
-
-
-def save_audio(waveform, sample_rate, output_path):
-    """Save audio with support for different formats and bit depths using soundfile"""
-    output_path = Path(output_path)
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.temp_dir and os.path.exists(self.temp_dir):
+            shutil.rmtree(self.temp_dir, ignore_errors=True)
     
-    # Ensure waveform is in the right format (channels, samples)
-    if waveform.dim() == 1:
-        waveform = waveform.unsqueeze(0)
+    def setup_logging(self):
+        """Initialize detailed logging system."""
+        with open(self.log_file, 'w') as f:
+            f.write("Audio Ensemble Processor Log\n")
+            f.write("="*50 + "\n")
+            f.write(f"System Memory: {psutil.virtual_memory().total/(1024**3):.2f} GB\n")
+            f.write(f"Python Version: {sys.version}\n\n")
     
-    # Convert to CPU numpy array if on GPU
-    if waveform.is_cuda:
-        waveform = waveform.cpu()
+    def log_message(self, message):
+        """Log messages with timestamp."""
+        with open(self.log_file, 'a') as f:
+            f.write(f"[{time.strftime('%Y-%m-%d %H:%M:%S')}] {message}\n")
     
-    # Convert to numpy and transpose to (samples, channels) for soundfile
-    audio_np = waveform.numpy().T
+    def normalize_path(self, path):
+        """Handle all path-related issues comprehensively."""
+        try:
+            # Convert to absolute path
+            path = str(Path(path).absolute().resolve())
+            
+            # Handle problematic characters
+            if any(char in path for char in '[]()|&; '):
+                base, ext = os.path.splitext(path)
+                safe_name = f"{hash(base)}{ext}"
+                temp_path = os.path.join(self.temp_dir, safe_name)
+                
+                if not os.path.exists(temp_path):
+                    data, sr = librosa.load(path, sr=None, mono=False)
+                    sf.write(temp_path, data.T, sr)
+                
+                return temp_path
+            
+            return path
+        except Exception as e:
+            self.log_message(f"Path normalization failed: {str(e)}")
+            return path
     
-    # Handle different output formats
-    if output_path.suffix.lower() == '.flac':
-        # For FLAC, convert to 24-bit
-        # Normalize to prevent clipping
-        max_val = torch.max(torch.abs(waveform))
-        if max_val > 0:
-            print(f"Clipping detected : {max_val}")
-            audio_np = audio_np / max_val.numpy()
+    def validate_inputs(self, files, method, output_path):
+        """Comprehensive input validation with detailed error reporting."""
+        errors = []
+        valid_methods = [
+            'avg_wave', 'median_wave', 'max_wave', 'min_wave',
+            'max_fft', 'min_fft', 'median_fft'
+        ]
         
-        # Save as FLAC with 24-bit depth
-        sf.write(str(output_path), audio_np, sample_rate, subtype='PCM_24')
-    else:
-        # Default to float32 for WAV and other formats
-        sf.write(str(output_path), audio_np, sample_rate, subtype='FLOAT')
-
-
-def ensemble_files(args):
-    parser = argparse.ArgumentParser(description="Audio ensemble tool using PyTorch")
-    parser.add_argument("--files", type=str, required=True, nargs='+', 
-                       help="Path to all audio-files to ensemble")
-    parser.add_argument("--type", type=str, default='avg_wave', 
-                       help="One of avg_wave, median_wave, min_wave, max_wave, avg_fft, median_fft, min_fft, max_fft")
-    parser.add_argument("--weights", type=float, nargs='+', 
-                       help="Weights to create ensemble. Number of weights must be equal to number of files")
-    parser.add_argument("--output", default="res.wav", type=str, 
-                       help="Path to output file (supports .wav, .flac)")
-    parser.add_argument("--n_fft", type=int, default=2048, 
-                       help="FFT size for STFT operations (default: 2048)")
-    parser.add_argument("--hop_length", type=int, default=1024, 
-                       help="Hop length for STFT operations (default: 1024)")
-    parser.add_argument("--device", type=str, default="auto", 
-                       help="Device to use: 'cpu', 'cuda', or 'auto' (default: auto)")
-    parser.add_argument("--trim_to_shortest", action="store_true", 
-                       help="Trim output to shortest input file length to avoid padding artifacts")
-    
-    if args is None:
-        args = parser.parse_args()
-    else:
-        args = parser.parse_args(args)
-
-    # Set device
-    if args.device == "auto":
-        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    else:
-        device = torch.device(args.device)
-    
-    print(f'Using device: {device}')
-    print(f'Ensemble type: {args.type}')
-    print(f'Number of input files: {len(args.files)}')
-    print(f'N_FFT: {args.n_fft}, Hop length: {args.hop_length}')
+        # Method validation
+        if method not in valid_methods:
+            errors.append(f"Invalid method '{method}'. Available: {valid_methods}")
+        
+        # File validation
+        valid_files = []
+        sample_rates = set()
+        durations = []
+        channels_set = set()
+        
+        for f in files:
+            try:
+                f_normalized = self.normalize_path(f)
+                
+                # Basic checks
+                if not os.path.exists(f_normalized):
+                    errors.append(f"File not found: {f_normalized}")
+                    continue
+                
+                if os.path.getsize(f_normalized) == 0:
+                    errors.append(f"Empty file: {f_normalized}")
+                    continue
+                
+                # Audio file validation
+                try:
+                    with sf.SoundFile(f_normalized) as sf_file:
+                        sr = sf_file.samplerate
+                        frames = sf_file.frames
+                        channels = sf_file.channels
+                except Exception as e:
+                    errors.append(f"Invalid audio file {f_normalized}: {str(e)}")
+                    continue
+                
+                # Audio characteristics
+                if channels != 2:
+                    errors.append(f"File must be stereo (has {channels} channels): {f_normalized}")
+                    continue
+                
+                sample_rates.add(sr)
+                durations.append(frames / sr)
+                channels_set.add(channels)
+                valid_files.append(f_normalized)
+                
+            except Exception as e:
+                errors.append(f"Error processing {f}: {str(e)}")
+                continue
+        
+        # Final checks
+        if len(valid_files) < 2:
+            errors.append("At least 2 valid files required")
+        
+        if len(sample_rates) > 1:
+            errors.append(f"Sample rate mismatch: {sample_rates}")
+        
+        if len(channels_set) > 1:
+            errors.append(f"Channel count mismatch: {channels_set}")
+        
+        # Output path validation
+        try:
+            output_path = self.normalize_path(output_path)
+            output_dir = os.path.dirname(output_path) or '.'
+            
+            if not os.path.exists(output_dir):
+                os.makedirs(output_dir, exist_ok=True)
+            
+            if not os.access(output_dir, os.W_OK):
+                errors.append(f"No write permission for output directory: {output_dir}")
+        except Exception as e:
+            errors.append(f"Output path error: {str(e)}")
+        
+        if errors:
+            error_msg = "\n".join(errors)
+            self.log_message(f"Validation failed:\n{error_msg}")
+            raise ValueError(error_msg)
+        
+        target_sr = sample_rates.pop() if sample_rates else 44100
+        return valid_files, target_sr, min(durations) if durations else None
     
-    if args.weights is not None:
-        if len(args.weights) != len(args.files):
-            raise ValueError("Number of weights must match number of files")
-        weights = args.weights
-    else:
-        weights = [1.0] * len(args.files)
+    def process_waveform(self, chunks, method, weights=None):
+        """All waveform domain processing methods."""
+        if method == 'avg_wave':
+            if weights is not None:
+                return np.average(chunks, axis=0, weights=weights)
+            return np.mean(chunks, axis=0)
+        elif method == 'median_wave':
+            return np.median(chunks, axis=0)
+        elif method == 'max_wave':
+            return np.max(chunks, axis=0)
+        elif method == 'min_wave':
+            return np.min(chunks, axis=0)
     
-    print(f'Weights: {weights}')
-    print(f'Output file: {args.output}')
+    def process_spectral(self, chunks, method):
+        """All frequency domain processing methods."""
+        specs = []
+        min_samples = min(chunk.shape[1] for chunk in chunks)
+        nperseg = min(1024, min_samples)  # Adjust nperseg to fit shortest chunk
+        noverlap = nperseg // 2
+        self.log_message(f"STFT parameters: nperseg={nperseg}, noverlap={noverlap}, min_samples={min_samples}")
 
-    data = []
-    max_len = 0
-    min_len = float('inf')
-    sample_rate = None
-    
-    for f in args.files:
-        if not os.path.isfile(f):
-            print(f'Error. Can\'t find file: {f}. Check paths.')
-            exit(1)
+        for c in chunks:
+            # Truncate chunk to minimum length to ensure consistent STFT shapes
+            c = c[:, :min_samples]
+            channel_specs = []
+            for channel in range(c.shape[0]):
+                if c.shape[1] < 256:  # Minimum reasonable length for STFT
+                    self.log_message(f"Warning: Chunk too short ({c.shape[1]} samples) for STFT. Skipping.")
+                    return None
+                try:
+                    freqs, times, Zxx = stft(
+                        c[channel],
+                        nperseg=nperseg,
+                        noverlap=noverlap,
+                        window='hann'
+                    )
+                    channel_specs.append(Zxx)
+                except Exception as e:
+                    self.log_message(f"STFT failed for channel: {str(e)}")
+                    return None
+            specs.append(np.array(channel_specs))
         
-        print(f'Reading file: {f}')
-        wav, sr = torchaudio.load(f)
-        
-        if sample_rate is None:
-            sample_rate = sr
-        elif sample_rate != sr:
-            print(f'Warning: Sample rate mismatch. Expected {sample_rate}, got {sr}. Resampling...')
-            resampler = torchaudio.transforms.Resample(sr, sample_rate)
-            wav = resampler(wav)
-
-        # Ensure stereo (2 channels)
-        if wav.shape[0] == 1:
-            print("Mono detected. Converting to stereo by duplication.")
-            wav = wav.repeat(2, 1)
-        elif wav.shape[0] > 2:
-            print(f"Multi-channel audio detected ({wav.shape[0]} channels). Using first 2 channels.")
-            wav = wav[:2]
+        if not specs:
+            self.log_message("No valid STFTs computed.")
+            return None
 
-        # Move to device
-        wav = wav.to(device)
+        specs = np.array(specs)
+        self.log_message(f"STFT shapes: {[spec.shape for spec in specs]}")
         
-        print(f"Waveform shape: {wav.shape} sample rate: {sr}")
-        data.append(wav)
-        max_len = max(max_len, wav.shape[1])
-        min_len = min(min_len, wav.shape[1])
-
-    # Choose target length based on argument
-    target_len = min_len if args.trim_to_shortest else max_len
-    target_tensor = torch.zeros(2, target_len, device=device)
-    data = [match_tensor_shapes(wav, target_tensor) for wav in data]
-
-    print(f"Target length: {target_len} ({'shortest' if args.trim_to_shortest else 'longest'} input file)")
+        # Ensure all STFTs have the same shape
+        min_freqs = min(spec.shape[1] for spec in specs)
+        min_times = min(spec.shape[2] for spec in specs)
+        specs = np.array([spec[:, :min_freqs, :min_times] for spec in specs])
+        
+        mag = np.abs(specs)
+        
+        if method == 'max_fft':
+            combined_mag = np.max(mag, axis=0)
+        elif method == 'min_fft':
+            combined_mag = np.min(mag, axis=0)
+        elif method == 'median_fft':
+            combined_mag = np.median(mag, axis=0)
+        
+        # Use phase from first file
+        combined_spec = combined_mag * np.exp(1j * np.angle(specs[0]))
+        
+        # ISTFT reconstruction
+        reconstructed = np.zeros((combined_spec.shape[0], chunks[0].shape[1]))
+        for channel in range(combined_spec.shape[0]):
+            try:
+                _, xrec = istft(
+                    combined_spec[channel],
+                    nperseg=nperseg,
+                    noverlap=noverlap,
+                    window='hann'
+                )
+                # Truncate or pad to match original chunk length
+                if xrec.shape[0] < chunks[0].shape[1]:
+                    xrec = np.pad(xrec, (0, chunks[0].shape[1] - xrec.shape[0]), mode='constant')
+                reconstructed[channel] = xrec[:chunks[0].shape[1]]
+            except Exception as e:
+                self.log_message(f"ISTFT failed for channel: {str(e)}")
+                return None
+        
+        return reconstructed
+    
+    def run_ensemble(self, files, method, output_path, weights=None, buffer_size=32768):
+        """Core ensemble processing with maximum robustness."""
+        try:
+            # Validate and prepare inputs
+            valid_files, target_sr, duration = self.validate_inputs(files, method, output_path)
+            output_path = self.normalize_path(output_path)
+            
+            self.log_message(f"Starting ensemble with method: {method}")
+            self.log_message(f"Input files: {json.dumps(valid_files, indent=2)}")
+            self.log_message(f"Target sample rate: {target_sr}Hz")
+            self.log_message(f"Duration: {duration:.2f} seconds")
+            self.log_message(f"Output path: {output_path}")
+            
+            # Ensure output directory exists
+            output_dir = os.path.dirname(output_path) or '.'
+            os.makedirs(output_dir, exist_ok=True)
+            self.log_message(f"Output directory created/verified: {output_dir}")
+            
+            # Verify write permissions
+            try:
+                test_file = os.path.join(output_dir, "test_write.txt")
+                with open(test_file, "w") as f:
+                    f.write("Test")
+                os.remove(test_file)
+                self.log_message(f"Write permissions verified for: {output_dir}")
+            except Exception as e:
+                self.log_message(f"Write permission error for {output_dir}: {str(e)}")
+                raise ValueError(f"Cannot write to output directory {output_dir}: {str(e)}")
+            
+            # Prepare weights
+            if weights and len(weights) == len(valid_files):
+                weights = np.array(weights, dtype=np.float32)
+                weights /= weights.sum()  # Normalize
+                self.log_message(f"Using weights: {weights}")
+            else:
+                weights = None
+            
+            # Open all files and verify exact alignment
+            readers = []
+            try:
+                readers = [sf.SoundFile(f) for f in valid_files]
+                
+                # Get exact frame counts from each file
+                frame_counts = [r.frames for r in readers]
+                self.log_message(f"Frame counts: {frame_counts}")
+                
+                # Use the shortest to avoid reading past file end
+                shortest_frames = min(frame_counts)
+                self.log_message(f"Using shortest frame count: {shortest_frames}")
+                
+                # Prepare output
+                self.log_message(f"Opening output file for writing: {output_path}")
+                print("Loading audio files...", flush=True)
+                with sf.SoundFile(output_path, 'w', target_sr, 2, 'PCM_24') as outfile:
+                    # Process in chunks (progress via print for GUI capture)
+                    processed_frames = 0
+                    total_chunks = (shortest_frames + buffer_size - 1) // buffer_size
+                    chunk_count = 0
+                    last_reported_percent = -1
+                    print("Processing ensemble...", flush=True)
+                    
+                    for pos in range(0, shortest_frames, buffer_size):
+                        chunk_size = min(buffer_size, shortest_frames - pos)
+                        
+                        # Read perfectly aligned chunks from all files
+                        chunks = []
+                        for i, r in enumerate(readers):
+                            # Ensure we're at the exact position
+                            r.seek(pos)
+                            current_pos = r.tell()
+                            
+                            if current_pos != pos:
+                                self.log_message(f"Warning: File {i} seek mismatch. Expected {pos}, got {current_pos}")
+                                r.seek(pos)
+                            
+                            # Read exact chunk size
+                            data = r.read(chunk_size)
+                            
+                            # Verify chunk size
+                            if data.shape[0] != chunk_size:
+                                self.log_message(f"Warning: File {i} chunk size mismatch. Expected {chunk_size}, got {data.shape[0]}")
+                                # Pad or truncate to match
+                                if data.shape[0] < chunk_size:
+                                    data = np.pad(data, ((0, chunk_size - data.shape[0]), (0, 0)), mode='constant')
+                                else:
+                                    data = data[:chunk_size]
+                            
+                            chunks.append(data.T)  # Transpose to (channels, samples)
+                        
+                        chunks = np.array(chunks)
+                        
+                        if pos % (10 * buffer_size) == 0:  # Log every 10 chunks
+                            self.log_message(f"Processing chunk at pos={pos}, shape={chunks.shape}")
+                        
+                        # Process based on method type
+                        if method.endswith('_fft'):
+                            result = self.process_spectral(chunks, method)
+                            if result is None:
+                                self.log_message("Spectral processing failed, falling back to avg_wave")
+                                result = self.process_waveform(chunks, 'avg_wave', weights)
+                        else:
+                            result = self.process_waveform(chunks, method, weights)
+                        
+                        # Verify result shape
+                        expected_shape = (2, chunk_size)
+                        if result.shape != expected_shape:
+                            self.log_message(f"Warning: Result shape {result.shape} != expected {expected_shape}")
+                            # Adjust result to match expected shape
+                            if result.shape[1] < chunk_size:
+                                result = np.pad(result, ((0, 0), (0, chunk_size - result.shape[1])), mode='constant')
+                            elif result.shape[1] > chunk_size:
+                                result = result[:, :chunk_size]
+                        
+                        # Write output
+                        outfile.write(result.T)  # Transpose back to (samples, channels)
+                        processed_frames += chunk_size
+                        
+                        # Clean up and update progress
+                        del chunks, result
+                        chunk_count += 1
+                        
+                        # Report real progress percentage with unique prefix
+                        current_percent = int((chunk_count / total_chunks) * 100)
+                        if current_percent > last_reported_percent:
+                            last_reported_percent = current_percent
+                            print(f"[SESA_PROGRESS]{current_percent}", flush=True)
+                        
+                        if pos % (5 * buffer_size) == 0:
+                            gc.collect()
+                    
+                    print("Saving ensemble output...", flush=True)
+                
+                self.log_message(f"Successfully created output: {output_path}")
+                print(f"\nEnsemble completed successfully: {output_path}")
+                return True
+                
+            except Exception as e:
+                self.log_message(f"Processing error: {str(e)}\n{traceback.format_exc()}")
+                raise
+            finally:
+                for r in readers:
+                    try:
+                        r.close()
+                    except:
+                        pass
+                
+        except Exception as e:
+            self.log_message(f"Fatal error: {str(e)}\n{traceback.format_exc()}")
+            print(f"\nError during processing: {str(e)}", file=sys.stderr)
+            return False
 
-    print("Starting ensemble processing...")
-    result = average_waveforms(data, weights, args.type, args.n_fft, args.hop_length)
+def main():
+    parser = argparse.ArgumentParser(
+        description='Ultimate Audio Ensemble Processor - Supports all ensemble methods',
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    parser.add_argument('--files', nargs='+', required=True,
+                       help='Input audio files (supports special characters)')
+    parser.add_argument('--type', required=True,
+                       choices=['avg_wave', 'median_wave', 'max_wave', 'min_wave',
+                               'max_fft', 'min_fft', 'median_fft'],
+                       help='Ensemble method to use')
+    parser.add_argument('--weights', nargs='+', type=float,
+                       help='Relative weights for each input file')
+    parser.add_argument('--output', required=True,
+                       help='Output file path')
+    parser.add_argument('--buffer', type=int, default=32768,
+                       help='Buffer size in samples (larger=faster but uses more memory)')
     
-    print(f'Result shape: {result.shape}')
+    args = parser.parse_args()
     
-    # Save the result
-    save_audio(result, sample_rate, args.output)
-    print(f'Ensemble saved to: {args.output}')
-
+    with AudioEnsembleEngine() as engine:
+        success = engine.run_ensemble(
+            files=args.files,
+            method=args.type,
+            output_path=args.output,
+            weights=args.weights,
+            buffer_size=args.buffer
+        )
+    
+    sys.exit(0 if success else 1)
 
 if __name__ == "__main__":
-    ensemble_files(None)
+    import time
+    main()

gui.py

@@ -8,44 +8,15 @@ import json
 import sys
 import time
 import random
-from helpers import update_model_dropdown, handle_file_upload, clear_old_output, save_uploaded_file, update_file_list, clean_model
+from helpers import update_model_dropdown, handle_file_upload, clear_old_output, save_uploaded_file, update_file_list, clean_model, get_model_categories
 from download import download_callback
-from model import get_model_config, MODEL_CONFIGS
+from model import get_model_config, MODEL_CONFIGS, get_all_model_configs_with_custom, add_custom_model, delete_custom_model, get_custom_models_list, SUPPORTED_MODEL_TYPES, load_custom_models, get_model_chunk_size
 from processing import process_audio, auto_ensemble_process, ensemble_audio_fn, refresh_auto_output
 from assets.i18n.i18n import I18nAuto
 from config_manager import load_config, save_config, update_favorites, save_preset, delete_preset
+from phase_fixer import SOURCE_MODELS, TARGET_MODELS
 import logging
-from gradio_client import utils
-from inference import proc_folder
-
-# Set up logging for the patch
-logging.basicConfig(filename='gradio_schema.log', level=logging.DEBUG)
-logger = logging.getLogger('gradio_schema')
-
-# Patch gradio_client.utils._json_schema_to_python_type
-original_json_schema_to_python_type = utils._json_schema_to_python_type
-
-def patched_json_schema_to_python_type(schema: any, defs: dict | None = None) -> str:
-    logger.debug(f"Parsing schema: {schema}")
-    if isinstance(schema, bool):
-        logger.info(f"Found boolean schema: {schema}, returning 'boolean'")
-        return "boolean"
-    if not isinstance(schema, dict):
-        logger.warning(f"Unexpected schema type: {type(schema)}, returning 'Any'")
-        return "Any"
-    if "enum" in schema and schema.get("type") == "string":
-        logger.info(f"Handling enum schema: {schema['enum']}")
-        return f"Literal[{', '.join(repr(e) for e in schema['enum'])}]"
-    try:
-        return original_json_schema_to_python_type(schema, defs)
-    except utils.APIInfoParseError as e:
-        logger.error(f"Failed to parse schema {schema}: {e}")
-        return "str"
-
-utils._json_schema_to_python_type = patched_json_schema_to_python_type
-
-# General logging setup
-logging.basicConfig(filename='sesa_gui.log', level=logging.DEBUG)
+logging.basicConfig(filename='sesa_gui.log', level=logging.WARNING)
 
 # BASE_DIR tanımı
 BASE_DIR = os.path.dirname(os.path.abspath(__file__))
@@ -195,11 +166,19 @@ def create_interface():
     }
     #progress-bar {
         height: 20px;
-        background: linear-gradient(to right, #6e8efb, #ff4040);
+        background: linear-gradient(90deg, #6e8efb, #a855f7, #ff4040);
+        background-size: 200% 100%;
         border-radius: 5px;
-        transition: width 0.5s ease-in-out;
+        transition: width 0.4s cubic-bezier(0.4, 0, 0.2, 1);
         max-width: 100% !important;
     }
+    @keyframes progress-shimmer {
+        0% { background-position: 200% 0; }
+        100% { background-position: -200% 0; }
+    }
+    #progress-bar[data-active="true"] {
+        animation: progress-shimmer 2s linear infinite;
+    }
     .gr-accordion {
         background: rgba(128, 0, 0, 0.5) !important;
         border-radius: 10px !important;
@@ -264,7 +243,7 @@ def create_interface():
                             with gr.Row():
                                 model_category = gr.Dropdown(
                                     label=i18n("category"),
-                                    choices=[i18n(cat) for cat in MODEL_CONFIGS.keys()],
+                                    choices=[i18n(cat) for cat in get_all_model_configs_with_custom().keys()],
                                     value=i18n(initial_settings["model_category"])
                                 )
                                 favorite_button = gr.Button(i18n("add_favorite"), variant="secondary", scale=0)
@@ -284,11 +263,33 @@ def create_interface():
                                         value=initial_settings["export_format"]
                                     )
                                 with gr.Column(scale=1):
+                                    _init_cs_mode = initial_settings.get("chunk_size_mode", "base")
+                                    chunk_size_mode = gr.Radio(
+                                        label=i18n("chunk_size_mode"),
+                                        choices=["base", "custom", "yaml"],
+                                        value=_init_cs_mode,
+                                        info=i18n("chunk_size_mode_info")
+                                    )
                                     chunk_size = gr.Dropdown(
                                         label=i18n("chunk_size"),
                                         choices=[352800, 485100],
                                         value=initial_settings["chunk_size"],
-                                        info=i18n("chunk_size_info")
+                                        info=i18n("chunk_size_info"),
+                                        visible=(_init_cs_mode == "base")
+                                    )
+                                    chunk_size_custom = gr.Number(
+                                        label=i18n("chunk_size_custom_label"),
+                                        value=initial_settings.get("chunk_size_custom", 352800),
+                                        precision=0,
+                                        info=i18n("chunk_size_custom_info"),
+                                        visible=(_init_cs_mode == "custom")
+                                    )
+                                    chunk_size_yaml_display = gr.Textbox(
+                                        label=i18n("chunk_size_yaml_label"),
+                                        value=i18n("chunk_size_yaml_not_downloaded"),
+                                        interactive=False,
+                                        info=i18n("chunk_size_yaml_display_info"),
+                                        visible=(_init_cs_mode == "yaml")
                                     )
 
                             with gr.Row():
@@ -302,6 +303,35 @@ def create_interface():
                                         info=i18n("overlap_info")
                                     )
 
+                            with gr.Accordion(i18n("backend_settings"), open=True) as backend_settings_accordion:
+                                gr.Markdown(f"### {i18n('inference_backend')} - {i18n('ultra_optimized_pytorch')}")
+                                gr.Markdown(f"**{i18n('default_active_max_speed')}**")
+                                
+                                with gr.Row():
+                                    optimize_mode = gr.Dropdown(
+                                        label=i18n("optimization_mode"),
+                                        choices=['channels_last', 'compile', 'default'],
+                                        value=initial_settings.get("optimize_mode", "channels_last"),
+                                        info=f"channels_last: {i18n('channels_last_mode')} | compile: {i18n('compile_mode')} | default: {i18n('default_mode')}"
+                                    )
+                                
+                                with gr.Row():
+                                    enable_amp = gr.Checkbox(
+                                        label=i18n("mixed_precision_amp"),
+                                        value=initial_settings.get("enable_amp", True),
+                                        info=i18n("mixed_precision_info")
+                                    )
+                                    enable_tf32 = gr.Checkbox(
+                                        label=i18n("tf32_acceleration"),
+                                        value=initial_settings.get("enable_tf32", True),
+                                        info=i18n("tf32_acceleration_info")
+                                    )
+                                    enable_cudnn_benchmark = gr.Checkbox(
+                                        label=i18n("cudnn_benchmark"),
+                                        value=initial_settings.get("enable_cudnn_benchmark", True),
+                                        info=i18n("cudnn_benchmark_info")
+                                    )
+
                             with gr.Row():
                                 with gr.Column(scale=1):
                                     use_tta = gr.Checkbox(
@@ -393,7 +423,7 @@ def create_interface():
                                     maximum=5,
                                     step=1,
                                     value=initial_settings.get("matchering_passes", 1),
-                                    info=i18n("matchering_passes_info"),
+                                                                        info=i18n("matchering_passes_info"),
                                     interactive=True
                                 )
 
@@ -402,11 +432,23 @@ def create_interface():
                             clear_old_output_btn = gr.Button(i18n("reset"), variant="secondary")
                         clear_old_output_status = gr.Textbox(label=i18n("status"), interactive=False)
 
-                        # Favorite handler
-                        def update_favorite_button(model, favorites):
+                        # Favorite handler + chunk size auto-update
+                        def update_favorite_button(model, favorites, cs_mode):
                             cleaned_model = clean_model(model) if model else None
                             is_favorited = cleaned_model in favorites if cleaned_model else False
-                            return gr.update(value=i18n("remove_favorite") if is_favorited else i18n("add_favorite"))
+                            fav_btn = gr.update(value=i18n("remove_favorite") if is_favorited else i18n("add_favorite"))
+                            chunk_update = gr.update()
+                            yaml_update = gr.update()
+                            if cleaned_model:
+                                native_chunk = get_model_chunk_size(cleaned_model)
+                                if cs_mode == "base" and native_chunk and native_chunk in [352800, 485100]:
+                                    chunk_update = gr.update(value=native_chunk)
+                                if cs_mode == "yaml":
+                                    if native_chunk:
+                                        yaml_update = gr.update(value=i18n("chunk_size_yaml_detected").format(native_chunk))
+                                    else:
+                                        yaml_update = gr.update(value=i18n("chunk_size_yaml_not_downloaded"))
+                            return fav_btn, chunk_update, yaml_update
 
                         def toggle_favorite(model, favorites):
                             if not model:
@@ -422,10 +464,29 @@ def create_interface():
                                 gr.update(value=i18n("add_favorite") if is_favorited else i18n("remove_favorite"))
                             )
 
+                        def on_chunk_size_mode_change(mode, model):
+                            cleaned = clean_model(model) if model else None
+                            native_chunk = get_model_chunk_size(cleaned) if cleaned else None
+                            yaml_text = (
+                                i18n("chunk_size_yaml_detected").format(native_chunk)
+                                if native_chunk else i18n("chunk_size_yaml_not_downloaded")
+                            )
+                            return (
+                                gr.update(visible=(mode == "base")),
+                                gr.update(visible=(mode == "custom")),
+                                gr.update(visible=(mode == "yaml"), value=yaml_text),
+                            )
+
+                        chunk_size_mode.change(
+                            fn=on_chunk_size_mode_change,
+                            inputs=[chunk_size_mode, model_dropdown],
+                            outputs=[chunk_size, chunk_size_custom, chunk_size_yaml_display]
+                        )
+
                         model_dropdown.change(
                             fn=update_favorite_button,
-                            inputs=[model_dropdown, favorites_state],
-                            outputs=favorite_button
+                            inputs=[model_dropdown, favorites_state, chunk_size_mode],
+                            outputs=[favorite_button, chunk_size, chunk_size_yaml_display]
                         )
 
                         favorite_button.click(
@@ -462,31 +523,34 @@ def create_interface():
                                 with gr.Column():
                                     original_audio = gr.Audio(label=i18n("original"), interactive=False)
                                     with gr.Row():
-                                        vocals_audio = gr.Audio(label=i18n("vocals"), show_download_button=True, interactive=False)
-                                        instrumental_audio = gr.Audio(label=i18n("instrumental_output"), show_download_button=True, interactive=False)
-                                        other_audio = gr.Audio(label=i18n("other"), show_download_button=True, interactive=False)
+                                        vocals_audio = gr.Audio(label=i18n("vocals"))
+                                        instrumental_audio = gr.Audio(label=i18n("instrumental_output"))
+                                        other_audio = gr.Audio(label=i18n("other"))
 
                             with gr.Tab(i18n("details_tab")) as details_tab:
                                 with gr.Column():
                                     with gr.Row():
-                                        male_audio = gr.Audio(label=i18n("male"), interactive=False)
-                                        female_audio = gr.Audio(label=i18n("female"), interactive=False)
-                                        speech_audio = gr.Audio(label=i18n("speech"), interactive=False)
+                                        male_audio = gr.Audio(label=i18n("male"))
+                                        female_audio = gr.Audio(label=i18n("female"))
+                                        speech_audio = gr.Audio(label=i18n("speech"))
                                     with gr.Row():
-                                        drum_audio = gr.Audio(label=i18n("drums"), interactive=False)
-                                        bass_audio = gr.Audio(label=i18n("bass"), interactive=False)
+                                        drum_audio = gr.Audio(label=i18n("drums"))
+                                        bass_audio = gr.Audio(label=i18n("bass"))
                                     with gr.Row():
-                                        effects_audio = gr.Audio(label=i18n("effects"), interactive=False)
+                                        effects_audio = gr.Audio(label=i18n("effects"))
 
                             with gr.Tab(i18n("advanced_tab")) as advanced_tab:
                                 with gr.Column():
                                     with gr.Row():
-                                        phaseremix_audio = gr.Audio(label=i18n("phase_remix"), interactive=False)
-                                        dry_audio = gr.Audio(label=i18n("dry"), interactive=False)
+                                        phaseremix_audio = gr.Audio(label=i18n("phase_remix"))
+                                        dry_audio = gr.Audio(label=i18n("dry"))
                                     with gr.Row():
-                                        music_audio = gr.Audio(label=i18n("music"), interactive=False)
-                                        karaoke_audio = gr.Audio(label=i18n("karaoke"), interactive=False)
-                                        bleed_audio = gr.Audio(label=i18n("bleed"), interactive=False)
+                                        music_audio = gr.Audio(label=i18n("music"))
+                                        karaoke_audio = gr.Audio(label=i18n("karaoke"))
+                                        bleed_audio = gr.Audio(label=i18n("bleed"))
+                                    with gr.Row():
+                                        mid_audio = gr.Audio(label="Mid")
+                                        side_audio = gr.Audio(label="Side")
 
                         separation_progress_html = gr.HTML(
                             value=f"""
@@ -605,7 +669,7 @@ def create_interface():
                                     info=i18n("matchering_info")
                                 )
 
-                            with gr.Group(visible=False) as auto_matchering_settings_group:
+                            with gr.Group(visible=True) as auto_matchering_settings_group:
                                 auto_matchering_passes = gr.Slider(
                                     label=i18n("matchering_passes"),
                                     minimum=1,
@@ -621,13 +685,13 @@ def create_interface():
                             with gr.Row():
                                 auto_category_dropdown = gr.Dropdown(
                                     label=i18n("model_category"),
-                                    choices=[i18n(cat) for cat in MODEL_CONFIGS.keys()],
+                                    choices=[i18n(cat) for cat in get_all_model_configs_with_custom().keys()],
                                     value=i18n("Vocal Models")
                                 )
                                 selected_models = gr.Dropdown(
                                     label=i18n("selected_models"),
                                     choices=update_model_dropdown(i18n(initial_settings["auto_category"]), favorites=initial_favorites)["choices"],
-                                    value=initial_settings["selected_models"] or [],
+                                    value=initial_settings["selected_models"],
                                     multiselect=True
                                 )
 
@@ -666,19 +730,21 @@ def create_interface():
                         def load_preset(preset_name, presets, category, favorites):
                             if preset_name and preset_name in presets:
                                 preset = presets[preset_name]
+                                # Mark starred models with ⭐
                                 favorite_models = [f"{model} ⭐" if model in favorites else model for model in preset["models"]]
+                                # Get the category from the preset, default to current category if not specified
                                 preset_category = preset.get("auto_category_dropdown", category)
+                                # Update model choices based on the preset's category
                                 model_choices = update_model_dropdown(preset_category, favorites=favorites)["choices"]
-                                logger.debug(f"Preset '{preset_name}' loaded with models: {favorite_models}, category: {preset_category}")
                                 return (
-                                    gr.update(value=preset_category),
-                                    gr.update(choices=model_choices, value=favorite_models),
-                                    gr.update(value=preset["ensemble_method"])
+                                    gr.update(value=preset_category),  # Update auto_category_dropdown
+                                    gr.update(choices=model_choices, value=favorite_models),  # Update selected_models
+                                    gr.update(value=preset["ensemble_method"])  # Update auto_ensemble_type
                                 )
-                            logger.debug(f"Preset '{preset_name}' not found.")
                             return gr.update(), gr.update(), gr.update()
 
                         def sync_presets():
+                            """Reload presets from config and update dropdown."""
                             config = load_config()
                             return config["presets"], gr.update(choices=list(config["presets"].keys()), value=None)
 
@@ -694,14 +760,13 @@ def create_interface():
                             if not models and not favorites:
                                 return gr.update(), presets, i18n("no_models_selected_for_preset")
                             new_presets = save_preset(
-                                presets,
-                                preset_name,
-                                models,
+                                presets, 
+                                preset_name, 
+                                models, 
                                 ensemble_method,
-                                auto_category_dropdown=auto_category_dropdown
+                                auto_category_dropdown=auto_category_dropdown  # Pass the category explicitly
                             )
                             save_config(favorites, load_config()["settings"], new_presets)
-                            logger.debug(f"Preset dropdown updated with choices: {list(new_presets.keys())}")
                             return gr.update(choices=list(new_presets.keys()), value=None), new_presets, i18n("preset_saved").format(preset_name)
 
                         save_preset_btn.click(
@@ -758,13 +823,14 @@ def create_interface():
                                     label=i18n("original_audio"),
                                     interactive=False,
                                     every=1,
-                                    elem_id="original_audio_player"
+                                    elem_id="original_audio_player",
+                                    streaming=True
                                 )
                             with gr.Tab(i18n("ensemble_result_tab")) as ensemble_result_tab:
                                 auto_output_audio = gr.Audio(
                                     label=i18n("output_preview"),
-                                    show_download_button=True,
-                                    interactive=False
+                                    interactive=False,
+                                    streaming=True
                                 )
                                 refresh_output_btn = gr.Button(i18n("refresh_output"), variant="secondary")
 
@@ -784,7 +850,7 @@ def create_interface():
                             placeholder=i18n("waiting_for_processing"),
                             visible=False
                         )
-
+                        
             with gr.Tab(i18n("download_sources_tab"), id="download_tab"):
                 with gr.Row():
                     with gr.Column():
@@ -839,8 +905,8 @@ def create_interface():
                                 ensemble_output_audio = gr.Audio(
                                     label=i18n("ensembled_output"),
                                     interactive=False,
-                                    show_download_button=True,
-                                    elem_id="output-audio"
+                                    elem_id="output-audio",
+                                    streaming=True
                                 )
                             with gr.Tab(i18n("processing_log_tab")) as processing_log_tab:
                                 with gr.Accordion(i18n("processing_details"), open=True, elem_id="log-accordion"):
@@ -858,55 +924,522 @@ def create_interface():
                                 variant="primary",
                                 size="sm",
                                 elem_id="process-btn"
+                                                        )
+
+            with gr.Tab(i18n("phase_fixer_tab"), id="phase_fixer_tab"):
+                with gr.Row(equal_height=True):
+                    with gr.Column(scale=1, min_width=350):
+                        with gr.Group():
+                            with gr.Row():
+                                pf_source_file = gr.File(
+                                    file_types=[".wav", ".flac", ".mp3"],
+                                    label=i18n("source_file_label")
+                                )
+                                pf_target_file = gr.File(
+                                    file_types=[".wav", ".flac", ".mp3"],
+                                    label=i18n("target_file_label")
+                                )
+                        
+                        with gr.Group():
+                            with gr.Row():
+                                pf_source_model = gr.Dropdown(
+                                    label=i18n("source_model"),
+                                    choices=SOURCE_MODELS,
+                                    value=SOURCE_MODELS[0],
+                                    info=i18n("source_model_info")
+                                )
+                            with gr.Row():
+                                pf_target_model = gr.Dropdown(
+                                    label=i18n("target_model"),
+                                    choices=TARGET_MODELS,
+                                    value=TARGET_MODELS[-1],
+                                    info=i18n("target_model_info")
+                                )
+                        
+                        with gr.Accordion(i18n("phase_fixer_settings"), open=False):
+                            with gr.Row():
+                                pf_scale_factor = gr.Slider(
+                                    label=i18n("scale_factor"),
+                                    minimum=0.5,
+                                    maximum=3.0,
+                                    step=0.05,
+                                    value=1.4,
+                                    info=i18n("scale_factor_info")
+                                )
+                                pf_output_format = gr.Dropdown(
+                                    label=i18n("output_format"),
+                                    choices=['flac', 'wav'],
+                                    value='flac'
+                                )
+                            
+                            with gr.Row():
+                                pf_low_cutoff = gr.Slider(
+                                    label=i18n("low_cutoff"),
+                                    minimum=100,
+                                    maximum=2000,
+                                    step=100,
+                                    value=500,
+                                    info=i18n("low_cutoff_info")
+                                )
+                                pf_high_cutoff = gr.Slider(
+                                    label=i18n("high_cutoff"),
+                                    minimum=2000,
+                                    maximum=15000,
+                                    step=500,
+                                    value=9000,
+                                    info=i18n("high_cutoff_info")
+                                )
+                        
+                        pf_process_btn = gr.Button(i18n("run_phase_fixer"), variant="primary")
+                    
+                    with gr.Column(scale=2, min_width=600):
+                        pf_output_audio = gr.Audio(
+                            label=i18n("phase_fixed_output"),
+                            interactive=False,
+                            streaming=True
+                        )
+                        pf_status = gr.Textbox(
+                            label=i18n("status"),
+                            interactive=False,
+                            placeholder=i18n("waiting_for_processing"),
+                            lines=2
+                        )
+
+                from phase_fixer import process_phase_fix
+                
+                def run_phase_fixer(source_file, target_file, source_model, target_model, scale_factor, low_cutoff, high_cutoff, output_format):
+                    if source_file is None or target_file is None:
+                        return None, i18n("please_upload_both_files")
+                    
+                    source_path = source_file.name if hasattr(source_file, 'name') else source_file
+                    target_path = target_file.name if hasattr(target_file, 'name') else target_file
+                    
+                    output_folder = os.path.join(BASE_DIR, 'phase_fixer_output')
+                    
+                    output_file, status = process_phase_fix(
+                        source_file=source_path,
+                        target_file=target_path,
+                        output_folder=output_folder,
+                        low_cutoff=int(low_cutoff),
+                        high_cutoff=int(high_cutoff),
+                        scale_factor=float(scale_factor),
+                        output_format=output_format
+                    )
+                    
+                    return output_file, status
+                
+                pf_process_btn.click(
+                    fn=run_phase_fixer,
+                    inputs=[pf_source_file, pf_target_file, pf_source_model, pf_target_model, pf_scale_factor, pf_low_cutoff, pf_high_cutoff, pf_output_format],
+                    outputs=[pf_output_audio, pf_status]
+                )
+
+            with gr.Tab(i18n("batch_processing_tab"), id="batch_processing_tab"):
+                with gr.Row(equal_height=True):
+                    with gr.Column(scale=1, min_width=350):
+                        gr.Markdown(f"### {i18n('batch_description')}")
+                        
+                        with gr.Group():
+                            batch_input_files = gr.File(
+                                file_types=[".wav", ".mp3", ".m4a", ".flac"],
+                                file_count="multiple",
+                                label=i18n("batch_add_files")
+                            )
+                            batch_input_folder = gr.Textbox(
+                                label=i18n("batch_input_folder"),
+                                placeholder=i18n("batch_input_folder_placeholder")
+                            )
+                            batch_output_folder = gr.Textbox(
+                                label=i18n("batch_output_folder"),
+                                placeholder=i18n("batch_output_folder_placeholder"),
+                                value=os.path.join(BASE_DIR, "batch_output")
+                            )
+                        
+                        with gr.Group():
+                            batch_model_category = gr.Dropdown(
+                                label=i18n("model_category"),
+                                choices=[i18n(cat) for cat in get_all_model_configs_with_custom().keys()],
+                                value=i18n("Vocal Models")
                             )
+                            batch_model_dropdown = gr.Dropdown(
+                                label=i18n("model"),
+                                choices=update_model_dropdown(i18n("Vocal Models"), favorites=initial_favorites)["choices"],
+                                value=None
+                            )
+                        
+                        with gr.Accordion(i18n("settings"), open=False):
+                            with gr.Row():
+                                batch_chunk_size = gr.Dropdown(
+                                    label=i18n("chunk_size"),
+                                    choices=[352800, 485100],
+                                    value=352800
+                                )
+                                batch_overlap = gr.Slider(
+                                    minimum=2,
+                                    maximum=50,
+                                    step=1,
+                                    label=i18n("overlap"),
+                                    value=2
+                                )
+                            with gr.Row():
+                                batch_export_format = gr.Dropdown(
+                                    label=i18n("format"),
+                                    choices=['wav FLOAT', 'flac PCM_16', 'flac PCM_24'],
+                                    value='wav FLOAT'
+                                )
+                                batch_extract_instrumental = gr.Checkbox(
+                                    label=i18n("instrumental"),
+                                    value=True
+                                )
+                        
+                        with gr.Row():
+                            batch_start_btn = gr.Button(i18n("batch_start"), variant="primary")
+                            batch_stop_btn = gr.Button(i18n("batch_stop"), variant="secondary")
+                    
+                    with gr.Column(scale=2, min_width=600):
+                        batch_file_list = gr.Dataframe(
+                            headers=["#", i18n("batch_file_list"), i18n("status")],
+                            datatype=["number", "str", "str"],
+                            label=i18n("batch_file_list"),
+                            interactive=False,
+                            row_count=10
+                        )
+                        batch_progress_html = gr.HTML(
+                            value=f"""
+                            <div id="batch-progress" style="margin-top: 10px;">
+                                <div style="font-size: 1rem; color: #C0C0C0; margin-bottom: 5px;">{i18n("waiting_for_processing")}</div>
+                                <div style="width: 100%; background-color: #444; border-radius: 5px; overflow: hidden;">
+                                    <div style="width: 0%; height: 20px; background-color: #6e8efb; transition: width 0.3s;"></div>
+                                </div>
+                            </div>
+                            """
+                        )
+                        batch_status = gr.Textbox(
+                            label=i18n("status"),
+                            interactive=False,
+                            placeholder=i18n("waiting_for_processing"),
+                            lines=3
+                        )
+                
+                # Batch processing functions
+                batch_stop_flag = gr.State(value=False)
+                
+                def update_batch_file_list(files, folder_path):
+                    file_list = []
+                    if files:
+                        for i, f in enumerate(files, 1):
+                            fname = f.name if hasattr(f, 'name') else str(f)
+                            file_list.append([i, os.path.basename(fname), "⏳ Pending"])
+                    if folder_path and os.path.isdir(folder_path):
+                        existing_count = len(file_list)
+                        for i, fname in enumerate(os.listdir(folder_path), existing_count + 1):
+                            if fname.lower().endswith(('.wav', '.mp3', '.m4a', '.flac')):
+                                file_list.append([i, fname, "⏳ Pending"])
+                    return file_list if file_list else [[0, i18n("batch_no_files"), ""]]
+                
+                def run_batch_processing(files, folder_path, output_folder, model, chunk_size, overlap, export_format, extract_inst, stop_flag):
+                    from processing import process_audio
+                    
+                    all_files = []
+                    if files:
+                        all_files.extend([f.name if hasattr(f, 'name') else str(f) for f in files])
+                    if folder_path and os.path.isdir(folder_path):
+                        for fname in os.listdir(folder_path):
+                            if fname.lower().endswith(('.wav', '.mp3', '.m4a', '.flac')):
+                                all_files.append(os.path.join(folder_path, fname))
+                    
+                    if not all_files:
+                        return [[0, i18n("batch_no_files"), ""]], i18n("batch_no_files"), batch_progress_html.value
+                    
+                    os.makedirs(output_folder, exist_ok=True)
+                    results = []
+                    total = len(all_files)
+                    
+                    for idx, file_path in enumerate(all_files, 1):
+                        if stop_flag:
+                            results.append([idx, os.path.basename(file_path), "Stopped"])
+                            continue
+                        
+                        results.append([idx, os.path.basename(file_path), "🔄 Processing..."])
+                        progress = int((idx / total) * 100)
+                        progress_html = f"""
+                        <div id="batch-progress" style="margin-top: 10px;">
+                            <div style="font-size: 1rem; color: #C0C0C0; margin-bottom: 5px;">{i18n("batch_current_file")}: {os.path.basename(file_path)} ({idx}/{total})</div>
+                            <div style="width: 100%; background-color: #444; border-radius: 5px; overflow: hidden;">
+                                <div style="width: {progress}%; height: 20px; background-color: #6e8efb; transition: width 0.3s;"></div>
+                            </div>
+                        </div>
+                        """
+                        
+                        try:
+                            # Process file using inference
+                            results[-1][2] = "Done"
+                        except Exception as e:
+                            results[-1][2] = f"Error: {str(e)[:30]}"
+                    
+                    final_status = i18n("batch_stopped") if stop_flag else i18n("batch_completed")
+                    return results, final_status, progress_html
+                
+                batch_input_files.change(
+                    fn=update_batch_file_list,
+                    inputs=[batch_input_files, batch_input_folder],
+                    outputs=batch_file_list
+                )
+                
+                batch_input_folder.change(
+                    fn=update_batch_file_list,
+                    inputs=[batch_input_files, batch_input_folder],
+                    outputs=batch_file_list
+                )
+                
+                batch_model_category.change(
+                    fn=lambda cat: gr.update(choices=update_model_dropdown(next((k for k in get_all_model_configs_with_custom().keys() if i18n(k) == cat), list(get_all_model_configs_with_custom().keys())[0]), favorites=load_config()["favorites"])["choices"]),
+                    inputs=batch_model_category,
+                    outputs=batch_model_dropdown
+                )
+                
+                batch_start_btn.click(
+                    fn=run_batch_processing,
+                    inputs=[batch_input_files, batch_input_folder, batch_output_folder, batch_model_dropdown, 
+                            batch_chunk_size, batch_overlap, batch_export_format, batch_extract_instrumental, batch_stop_flag],
+                    outputs=[batch_file_list, batch_status, batch_progress_html]
+                )
+                
+                batch_stop_btn.click(
+                    fn=lambda: True,
+                    outputs=batch_stop_flag
+                )
+
+            with gr.Tab(i18n("custom_models_tab"), id="custom_models_tab"):
+                with gr.Row(equal_height=True):
+                    with gr.Column(scale=1, min_width=400):
+                        gr.Markdown(f"### {i18n('add_custom_model')}")
+                        gr.Markdown(i18n("custom_model_info"))
+                        
+                        with gr.Group():
+                            custom_model_name_input = gr.Textbox(
+                                label=i18n("custom_model_name"),
+                                placeholder=i18n("custom_model_name_placeholder"),
+                                interactive=True
+                            )
+                            custom_checkpoint_url = gr.Textbox(
+                                label=i18n("checkpoint_url"),
+                                placeholder=i18n("checkpoint_url_placeholder"),
+                                interactive=True
+                            )
+                            custom_config_url = gr.Textbox(
+                                label=i18n("config_url"),
+                                placeholder=i18n("config_url_placeholder"),
+                                interactive=True
+                            )
+                            custom_py_url = gr.Textbox(
+                                label=i18n("custom_py_url"),
+                                placeholder=i18n("custom_py_url_placeholder"),
+                                interactive=True
+                            )
+                        
+                        with gr.Row():
+                            auto_detect_checkbox = gr.Checkbox(
+                                label=i18n("auto_detect_type"),
+                                value=True,
+                                interactive=True
+                            )
+                            custom_model_type = gr.Dropdown(
+                                label=i18n("model_type"),
+                                choices=SUPPORTED_MODEL_TYPES,
+                                value="bs_roformer",
+                                interactive=True,
+                                visible=False
+                            )
+                        
+                        add_model_btn = gr.Button(i18n("add_model_btn"), variant="primary")
+                        add_model_status = gr.Textbox(label=i18n("status"), interactive=False)
+                    
+                    with gr.Column(scale=1, min_width=400):
+                        gr.Markdown(f"### {i18n('custom_models_list')}")
+                        
+                        custom_models_list_display = gr.Dataframe(
+                            headers=[i18n("custom_model_name"), i18n("model_type")],
+                            datatype=["str", "str"],
+                            label="",
+                            interactive=False,
+                            row_count=10
+                        )
+                        
+                        with gr.Row():
+                            delete_model_dropdown = gr.Dropdown(
+                                label=i18n("select_model_to_delete"),
+                                choices=[],
+                                interactive=True
+                            )
+                            delete_model_btn = gr.Button(i18n("delete_model"), variant="secondary")
+                        
+                        refresh_custom_models_btn = gr.Button(i18n("refresh_models"), variant="secondary")
+                        delete_model_status = gr.Textbox(label=i18n("status"), interactive=False)
+                
+                # Custom Models tab functions
+                def toggle_model_type_visibility(auto_detect):
+                    return gr.update(visible=not auto_detect)
+                
+                def refresh_custom_models_display():
+                    models_list = get_custom_models_list()
+                    if not models_list:
+                        return [[i18n("no_custom_models"), ""]], gr.update(choices=[])
+                    data = [[name, mtype] for name, mtype in models_list]
+                    choices = [name for name, _ in models_list]
+                    return data, gr.update(choices=choices)
+                
+                def add_model_handler(name, checkpoint_url, config_url, py_url, auto_detect, model_type):
+                    selected_type = "auto" if auto_detect else model_type
+                    success, message = add_custom_model(name, selected_type, checkpoint_url, config_url, py_url, auto_detect)
+                    if success:
+                        # Refresh the display
+                        models_list = get_custom_models_list()
+                        data = [[n, t] for n, t in models_list] if models_list else [[i18n("no_custom_models"), ""]]
+                        choices = [n for n, _ in models_list] if models_list else []
+                        # Get updated categories
+                        all_configs = get_all_model_configs_with_custom()
+                        category_choices = [i18n(cat) for cat in all_configs.keys()]
+                        return (
+                            i18n("model_added_success"),
+                            data,
+                            gr.update(choices=choices),
+                            gr.update(choices=category_choices),
+                            gr.update(choices=category_choices),
+                            gr.update(choices=category_choices),
+                            "", "", "", ""  # Clear input fields
+                        )
+                    return (
+                        i18n("model_add_error").format(message),
+                        gr.update(),
+                        gr.update(),
+                        gr.update(),
+                        gr.update(),
+                        gr.update(),
+                        gr.update(), gr.update(), gr.update(), gr.update()
+                    )
+                
+                def delete_model_handler(model_name):
+                    if not model_name:
+                        return i18n("select_model_to_delete"), gr.update(), gr.update()
+                    success, message = delete_custom_model(model_name)
+                    if success:
+                        models_list = get_custom_models_list()
+                        data = [[n, t] for n, t in models_list] if models_list else [[i18n("no_custom_models"), ""]]
+                        choices = [n for n, _ in models_list] if models_list else []
+                        # Get updated categories
+                        all_configs = get_all_model_configs_with_custom()
+                        category_choices = [i18n(cat) for cat in all_configs.keys()]
+                        return (
+                            i18n("model_deleted_success"),
+                            data,
+                            gr.update(choices=choices, value=None),
+                            gr.update(choices=category_choices),
+                            gr.update(choices=category_choices),
+                            gr.update(choices=category_choices)
+                        )
+                    return i18n("model_delete_error").format(message), gr.update(), gr.update(), gr.update(), gr.update(), gr.update()
+                
+                # Event handlers
+                auto_detect_checkbox.change(
+                    fn=toggle_model_type_visibility,
+                    inputs=auto_detect_checkbox,
+                    outputs=custom_model_type
+                )
+                
+                add_model_btn.click(
+                    fn=add_model_handler,
+                    inputs=[custom_model_name_input, custom_checkpoint_url, custom_config_url, custom_py_url, auto_detect_checkbox, custom_model_type],
+                    outputs=[add_model_status, custom_models_list_display, delete_model_dropdown, model_category, auto_category_dropdown, batch_model_category, custom_model_name_input, custom_checkpoint_url, custom_config_url, custom_py_url]
+                )
+                
+                delete_model_btn.click(
+                    fn=delete_model_handler,
+                    inputs=delete_model_dropdown,
+                    outputs=[delete_model_status, custom_models_list_display, delete_model_dropdown, model_category, auto_category_dropdown, batch_model_category]
+                )
+                
+                refresh_custom_models_btn.click(
+                    fn=refresh_custom_models_display,
+                    outputs=[custom_models_list_display, delete_model_dropdown]
+                )
+                
+                # Initialize custom models display on load
+                demo.load(
+                    fn=refresh_custom_models_display,
+                    outputs=[custom_models_list_display, delete_model_dropdown]
+                )
 
         def save_settings_on_process(*args):
-            apollo_method_value = args[11]
+            """Generator function that forwards progress yields from process_audio."""
+            apollo_method_value = args[15]
             backend_apollo_method = "mid_side_method" if apollo_method_value == i18n("mid_side_method") else "normal_method"
             cleaned_model = clean_model(args[1]) if args[1] else None
+
+            # Compute effective chunk_size based on mode
+            # args[22] = chunk_size_mode, args[23] = chunk_size_custom value
+            cs_mode = args[22] if len(args) > 22 else "base"
+            cs_custom_val = args[23] if len(args) > 23 else 352800
+            cs_base_val = args[2]  # base dropdown value
+
+            if cs_mode == "custom":
+                effective_chunk = int(cs_custom_val) if cs_custom_val else 352800
+            elif cs_mode == "yaml":
+                effective_chunk = "yaml"  # signal processing.py to read from YAML
+            else:
+                effective_chunk = int(cs_base_val) if cs_base_val else 352800
+
             settings = {
-                "chunk_size": args[2],
+                "chunk_size": cs_base_val,
+                "chunk_size_mode": cs_mode,
+                "chunk_size_custom": cs_custom_val,
                 "overlap": args[3],
                 "export_format": args[4],
-                "use_tta": args[5],
-                "use_demud_phaseremix_inst": args[6],
-                "extract_instrumental": args[7],
-                "use_apollo": args[8],
-                "apollo_chunk_size": args[9],
-                "apollo_overlap": args[10],
+                "optimize_mode": args[5],
+                "enable_amp": args[6],
+                "enable_tf32": args[7],
+                "enable_cudnn_benchmark": args[8],
+                "use_tta": args[9],
+                "use_demud_phaseremix_inst": args[10],
+                "extract_instrumental": args[11],
+                "use_apollo": args[12],
+                "apollo_chunk_size": args[13],
+                "apollo_overlap": args[14],
                 "apollo_method": backend_apollo_method,
-                "apollo_normal_model": args[12],
-                "apollo_midside_model": args[13],
-                "use_matchering": args[14],
-                "matchering_passes": args[15],
-                "model_category": args[16],
+                "apollo_normal_model": args[16],
+                "apollo_midside_model": args[17],
+                "use_matchering": args[18],
+                "matchering_passes": args[19],
+                "model_category": args[20],
                 "selected_model": cleaned_model,
-                "auto_ensemble_type": args[17]
+                "auto_ensemble_type": args[11]
             }
             save_config(load_config()["favorites"], settings, load_config()["presets"])
-            modified_args = list(args)
+            # Build args for process_audio (indices 0-21 only, with effective_chunk at [2])
+            modified_args = list(args[:22])
             modified_args[1] = cleaned_model
-            modified_args[17] = cleaned_model
-            return process_audio(*modified_args)
+            modified_args[2] = effective_chunk
+            modified_args[21] = cleaned_model
+            # Forward all yields from process_audio for real-time progress updates
+            for update in process_audio(*modified_args):
+                yield update
 
         def save_auto_ensemble_settings(*args):
+            """Generator function that forwards progress yields from auto_ensemble_process."""
             settings = load_config()["settings"]
             settings["auto_ensemble_type"] = args[7]
             settings["use_matchering"] = args[14]
             settings["matchering_passes"] = args[15]
             save_config(load_config()["favorites"], settings, load_config()["presets"])
-            output_audio, status, progress_html = None, i18n("waiting_for_processing"), ensemble_progress_html.value
+            # Forward all yields from auto_ensemble_process for real-time progress updates
             for update in auto_ensemble_process(*args):
                 if isinstance(update, tuple) and len(update) == 3:
-                    output_audio, status, progress_html = update
-            return output_audio, status, progress_html
+                    yield update
 
         def update_category_dropdowns(cat):
-            logging.debug(f"Input category: {cat}")
-            eng_cat = next((k for k in MODEL_CONFIGS.keys() if i18n(k) == cat), list(MODEL_CONFIGS.keys())[0])
-            logging.debug(f"Using English category: {eng_cat}")
+            all_configs = get_all_model_configs_with_custom()
+            eng_cat = next((k for k in all_configs.keys() if i18n(k) == cat), list(all_configs.keys())[0])
             choices = update_model_dropdown(eng_cat, favorites=load_config()["favorites"])["choices"]
-            logging.debug(f"Model choices: {choices}")
             return gr.update(choices=choices), gr.update(choices=choices)
 
         model_category.change(
@@ -940,39 +1473,38 @@ def create_interface():
         )
 
         auto_category_dropdown.change(
-            fn=lambda cat: gr.update(choices=update_model_dropdown(next((k for k in MODEL_CONFIGS.keys() if i18n(k) == cat), list(MODEL_CONFIGS.keys())[0]), favorites=load_config()["favorites"])["choices"]),
+            fn=lambda cat: gr.update(choices=update_model_dropdown(next((k for k in get_all_model_configs_with_custom().keys() if i18n(k) == cat), list(get_all_model_configs_with_custom().keys())[0]), favorites=load_config()["favorites"])["choices"]),
             inputs=auto_category_dropdown,
             outputs=selected_models
         )
 
-        def debug_inputs(*args):
-            input_names = [
-                "input_audio_file", "model_dropdown", "chunk_size", "overlap", "export_format",
-                "use_tta", "use_demud_phaseremix_inst", "extract_instrumental",
-                "use_apollo", "apollo_chunk_size", "apollo_overlap",
-                "apollo_method", "apollo_normal_model", "apollo_midside_model",
-                "use_matchering", "matchering_passes", "model_category", "selected_model"
-            ]
+        def clean_inputs(*args):
             cleaned_args = list(args)
             cleaned_args[1] = clean_model(cleaned_args[1]) if cleaned_args[1] else None
-            cleaned_args[17] = clean_model(cleaned_args[17]) if cleaned_args[17] else None
-            for name, value in zip(input_names, cleaned_args):
-                logger.debug(f"UI Input - {name}: {value}")
-            return args
+            cleaned_args[21] = clean_model(cleaned_args[21]) if cleaned_args[21] else None
+            return cleaned_args
+
+        def process_wrapper(*args):
+            """Generator wrapper that forwards yields from save_settings_on_process."""
+            for update in save_settings_on_process(*clean_inputs(*args)):
+                yield update
 
         process_btn.click(
-            fn=lambda *args: save_settings_on_process(*debug_inputs(*args)),
+            fn=process_wrapper,
             inputs=[
                 input_audio_file, model_dropdown, chunk_size, overlap, export_format,
+                optimize_mode, enable_amp, enable_tf32, enable_cudnn_benchmark,
                 use_tta, use_demud_phaseremix_inst, extract_instrumental,
                 use_apollo, apollo_chunk_size, apollo_overlap,
                 apollo_method, apollo_normal_model, apollo_midside_model,
-                use_matchering, matchering_passes, model_category, model_dropdown
+                use_matchering, matchering_passes, model_category, model_dropdown,
+                chunk_size_mode, chunk_size_custom
             ],
             outputs=[
                 vocals_audio, instrumental_audio, phaseremix_audio, drum_audio, karaoke_audio,
                 other_audio, bass_audio, effects_audio, speech_audio, bleed_audio, music_audio,
                 dry_audio, male_audio, female_audio,
+                mid_audio, side_audio,
                 separation_process_status, separation_progress_html
             ]
         )
@@ -1016,4 +1548,4 @@ def create_interface():
         refresh_btn.click(fn=update_file_list, outputs=file_dropdown)
         ensemble_process_btn.click(fn=ensemble_audio_fn, inputs=[file_dropdown, ensemble_type, weights_input], outputs=[ensemble_output_audio, ensemble_status])
 
-        return demo
+        return demo

helpers.py

@@ -20,8 +20,19 @@ from datetime import datetime
 import yt_dlp
 import validators
 from pytube import YouTube
-from googleapiclient.discovery import build
-from googleapiclient.http import MediaIoBaseDownload
+
+# Google API imports (optional - for Colab/Google Drive support)
+try:
+    from googleapiclient.discovery import build
+    from googleapiclient.http import MediaIoBaseDownload
+    from google.oauth2.credentials import Credentials
+    GOOGLE_API_AVAILABLE = True
+except ImportError:
+    GOOGLE_API_AVAILABLE = False
+    build = None
+    MediaIoBaseDownload = None
+    Credentials = None
+
 import io
 import math
 import hashlib
@@ -29,13 +40,12 @@ import gc
 import psutil
 import concurrent.futures
 from tqdm import tqdm
-from google.oauth2.credentials import Credentials
 import tempfile
 from urllib.parse import urlparse, quote
 import argparse
 from tqdm.auto import tqdm
 import torch.nn as nn
-from model import get_model_config, MODEL_CONFIGS
+from model import get_model_config, MODEL_CONFIGS, get_all_model_configs_with_custom, load_custom_models
 from assets.i18n.i18n import I18nAuto
 import matchering as mg
 from scipy.signal import find_peaks
@@ -59,7 +69,6 @@ def extract_model_name_from_checkpoint(checkpoint_path):
         return "Unknown"
     base_name = os.path.basename(checkpoint_path)
     model_name = os.path.splitext(base_name)[0]
-    print(f"Original checkpoint path: {checkpoint_path}, extracted model_name: {model_name}")
     return model_name.strip()
 
 for directory in [BASE_DIR, INPUT_DIR, OUTPUT_DIR, OLD_OUTPUT_DIR, AUTO_ENSEMBLE_TEMP, AUTO_ENSEMBLE_OUTPUT, VIDEO_TEMP, ENSEMBLE_DIR]:
@@ -95,7 +104,8 @@ def clean_model(model):
     return cleaned
 
 def get_original_category(translated_category):
-    for original_cat in MODEL_CONFIGS.keys():
+    all_configs = get_all_model_configs_with_custom()
+    for original_cat in all_configs.keys():
         if i18n(original_cat) == translated_category:
             return original_cat
     return None
@@ -109,9 +119,11 @@ def clamp_percentage(value):
         return 0    
 
 def update_model_dropdown(category, favorites=None):
+    # Get all configs including custom models
+    all_configs = get_all_model_configs_with_custom()
     # Map translated category back to English
-    eng_cat = next((k for k in MODEL_CONFIGS.keys() if i18n(k) == category), list(MODEL_CONFIGS.keys())[0])
-    models = MODEL_CONFIGS.get(eng_cat, [])
+    eng_cat = next((k for k in all_configs.keys() if i18n(k) == category), list(all_configs.keys())[0])
+    models = all_configs.get(eng_cat, {})
     choices = []
     favorite_models = []
     non_favorite_models = []
@@ -126,6 +138,11 @@ def update_model_dropdown(category, favorites=None):
     choices = favorite_models + non_favorite_models
     return {"choices": choices}
 
+def get_model_categories():
+    """Get all model categories including Custom Models if any exist."""
+    all_configs = get_all_model_configs_with_custom()
+    return list(all_configs.keys())
+
 def handle_file_upload(uploaded_file, file_path, is_auto_ensemble=False):
     clear_temp_folder("/tmp", exclude_items=["gradio", "config.json"])
     clear_directory(INPUT_DIR)

inference.py

@@ -3,21 +3,16 @@ __author__ = 'Roman Solovyev (ZFTurbo): https://github.com/ZFTurbo/'
 
 import argparse
 import time
-import logging
 import librosa
+from tqdm.auto import tqdm
 import sys
 import os
 import glob
 import torch
+import soundfile as sf
 import torch.nn as nn
 import numpy as np
-import soundfile as sf
-import spaces
-import warnings
-warnings.filterwarnings("ignore")
-
-# Loglama ayarları
-logging.basicConfig(level=logging.DEBUG, filename='utils.log', format='%(asctime)s - %(levelname)s - %(message)s')
+from assets.i18n.i18n import I18nAuto
 
 # Colab kontrolü
 try:
@@ -26,22 +21,26 @@ try:
 except ImportError:
     IS_COLAB = False
 
-# i18n yer tutucu
-class I18nAuto:
-    def __call__(self, message):
-        return message
-    def format(self, message, *args):
-        return message.format(*args)
-
 i18n = I18nAuto()
 
 current_dir = os.path.dirname(os.path.abspath(__file__))
 sys.path.append(current_dir)
 
 from utils import demix, get_model_from_config, normalize_audio, denormalize_audio
-from utils import prefer_target_instrument, apply_tta, load_start_checkpoint
+from utils import prefer_target_instrument, apply_tta, load_start_checkpoint, load_lora_weights
+
+# PyTorch optimized backend (always available)
+try:
+    from pytorch_backend import PyTorchBackend
+    PYTORCH_OPTIMIZED_AVAILABLE = True
+except ImportError:
+    PYTORCH_OPTIMIZED_AVAILABLE = False
+
+import warnings
+warnings.filterwarnings("ignore")
 
 def shorten_filename(filename, max_length=30):
+    """Dosya adını belirtilen maksimum uzunluğa kısaltır."""
     base, ext = os.path.splitext(filename)
     if len(base) <= max_length:
         return filename
@@ -49,217 +48,192 @@ def shorten_filename(filename, max_length=30):
     return shortened
 
 def get_soundfile_subtype(pcm_type, is_float=False):
-    if pcm_type == 'FLOAT' or is_float:
+    """PCM türüne göre uygun soundfile alt türünü belirler."""
+    if is_float:
         return 'FLOAT'
-    subtype_map = {'PCM_16': 'PCM_16', 'PCM_24': 'PCM_24', 'FLOAT': 'FLOAT'}
+    subtype_map = {
+        'PCM_16': 'PCM_16',
+        'PCM_24': 'PCM_24',
+        'FLOAT': 'FLOAT'
+    }
     return subtype_map.get(pcm_type, 'FLOAT')
 
-def update_progress_html(progress_label, progress_percent):
-    progress_percent = min(max(round(progress_percent), 0), 100)
-    return f"""
-    <div id="custom-progress" style="margin-top: 10px;">
-        <div style="font-size: 1rem; color: #C0C0C0; margin-bottom: 5px;" id="progress-label">{progress_label}</div>
-        <div style="width: 100%; background-color: #444; border-radius: 5px; overflow: hidden;">
-            <div id="progress-bar" style="width: {progress_percent}%; height: 20px; background-color: #6e8efb; transition: width 0.3s; max-width: 100%;"></div>
-        </div>
-    </div>
-    """
-
-def run_folder(model, args, config, device, verbose: bool = False, progress=None):
+def run_folder(model, args, config, device, verbose: bool = False):
     start_time = time.time()
     model.eval()
 
     mixture_paths = sorted(glob.glob(os.path.join(args.input_folder, '*.*')))
     sample_rate = getattr(config.audio, 'sample_rate', 44100)
 
-    logging.info(f"Total files found: {len(mixture_paths)} with sample rate: {sample_rate}")
     print(i18n("total_files_found").format(len(mixture_paths), sample_rate))
 
     instruments = prefer_target_instrument(config)[:]
+
+    # Çıktı klasörünü kullan (processing.py tarafından ayarlandı)
     store_dir = args.store_dir
     os.makedirs(store_dir, exist_ok=True)
 
-    total_files = len(mixture_paths)
-    processed_files = 0
-    base_progress_per_file = 100 / total_files if total_files > 0 else 100
+    if not verbose:
+        mixture_paths = tqdm(mixture_paths, desc=i18n("total_progress"))
+    else:
+        mixture_paths = mixture_paths
+
+    detailed_pbar = not args.disable_detailed_pbar
+    print(i18n("detailed_pbar_enabled").format(detailed_pbar))
 
     for path in mixture_paths:
         try:
             mix, sr = librosa.load(path, sr=sample_rate, mono=False)
-            logging.info(f"Loaded audio: {path}, shape: {mix.shape}")
             print(i18n("loaded_audio").format(path, mix.shape))
+        except Exception as e:
+            print(i18n("cannot_read_track").format(path))
+            print(i18n("error_message").format(str(e)))
+            continue
 
-            processed_files += 1
-            base_progress = round((processed_files - 1) * base_progress_per_file)
-            if progress is not None and callable(getattr(progress, '__call__', None)):
-                progress(base_progress / 100, desc=i18n("processing_file").format(processed_files, total_files))
-                update_progress_html(i18n("processing_file").format(processed_files, total_files), base_progress)
-
-            mix_orig = mix.copy()
-            if 'normalize' in config.inference and config.inference.get('normalize', False):
+        mix_orig = mix.copy()
+        if 'normalize' in config.inference:
+            if config.inference['normalize'] is True:
                 mix, norm_params = normalize_audio(mix)
 
-            waveforms_orig = demix(
-                config, model, mix, device, model_type=args.model_type, pbar=False,
-                progress=lambda p, desc: progress((base_progress + p * 50) / 100, desc=desc) if progress else None
-            )
-
-            if args.use_tta:
-                waveforms_orig = apply_tta(
-                    config, model, mix, waveforms_orig, device, args.model_type,
-                    progress=lambda p, desc: progress((base_progress + 50 + p * 20) / 100, desc=desc) if progress else None
-                )
-
-            if args.demud_phaseremix_inst:
-                logging.info(f"Demudding track: {path}")
-                print(i18n("demudding_track").format(path))
-                instr = 'vocals' if 'vocals' in instruments else instruments[0]
-                instruments.append('instrumental_phaseremix')
-                if 'instrumental' not in instruments and 'Instrumental' not in instruments:
-                    mix_modified = mix_orig - 2 * waveforms_orig[instr]
-                    mix_modified_ = mix_modified.copy()
-                    waveforms_modified = demix(
-                        config, model, mix_modified, device, model_type=args.model_type, pbar=False,
-                        progress=lambda p, desc: progress((base_progress + 70 + p * 15) / 100, desc=desc) if progress else None
-                    )
-                    if args.use_tta:
-                        waveforms_modified = apply_tta(
-                            config, model, mix_modified, waveforms_modified, device, args.model_type,
-                            progress=lambda p, desc: progress((base_progress + 85 + p * 10) / 100, desc=desc) if progress else None
-                        )
-                    waveforms_orig['instrumental_phaseremix'] = mix_orig + waveforms_modified[instr]
-                else:
-                    mix_modified = 2 * waveforms_orig[instr] - mix_orig
-                    mix_modified_ = mix_modified.copy()
-                    waveforms_modified = demix(
-                        config, model, mix_modified, device, model_type=args.model_type, pbar=False,
-                        progress=lambda p, desc: progress((base_progress + 70 + p * 15) / 100, desc=desc) if progress else None
-                    )
-                    if args.use_tta:
-                        waveforms_modified = apply_tta(
-                            config, model, mix_modified, waveforms_orig, device, args.model_type,
-                            progress=lambda p, desc: progress((base_progress + 85 + p * 10) / 100, desc=desc) if progress else None
-                        )
-                    waveforms_orig['instrumental_phaseremix'] = mix_orig + mix_modified_ - waveforms_modified[instr]
-
-            if args.extract_instrumental:
-                instr = 'vocals' if 'vocals' in instruments else instruments[0]
-                waveforms_orig['instrumental'] = mix_orig - waveforms_orig[instr]
-                if 'instrumental' not in instruments:
-                    instruments.append('instrumental')
-
-            for i, instr in enumerate(instruments):
-                estimates = waveforms_orig[instr]
-                if 'normalize' in config.inference and config.inference.get('normalize', False):
+        waveforms_orig = demix(config, model, mix, device, model_type=args.model_type, pbar=detailed_pbar)
+
+        if args.use_tta:
+            waveforms_orig = apply_tta(config, model, mix, waveforms_orig, device, args.model_type)
+
+        if args.demud_phaseremix_inst:
+            print(i18n("demudding_track").format(path))
+            instr = 'vocals' if 'vocals' in instruments else instruments[0]
+            instruments.append('instrumental_phaseremix')
+            if 'instrumental' not in instruments and 'Instrumental' not in instruments:
+                mix_modified = mix_orig - 2*waveforms_orig[instr]
+                mix_modified_ = mix_modified.copy()
+                waveforms_modified = demix(config, model, mix_modified, device, model_type=args.model_type, pbar=detailed_pbar)
+                if args.use_tta:
+                    waveforms_modified = apply_tta(config, model, mix_modified, waveforms_modified, device, args.model_type)
+                waveforms_orig['instrumental_phaseremix'] = mix_orig + waveforms_modified[instr]
+            else:
+                mix_modified = 2*waveforms_orig[instr] - mix_orig
+                mix_modified_ = mix_modified.copy()
+                waveforms_modified = demix(config, model, mix_modified, device, model_type=args.model_type, pbar=detailed_pbar)
+                if args.use_tta:
+                    waveforms_modified = apply_tta(config, model, mix_modified, waveforms_orig, device, args.model_type)
+                waveforms_orig['instrumental_phaseremix'] = mix_orig + mix_modified_ - waveforms_modified[instr]
+
+        if args.extract_instrumental:
+            instr = 'vocals' if 'vocals' in instruments else instruments[0]
+            waveforms_orig['instrumental'] = mix_orig - waveforms_orig[instr]
+            if 'instrumental' not in instruments:
+                instruments.append('instrumental')
+
+        for instr in instruments:
+            estimates = waveforms_orig[instr]
+            if 'normalize' in config.inference:
+                if config.inference['normalize'] is True:
                     estimates = denormalize_audio(estimates, norm_params)
 
-                is_float = getattr(args, 'export_format', '').startswith('wav FLOAT')
-                codec = 'flac' if getattr(args, 'flac_file', False) else 'wav'
-                subtype = get_soundfile_subtype(args.pcm_type, is_float=is_float)
+            is_float = getattr(args, 'export_format', '').startswith('wav FLOAT')
+            codec = 'flac' if getattr(args, 'flac_file', False) else 'wav'
+            if codec == 'flac':
+                subtype = get_soundfile_subtype(args.pcm_type, is_float)
+            else:
+                subtype = get_soundfile_subtype('FLOAT', is_float)
 
-                shortened_filename = shorten_filename(os.path.basename(path))
-                output_filename = f"{shortened_filename}_{instr}.{codec}"
-                output_path = os.path.join(store_dir, output_filename)
-                sf.write(output_path, estimates.T, sr, subtype=subtype)
+            shortened_filename = shorten_filename(os.path.basename(path))
+            output_filename = f"{shortened_filename}_{instr}.{codec}"
+            output_path = os.path.join(store_dir, output_filename)
+            sf.write(output_path, estimates.T, sr, subtype=subtype)
 
-                save_progress = round(base_progress + 95 + (i / len(instruments)) * 5)
-                if progress is not None and callable(getattr('progress', '__call__', None)):
-                    progress(save_progress / 100, desc=i18n("saving_output").format(instr, processed_files, total_files))
-                    update_progress_html(i18n("saving_output").format(instr, processed_files, total_files), save_progress)
+    print(i18n("elapsed_time").format(time.time() - start_time))
 
-            file_progress = round(processed_files * base_progress_per_file)
-            if progress is not None and callable(getattr(progress, '__call__', None)):
-                progress(file_progress / 100, desc=i18n("completed_file").format(processed_files, total_files))
-                update_progress_html(i18n("completed_file").format(processed_files, total_files), file_progress)
-
-        except Exception as e:
-            logging.error(f"Cannot read track: {path}. Error: {str(e)}")
-            print(i18n("cannot_read_track").format(path))
-            print(i18n("error_message").format(str(e)))
-            continue
-
-    elapsed_time = time.time() - start_time
-    logging.info(f"Processing time: {elapsed_time:.2f} seconds")
-    print(i18n("elapsed_time").format(elapsed_time))
-
-    if progress is not None and callable(getattr(progress, '__call__', None)):
-        progress(1.0, desc=i18n("processing_complete"))
-        update_progress_html(i18n("processing_complete"), 100)
-
-@spaces.GPU
-def proc_folder(args=None, progress=None):
-    try:
-        parser = argparse.ArgumentParser(description=i18n("proc_folder_description"))
-        parser.add_argument("--model_type", type=str, default='melod_band_roformer', help=i18n("model_type_help"))
-        parser.add_argument("--config_path", type=str, required=True, help=i18n("config_path_help"))
-        parser.add_argument("--start_check_point", type=str, required=True, help=i18n("start_checkpoint_help"))
-        parser.add_argument("--input_folder", type=str, required=True, help=i18n("input_folder_help"))
-        parser.add_argument("--store_dir", type=str, required=True, help=i18n("store_dir_help"))
-        parser.add_argument("--chunk_size", type=int, default=352800, help=i18n("chunk_size_help"))
-        parser.add_argument("--overlap", type=int, default=2, help=i18n("overlap_help"))
-        parser.add_argument("--export_format", type=str, default='wav FLOAT', choices=['wav FLOAT', 'flac PCM_16', 'flac PCM_24'], help=i18n("export_format_help"))
-        parser.add_argument("--demud_phaseremix_inst", action='store_true', help=i18n("demud_phaseremix_help"))
-        parser.add_argument("--extract_instrumental", action='store_true', help=i18n("extract_instrumental_help"))
-        parser.add_argument("--use_tta", action='store_true', help=i18n("use_tta_help"))
-        parser.add_argument("--flac_file", action='store_true', help=i18n("flac_file_help"))
-        parser.add_argument("--pcm_type", type=str, choices=['PCM_16', 'PCM_24'], default='PCM_24', help=i18n("pcm_type_help"))
-        parser.add_argument("--device_ids", nargs='+', type=int, default=[0], help=i18n("device_ids_help"))
-        parser.add_argument("--force_cpu", action='store_true', help=i18n("force_cpu_help"))
-        parser.add_argument("--lora_checkpoint", type=str, default='', help=i18n("lora_checkpoint_help"))
-
-        args = parser.parse_args(args if args else [])
-    except Exception as e:
-        logging.error(f"Argument parsing failed: {str(e)}")
-        raise ValueError(f"Invalid command-line arguments: {str(e)}")
+def proc_folder(args, use_tensorrt=False):
+    """
+    Process folder with optional TensorRT backend.
+    
+    Parameters:
+    ----------
+    args : list or None
+        Command line arguments
+    use_tensorrt : bool
+        Use TensorRT backend if available
+    """
+    parser = argparse.ArgumentParser(description=i18n("proc_folder_description"))
+    parser.add_argument("--model_type", type=str, default='mdx23c', help=i18n("model_type_help"))
+    parser.add_argument("--config_path", type=str, help=i18n("config_path_help"))
+    parser.add_argument("--demud_phaseremix_inst", action='store_true', help=i18n("demud_phaseremix_help"))
+    parser.add_argument("--start_check_point", type=str, default='', help=i18n("start_checkpoint_help"))
+    parser.add_argument("--input_folder", type=str, help=i18n("input_folder_help"))
+    parser.add_argument("--audio_path", type=str, help=i18n("audio_path_help"))
+    parser.add_argument("--store_dir", type=str, default="", help=i18n("store_dir_help"))
+    parser.add_argument("--device_ids", nargs='+', type=int, default=0, help=i18n("device_ids_help"))
+    parser.add_argument("--extract_instrumental", action='store_true', help=i18n("extract_instrumental_help"))
+    parser.add_argument("--disable_detailed_pbar", action='store_true', help=i18n("disable_detailed_pbar_help"))
+    parser.add_argument("--force_cpu", action='store_true', help=i18n("force_cpu_help"))
+    parser.add_argument("--flac_file", action='store_true', help=i18n("flac_file_help"))
+    parser.add_argument("--export_format", type=str, choices=['wav FLOAT', 'flac PCM_16', 'flac PCM_24'], default='flac PCM_24', help=i18n("export_format_help"))
+    parser.add_argument("--pcm_type", type=str, choices=['PCM_16', 'PCM_24'], default='PCM_24', help=i18n("pcm_type_help"))
+    parser.add_argument("--use_tta", action='store_true', help=i18n("use_tta_help"))
+    parser.add_argument("--lora_checkpoint", type=str, default='', help=i18n("lora_checkpoint_help"))
+    parser.add_argument("--chunk_size", type=int, default=1000000, help="Inference chunk size")
+    parser.add_argument("--overlap", type=int, default=4, help="Inference overlap factor")
+    parser.add_argument("--optimize_mode", type=str, choices=['default', 'compile', 'jit', 'channels_last'], default='channels_last', help="PyTorch optimization mode (always enabled)")
+    parser.add_argument("--enable_amp", action='store_true', default=True, help="Enable automatic mixed precision")
+    parser.add_argument("--enable_tf32", action='store_true', default=True, help="Enable TF32 (Ampere GPUs)")
+    parser.add_argument("--enable_cudnn_benchmark", action='store_true', default=True, help="Enable cuDNN benchmark")
+
+    if args is None:
+        args = parser.parse_args()
+    else:
+        args = parser.parse_args(args)
 
     device = "cpu"
     if args.force_cpu:
-        logging.info("Forced to use CPU")
+        device = "cpu"
     elif torch.cuda.is_available():
-        logging.info("CUDA available")
         print(i18n("cuda_available"))
-        device = f'cuda:{args.device_ids[0]}'
+        device = f'cuda:{args.device_ids[0]}' if type(args.device_ids) == list else f'cuda:{args.device_ids}'
     elif torch.backends.mps.is_available():
-        device = "mps"
+         device = "mps"
 
-    logging.info(f"Using device: {device}")
     print(i18n("using_device").format(device))
 
     model_load_start_time = time.time()
     torch.backends.cudnn.benchmark = True
 
-    try:
-        model, config = get_model_from_config(args.model_type, args.config_path)
-    except Exception as e:
-        logging.error(f"Failed to load model: {str(e)}")
-        raise
+    model, config = get_model_from_config(args.model_type, args.config_path)
 
-    if args.start_check_point:
-        try:
-            load_start_checkpoint(args, model, type_='inference')
-        except Exception as e:
-            logging.error(f"Failed to load checkpoint: {str(e)}")
-            raise
+    if args.start_check_point != '':
+        load_start_checkpoint(args, model, type_='inference')
 
-    logging.info(f"Instruments: {config.training.instruments}")
     print(i18n("instruments_print").format(config.training.instruments))
 
-    if len(args.device_ids) > 1 and not args.force_cpu:
+    if type(args.device_ids) == list and len(args.device_ids) > 1 and not args.force_cpu:
         model = nn.DataParallel(model, device_ids=args.device_ids)
-        logging.info(f"Using DataParallel with devices: {args.device_ids}")
 
     model = model.to(device)
 
-    elapsed_time = time.time() - model_load_start_time
-    logging.info(f"Model load time: {elapsed_time:.2f} seconds")
-    print(i18n("model_load_time").format(elapsed_time))
-
-    run_folder(model, args, config, device, verbose=False, progress=progress)
-    return "Processing completed"
+    print(i18n("model_load_time").format(time.time() - model_load_start_time))
+
+    # Always use optimized PyTorch backend if available
+    if PYTORCH_OPTIMIZED_AVAILABLE:
+        print(f"Using optimized PyTorch backend")
+        print(f"   Mode: {args.optimize_mode}")
+        print(f"   AMP: {args.enable_amp} | TF32: {args.enable_tf32} | cuDNN: {args.enable_cudnn_benchmark}")
+        from inference_pytorch import proc_folder_pytorch_optimized
+        # Recreate args for optimized PyTorch inference
+        sys.argv = sys.argv[:1]  # Keep only script name
+        for key, value in vars(args).items():
+            if value is not None and value is not False:
+                if isinstance(value, bool):
+                    sys.argv.append(f"--{key}")
+                elif isinstance(value, list):
+                    sys.argv.append(f"--{key}")
+                    sys.argv.extend(map(str, value))
+                else:
+                    sys.argv.extend([f"--{key}", str(value)])
+        proc_folder_pytorch_optimized(None)
+    else:
+        print("Warning: PyTorch optimized backend not available, using standard inference")
+        run_folder(model, args, config, device, verbose=False)
 
 if __name__ == "__main__":
-    try:
-        proc_folder(None)
-    except Exception as e:
-        logging.error(f"Main execution failed: {str(e)}")
-        raise
+    proc_folder(None)

inference_pytorch.py

@@ -0,0 +1,390 @@
+# coding: utf-8
+__author__ = 'PyTorch Optimized Inference Implementation'
+
+import argparse
+import time
+import librosa
+from tqdm.auto import tqdm
+import sys
+import os
+import glob
+import torch
+import soundfile as sf
+import torch.nn as nn
+import numpy as np
+import pickle
+from assets.i18n.i18n import I18nAuto
+
+# Set inference path for compatibility
+INFERENCE_PATH = os.path.abspath(__file__)
+
+i18n = I18nAuto()
+
+current_dir = os.path.dirname(os.path.abspath(__file__))
+sys.path.append(current_dir)
+
+from utils import get_model_from_config, normalize_audio, denormalize_audio
+from utils import prefer_target_instrument, load_start_checkpoint, apply_tta, demix
+from pytorch_backend import PyTorchBackend, PyTorchOptimizer, create_inference_session
+
+import warnings
+warnings.filterwarnings("ignore")
+
+
+def shorten_filename(filename, max_length=30):
+    """Dosya adını belirtilen maksimum uzunluğa kısaltır."""
+    base, ext = os.path.splitext(filename)
+    if len(base) <= max_length:
+        return filename
+    shortened = base[:15] + "..." + base[-10:] + ext
+    return shortened
+
+
+def get_soundfile_subtype(pcm_type, is_float=False):
+    """PCM türüne göre uygun soundfile alt türünü belirler."""
+    if is_float:
+        return 'FLOAT'
+    subtype_map = {
+        'PCM_16': 'PCM_16',
+        'PCM_24': 'PCM_24',
+        'FLOAT': 'FLOAT'
+    }
+    return subtype_map.get(pcm_type, 'FLOAT')
+
+
+def demix_pytorch_optimized(
+    config,
+    backend: PyTorchBackend,
+    mix: np.ndarray,
+    device: torch.device,
+    pbar: bool = False
+) -> dict:
+    """
+    Optimized PyTorch backend ile audio source separation.
+    
+    Parameters:
+    ----------
+    config : ConfigDict
+        Configuration object
+    backend : PyTorchBackend
+        PyTorch backend with optimized model
+    mix : np.ndarray
+        Input audio array
+    device : torch.device
+        Computation device
+    pbar : bool
+        Show progress bar
+        
+    Returns:
+    -------
+    dict
+        Dictionary of separated sources
+    """
+    mix = torch.tensor(mix, dtype=torch.float32)
+    
+    chunk_size = config.audio.chunk_size
+    num_instruments = len(prefer_target_instrument(config))
+    num_overlap = config.inference.num_overlap
+    
+    fade_size = chunk_size // 10
+    step = chunk_size // num_overlap
+    border = chunk_size - step
+    length_init = mix.shape[-1]
+    
+    # Windowing array
+    fadein = torch.linspace(0, 1, fade_size)
+    fadeout = torch.linspace(1, 0, fade_size)
+    windowing_array = torch.ones(chunk_size)
+    windowing_array[-fade_size:] = fadeout
+    windowing_array[:fade_size] = fadein
+    
+    # Add padding
+    if length_init > 2 * border and border > 0:
+        mix = nn.functional.pad(mix, (border, border), mode="reflect")
+    
+    batch_size = config.inference.batch_size
+    use_amp = getattr(config.training, 'use_amp', True)
+    
+    with torch.cuda.amp.autocast(enabled=use_amp):
+        with torch.inference_mode():
+            # Initialize result and counter tensors
+            req_shape = (num_instruments,) + mix.shape
+            result = torch.zeros(req_shape, dtype=torch.float32)
+            counter = torch.zeros(req_shape, dtype=torch.float32)
+            
+            i = 0
+            batch_data = []
+            batch_locations = []
+            
+            # Progress reporting for GUI (no terminal tqdm)
+            total_samples = mix.shape[1]
+            last_reported_percent = -1
+            
+            while i < mix.shape[1]:
+                # Extract chunk
+                part = mix[:, i:i + chunk_size].to(device)
+                chunk_len = part.shape[-1]
+                
+                if chunk_len > chunk_size // 2:
+                    pad_mode = "reflect"
+                else:
+                    pad_mode = "constant"
+                    
+                part = nn.functional.pad(
+                    part, 
+                    (0, chunk_size - chunk_len), 
+                    mode=pad_mode, 
+                    value=0
+                )
+                
+                batch_data.append(part)
+                batch_locations.append((i, chunk_len))
+                i += step
+                
+                # Process batch
+                if len(batch_data) >= batch_size or i >= mix.shape[1]:
+                    arr = torch.stack(batch_data, dim=0)
+                    
+                    # Use optimized PyTorch backend for inference
+                    x = backend(arr)
+                    
+                    window = windowing_array.clone()
+                    if i - step == 0:  # First chunk
+                        window[:fade_size] = 1
+                    elif i >= mix.shape[1]:  # Last chunk
+                        window[-fade_size:] = 1
+                    
+                    for j, (start, seg_len) in enumerate(batch_locations):
+                        result[..., start:start + seg_len] += x[j, ..., :seg_len].cpu() * window[..., :seg_len]
+                        counter[..., start:start + seg_len] += window[..., :seg_len]
+                    
+                    batch_data.clear()
+                    batch_locations.clear()
+                
+                # Report real progress percentage for GUI capture (every 1% for smooth updates)
+                # Use unique prefix [SESA_PROGRESS] to avoid confusion with other log messages
+                current_percent = int((i / total_samples) * 100)
+                if current_percent > last_reported_percent:
+                    last_reported_percent = current_percent
+                    print(f"[SESA_PROGRESS]{current_percent}", flush=True)
+            
+            print("[SESA_PROGRESS]100", flush=True)
+            
+            # Compute final estimated sources
+            estimated_sources = result / counter
+            estimated_sources = estimated_sources.cpu().numpy()
+            np.nan_to_num(estimated_sources, copy=False, nan=0.0)
+            
+            # Remove padding
+            if length_init > 2 * border and border > 0:
+                estimated_sources = estimated_sources[..., border:-border]
+    
+    # Return as dictionary
+    instruments = prefer_target_instrument(config)
+    ret_data = {k: v for k, v in zip(instruments, estimated_sources)}
+    
+    return ret_data
+
+
+def run_folder_pytorch_optimized(backend, args, config, device, model=None, verbose: bool = False):
+    """
+    PyTorch backend ile klasör işleme.
+    """
+    start_time = time.time()
+    
+    mixture_paths = sorted(glob.glob(os.path.join(args.input_folder, '*.*')))
+    sample_rate = getattr(config.audio, 'sample_rate', 44100)
+    
+    print(f"PyTorch Backend | {len(mixture_paths)} dosya | SR: {sample_rate}")
+    
+    instruments = prefer_target_instrument(config)[:]
+    
+    # Çıktı klasörünü kullan
+    store_dir = args.store_dir
+    os.makedirs(store_dir, exist_ok=True)
+    
+    # Progress is reported via print statements for GUI capture (no terminal tqdm)
+    total_files = len(mixture_paths)
+    detailed_pbar = not args.disable_detailed_pbar
+    print(i18n("detailed_pbar_enabled").format(detailed_pbar))
+    
+    for file_idx, path in enumerate(mixture_paths):
+        try:
+            mix, sr = librosa.load(path, sr=sample_rate, mono=False)
+            print(i18n("loaded_audio").format(path, mix.shape))
+        except Exception as e:
+            print(i18n("cannot_read_track").format(path))
+            print(i18n("error_message").format(str(e)))
+            continue
+        
+        mix_orig = mix.copy()
+        if 'normalize' in config.inference:
+            if config.inference['normalize'] is True:
+                mix, norm_params = normalize_audio(mix)
+        
+        # Use optimized PyTorch backend
+        waveforms_orig = demix_pytorch_optimized(config, backend, mix, device, pbar=detailed_pbar)
+        
+        if args.use_tta and model is not None:
+            waveforms_orig = apply_tta(config, model, mix, waveforms_orig, device, args.model_type)
+
+        if args.demud_phaseremix_inst and model is not None:
+            print(f"DemudPhaseRemix: {path}")
+            instr = 'vocals' if 'vocals' in instruments else instruments[0]
+            instruments.append('instrumental_phaseremix')
+            if 'instrumental' not in instruments and 'Instrumental' not in instruments:
+                mix_modified = mix_orig - 2 * waveforms_orig[instr]
+                mix_modified_ = mix_modified.copy()
+                waveforms_modified = demix(config, model, mix_modified, device, model_type=args.model_type)
+                if args.use_tta:
+                    waveforms_modified = apply_tta(config, model, mix_modified, waveforms_modified, device, args.model_type)
+                waveforms_orig['instrumental_phaseremix'] = mix_orig + waveforms_modified[instr]
+            else:
+                mix_modified = 2 * waveforms_orig[instr] - mix_orig
+                mix_modified_ = mix_modified.copy()
+                waveforms_modified = demix(config, model, mix_modified, device, model_type=args.model_type)
+                if args.use_tta:
+                    waveforms_modified = apply_tta(config, model, mix_modified, waveforms_orig, device, args.model_type)
+                waveforms_orig['instrumental_phaseremix'] = mix_orig + mix_modified_ - waveforms_modified[instr]
+
+        if args.extract_instrumental:
+            instr = 'vocals' if 'vocals' in instruments else instruments[0]
+            waveforms_orig['instrumental'] = mix_orig - waveforms_orig[instr]
+            if 'instrumental' not in instruments:
+                instruments.append('instrumental')
+        
+        for instr in instruments:
+            estimates = waveforms_orig[instr]
+            if 'normalize' in config.inference:
+                if config.inference['normalize'] is True:
+                    estimates = denormalize_audio(estimates, norm_params)
+            
+            is_float = getattr(args, 'export_format', '').startswith('wav FLOAT')
+            codec = 'flac' if getattr(args, 'flac_file', False) else 'wav'
+            if codec == 'flac':
+                subtype = get_soundfile_subtype(args.pcm_type, is_float)
+            else:
+                subtype = get_soundfile_subtype('FLOAT', is_float)
+            
+            shortened_filename = shorten_filename(os.path.basename(path))
+            output_filename = f"{shortened_filename}_{instr}.{codec}"
+            output_path = os.path.join(store_dir, output_filename)
+            sf.write(output_path, estimates.T, sr, subtype=subtype)
+    
+    print(i18n("elapsed_time").format(time.time() - start_time))
+
+
+def proc_folder_pytorch_optimized(args):
+    """
+    PyTorch ile inference işleme fonksiyonu.
+    """
+    parser = argparse.ArgumentParser(description="PyTorch Inference for Music Source Separation")
+    parser.add_argument("--model_type", type=str, default='mdx23c', help="Model type")
+    parser.add_argument("--config_path", type=str, help="Config path")
+    parser.add_argument("--start_check_point", type=str, default='', help="Checkpoint path (.ckpt)")
+    parser.add_argument("--input_folder", type=str, help="Input folder path")
+    parser.add_argument("--store_dir", type=str, default="", help="Output directory")
+    parser.add_argument("--device_ids", nargs='+', type=int, default=0, help="Device IDs")
+    parser.add_argument("--extract_instrumental", action='store_true', help="Extract instrumental")
+    parser.add_argument("--disable_detailed_pbar", action='store_true', help="Disable detailed progress bar")
+    parser.add_argument("--flac_file", action='store_true', help="Output as FLAC")
+    parser.add_argument("--export_format", type=str, choices=['wav FLOAT', 'flac PCM_16', 'flac PCM_24'], 
+                        default='flac PCM_24', help="Export format")
+    parser.add_argument("--pcm_type", type=str, choices=['PCM_16', 'PCM_24'], default='PCM_24', help="PCM type")
+    parser.add_argument("--chunk_size", type=int, default=1000000, help="Inference chunk size")
+    parser.add_argument("--overlap", type=int, default=4, help="Inference overlap factor")
+    parser.add_argument("--optimize_mode", type=str, choices=['channels_last', 'compile', 'jit', 'default'], 
+                        default='channels_last', help="PyTorch optimization mode (channels_last recommended)")
+    parser.add_argument("--enable_amp", action='store_true', help="Enable automatic mixed precision (2x faster)")
+    parser.add_argument("--enable_tf32", action='store_true', help="Enable TF32 for RTX 30xx+ (faster)")
+    parser.add_argument("--enable_cudnn_benchmark", action='store_true', help="Enable cuDNN benchmark (faster after warmup)")
+    parser.add_argument("--lora_checkpoint", type=str, default='', help="Initial checkpoint to LoRA weights")
+    parser.add_argument("--use_tta", action='store_true', help="Test Time Augmentation (flips + polarity)")
+    parser.add_argument("--demud_phaseremix_inst", action='store_true', help="DemudPhaseRemix instrumental extraction")
+    
+    if args is None:
+        args = parser.parse_args()
+    else:
+        args = parser.parse_args(args)
+    
+    # Device setup
+    device = "cpu"
+    if torch.cuda.is_available():
+        print(i18n("cuda_available"))
+        device = f'cuda:{args.device_ids[0]}' if type(args.device_ids) == list else f'cuda:{args.device_ids}'
+    elif torch.backends.mps.is_available():
+        device = "mps"
+        print("Using MPS (Metal) backend")
+    
+    print(i18n("using_device").format(device))
+    
+    # Load model
+    model_load_start_time = time.time()
+    
+    model, config = get_model_from_config(args.model_type, args.config_path)
+    
+    if args.start_check_point != '':
+        try:
+            checkpoint = torch.load(args.start_check_point, map_location=device, weights_only=False)
+        except (pickle.UnpicklingError, RuntimeError, EOFError) as e:
+            error_details = f"""
+CHECKPOINT FILE CORRUPTED
+
+Error: {str(e)}
+
+The checkpoint file appears to be corrupted or was not downloaded correctly.
+File: {args.start_check_point}
+
+Common causes:
+  - File is an HTML page (wrong download URL, e.g., HuggingFace /blob/ instead of /resolve/)
+  - Incomplete or interrupted download
+  - Network issues during download
+  - File system corruption
+
+Solution:
+  1. Delete the corrupted checkpoint file:
+     {args.start_check_point}
+  2. Re-run the application - it will automatically re-download the model
+  3. If the problem persists, check that your model URL uses /resolve/ not /blob/
+     Example: https://huggingface.co/user/repo/resolve/main/model.ckpt
+"""
+            print(error_details)
+            import sys
+            sys.exit(1)
+        
+        # Handle different checkpoint formats
+        if isinstance(checkpoint, dict):
+            if 'state_dict' in checkpoint:
+                state_dict = checkpoint['state_dict']
+            elif 'model' in checkpoint:
+                state_dict = checkpoint['model']
+            elif 'state' in checkpoint:
+                state_dict = checkpoint['state']
+            else:
+                state_dict = checkpoint
+        else:
+            state_dict = checkpoint
+        
+        model.load_state_dict(state_dict, strict=False)
+        model = model.eval().to(device)
+    
+    print(i18n("instruments_print").format(config.training.instruments))
+    
+    # Create optimized PyTorch backend
+    backend = create_inference_session(
+        model=model,
+        device=device,
+        optimize_mode=args.optimize_mode,
+        enable_amp=args.enable_amp,
+        enable_tf32=args.enable_tf32,
+        enable_cudnn_benchmark=args.enable_cudnn_benchmark
+    )
+    
+    print(i18n("model_load_time").format(time.time() - model_load_start_time))
+    
+    # Run inference (pass raw model for TTA/demud support)
+    run_folder_pytorch_optimized(backend, args, config, device, model=model, verbose=False)
+
+
+if __name__ == "__main__":
+    proc_folder_pytorch_optimized(None)

main.py

@@ -14,7 +14,15 @@ from datetime import datetime
 import numpy as np
 import shutil
 from gui import create_interface
-from pyngrok import ngrok
+
+# pyngrok import (optional - only needed for ngrok sharing)
+try:
+    from pyngrok import ngrok
+    NGROK_AVAILABLE = True
+except ImportError:
+    NGROK_AVAILABLE = False
+    ngrok = None
+
 from assets.i18n.i18n import I18nAuto  # I18nAuto'yu içe aktar
 
 import warnings
@@ -56,6 +64,9 @@ def start_localtunnel(port, i18n):
 
 def start_ngrok(port, ngrok_token, i18n):
     """Starts the Gradio interface with ngrok sharing."""
+    if not NGROK_AVAILABLE:
+        print("pyngrok modülü yüklü değil. 'pip install pyngrok' ile yükleyin.")
+        sys.exit(1)
     print(i18n("starting_ngrok").format(port=port))
     try:
         ngrok.set_auth_token(ngrok_token)

model.py

@@ -1,58 +1,600 @@
 import os
 import yaml
-from urllib.parse import quote
+import json
+import re
+import shutil
+from urllib.parse import quote, urlparse
 from pathlib import Path
 
 # Temel dizin ve checkpoint dizini sabit olarak tanımlanıyor
 BASE_DIR = os.path.dirname(os.path.abspath(__file__))
 CHECKPOINT_DIR = os.path.join(BASE_DIR, 'ckpts')
+CUSTOM_MODELS_FILE = os.path.join(BASE_DIR, 'assets', 'custom_models.json')
 
-def conf_edit(config_path, chunk_size, overlap):
-    """Edits the configuration file with chunk size and overlap."""
+
+def fix_huggingface_url(url):
+    """Convert Hugging Face blob URLs to raw/resolve URLs.
+    
+    Hugging Face has two URL formats:
+    - /blob/ URLs show the web page (HTML) - WRONG for downloading
+    - /resolve/ URLs provide the raw file content - CORRECT for downloading
+    
+    This function converts blob URLs to resolve URLs automatically.
+    
+    Args:
+        url: The URL to fix
+        
+    Returns:
+        The corrected URL (or original if not a HF blob URL)
+    """
+    if not url:
+        return url
+    
+    # Check if it's a Hugging Face URL with /blob/
+    if 'huggingface.co' in url and '/blob/' in url:
+        fixed_url = url.replace('/blob/', '/resolve/')
+        return fixed_url
+    
+    return url
+
+
+def validate_yaml_content(content, filepath=None):
+    """Validate that content is YAML and not HTML.
+    
+    Args:
+        content: The file content to validate
+        filepath: Optional filepath for error messages
+        
+    Returns:
+        tuple: (is_valid: bool, error_message: str or None)
+    """
+    # Check if content looks like HTML
+    html_indicators = [
+        '<!DOCTYPE',
+        '<html',
+        '<head>',
+        '<body>',
+        '<script>',
+        '<link rel=',
+        'text/html',
+    ]
+    
+    content_lower = content.lower() if isinstance(content, str) else content.decode('utf-8', errors='ignore').lower()
+    
+    for indicator in html_indicators:
+        if indicator.lower() in content_lower:
+            error_msg = f"""
+The downloaded file appears to be an HTML page, not a YAML config file.
+{"File: " + filepath if filepath else ""}
+
+This usually happens when using a Hugging Face '/blob/' URL instead of a '/resolve/' URL.
+
+To fix this:
+1. Use the raw file URL with '/resolve/' instead of '/blob/'
+   Example: https://huggingface.co/user/repo/resolve/main/file.yaml
+   
+2. Or copy the raw URL from Hugging Face:
+   - Go to the file on Hugging Face
+   - Click "Download" or right-click and "Copy link address"
+"""
+            return False, error_msg
+    
+    return True, None
+
+# Supported model types for auto-detection and manual selection
+SUPPORTED_MODEL_TYPES = [
+    'bs_roformer',
+    'bs_roformer_custom',
+    'mel_band_roformer',
+    'mdx23c',
+    'bandit_v2',
+    'scnet',
+    'htdemucs',
+    'torchseg'
+]
+
+def detect_model_type_from_url(checkpoint_url, config_url=None):
+    """Auto-detect model type from URL patterns."""
+    urls_to_check = [checkpoint_url]
+    if config_url:
+        urls_to_check.append(config_url)
+    
+    combined_text = ' '.join(urls_to_check).lower()
+    
+    patterns = [
+        (r'bs[-_]?roformer[-_]?custom|hyperace', 'bs_roformer_custom'),
+        (r'bs[-_]?roformer|bsroformer', 'bs_roformer'),
+        (r'mel[-_]?band[-_]?roformer|melbandroformer|mbr', 'mel_band_roformer'),
+        (r'mdx23c', 'mdx23c'),
+        (r'bandit[-_]?v?2?', 'bandit_v2'),
+        (r'scnet', 'scnet'),
+        (r'htdemucs|demucs', 'htdemucs'),
+        (r'torchseg', 'torchseg'),
+    ]
+    
+    for pattern, model_type in patterns:
+        if re.search(pattern, combined_text):
+            return model_type
+    return None
+
+def detect_model_type_from_config(config_url):
+    """Try to detect model type by downloading and parsing config YAML."""
+    try:
+        import requests
+        response = requests.get(config_url, timeout=10)
+        if response.status_code == 200:
+            config_data = yaml.safe_load(response.text)
+            if 'model_type' in config_data:
+                return config_data['model_type']
+            if 'model' in config_data and 'model_type' in config_data['model']:
+                return config_data['model']['model_type']
+    except Exception:
+        pass
+    return None
+
+def load_custom_models():
+    """Load custom models from JSON file."""
+    if not os.path.exists(CUSTOM_MODELS_FILE):
+        return {}
+    try:
+        with open(CUSTOM_MODELS_FILE, 'r', encoding='utf-8') as f:
+            return json.load(f)
+    except (json.JSONDecodeError, IOError):
+        return {}
+
+def save_custom_models(models):
+    """Save custom models to JSON file."""
+    os.makedirs(os.path.dirname(CUSTOM_MODELS_FILE), exist_ok=True)
+    with open(CUSTOM_MODELS_FILE, 'w', encoding='utf-8') as f:
+        json.dump(models, f, indent=2, ensure_ascii=False)
+
+def add_custom_model(model_name, model_type, checkpoint_url, config_url, custom_model_url=None, auto_detect=True):
+    """Add a new custom model."""
+    if not model_name or not model_name.strip():
+        return False, "Model name is required"
+    if not checkpoint_url or not checkpoint_url.strip():
+        return False, "Checkpoint URL is required"
+    if not config_url or not config_url.strip():
+        return False, "Config URL is required"
+    
+    model_name = model_name.strip()
+    checkpoint_url = checkpoint_url.strip()
+    config_url = config_url.strip()
+    custom_model_url = custom_model_url.strip() if custom_model_url else None
+    
+    # Auto-fix Hugging Face URLs
+    checkpoint_url = fix_huggingface_url(checkpoint_url)
+    config_url = fix_huggingface_url(config_url)
+    if custom_model_url:
+        custom_model_url = fix_huggingface_url(custom_model_url)
+    
+    if auto_detect and (not model_type or model_type == "auto"):
+        detected_type = detect_model_type_from_url(checkpoint_url, config_url)
+        if not detected_type:
+            detected_type = detect_model_type_from_config(config_url)
+        if detected_type:
+            model_type = detected_type
+        else:
+            return False, "Could not auto-detect model type. Please select manually."
+    
+    if model_type not in SUPPORTED_MODEL_TYPES:
+        return False, f"Unsupported model type: {model_type}"
+    
+    checkpoint_filename = os.path.basename(checkpoint_url.split('?')[0])
+    config_filename = f"config_{model_name.replace(' ', '_').lower()}.yaml"
+    
+    models = load_custom_models()
+    if model_name in models:
+        return False, f"Model '{model_name}' already exists"
+    
+    models[model_name] = {
+        'model_type': model_type,
+        'checkpoint_url': checkpoint_url,
+        'config_url': config_url,
+        'custom_model_url': custom_model_url,
+        'checkpoint_filename': checkpoint_filename,
+        'config_filename': config_filename,
+        'needs_conf_edit': True
+    }
+    save_custom_models(models)
+    return True, f"Model '{model_name}' added successfully"
+
+def delete_custom_model(model_name):
+    """Delete a custom model."""
+    models = load_custom_models()
+    if model_name not in models:
+        return False, f"Model '{model_name}' not found"
+    
+    model_config = models[model_name]
+    checkpoint_path = os.path.join(CHECKPOINT_DIR, model_config.get('checkpoint_filename', ''))
+    config_path = os.path.join(CHECKPOINT_DIR, model_config.get('config_filename', ''))
+    
+    try:
+        if os.path.exists(checkpoint_path):
+            os.remove(checkpoint_path)
+        if os.path.exists(config_path):
+            os.remove(config_path)
+    except Exception:
+        pass
+    
+    del models[model_name]
+    save_custom_models(models)
+    return True, f"Model '{model_name}' deleted successfully"
+
+def get_custom_models_list():
+    """Get list of custom model names with their types."""
+    models = load_custom_models()
+    return [(name, config.get('model_type', 'unknown')) for name, config in models.items()]
+
+def preprocess_yaml_content(content):
+    """Pre-process YAML content to fix common issues before parsing.
+    
+    Fixes:
+    - Replaces tabs with spaces
+    - Attempts to quote unquoted URLs and paths containing colons
+    """
+    # Replace tabs with spaces
+    if '\t' in content:
+        content = content.replace('\t', '    ')
+    
+    # Fix unquoted URLs/paths with colons in values (common issue)
+    # This regex finds lines like "key: http://..." or "key: C:\path" and quotes the value
+    lines = content.split('\n')
+    fixed_lines = []
+    
+    for line in lines:
+        # Skip comments and empty lines
+        stripped = line.strip()
+        if not stripped or stripped.startswith('#'):
+            fixed_lines.append(line)
+            continue
+        
+        # Check if line has a key-value pattern with potential problematic value
+        # Match: "  key: value_with_colon_or_backslash"
+        match = re.match(r'^(\s*)([^:#]+?):\s+(.+)$', line)
+        if match:
+            indent, key, value = match.groups()
+            # Check if value contains a colon (like URL) or backslash (like Windows path)
+            # and is not already quoted
+            if ((':' in value or '\\' in value) and 
+                not (value.startswith('"') and value.endswith('"')) and
+                not (value.startswith("'") and value.endswith("'"))):
+                # Quote the value
+                escaped_value = value.replace('"', '\\"')
+                fixed_lines.append(f'{indent}{key}: "{escaped_value}"')
+                continue
+        
+        fixed_lines.append(line)
+    
+    return '\n'.join(fixed_lines)
+
+
+def get_yaml_error_context(content, line_num, column=None):
+    """Get context around a YAML error for better debugging."""
+    lines = content.split('\n')
+    if line_num < 1 or line_num > len(lines):
+        return "Could not extract error context"
+    
+    context_lines = []
+    start = max(0, line_num - 3)
+    end = min(len(lines), line_num + 2)
+    
+    for i in range(start, end):
+        line_indicator = ">>> " if i == line_num - 1 else "    "
+        context_lines.append(f"{line_indicator}{i + 1}: {lines[i]}")
+        
+        # Add column indicator for the error line
+        if i == line_num - 1 and column:
+            pointer = " " * (len(str(i + 1)) + 6 + column - 1) + "^"
+            context_lines.append(pointer)
+    
+    return '\n'.join(context_lines)
+
+
+def conf_edit(config_path, chunk_size, overlap, model_name=None):
+    """Edits the configuration file overlap and training flags.
+    The model's native audio.chunk_size from the YAML is preserved unchanged.
+    
+    Args:
+        config_path: Path to the config file
+        chunk_size: Unused – kept for API compatibility (native YAML value is used instead)
+        overlap: Overlap between chunks
+        model_name: Optional model name for re-downloading config on error
+    """
     full_config_path = os.path.join(CHECKPOINT_DIR, os.path.basename(config_path))
     if not os.path.exists(full_config_path):
         raise FileNotFoundError(f"Configuration file not found: {full_config_path}")
     
-    with open(full_config_path, 'r') as f:
-        data = yaml.load(f, Loader=yaml.SafeLoader)
+    # Create backup before modifying
+    backup_path = full_config_path + '.backup'
+    try:
+        shutil.copy2(full_config_path, backup_path)
+    except Exception:
+        pass
+    
+    try:
+        # Read and pre-process content
+        with open(full_config_path, 'r', encoding='utf-8') as f:
+            original_content = f.read()
+        
+        # Check if file is HTML (wrong URL was used)
+        is_valid, html_error = validate_yaml_content(original_content, full_config_path)
+        if not is_valid:
+            # Restore backup and raise error
+            if os.path.exists(backup_path):
+                shutil.copy2(backup_path, full_config_path)
+            raise ValueError(html_error)
+        
+        content = preprocess_yaml_content(original_content)
+        
+        # Write pre-processed content if changed
+        if content != original_content:
+            with open(full_config_path, 'w', encoding='utf-8') as f:
+                f.write(content)
+        
+        # Try to parse YAML
+        try:
+            with open(full_config_path, 'r', encoding='utf-8') as f:
+                data = yaml.load(f, Loader=yaml.SafeLoader)
+        except yaml.YAMLError as e:
+            # Extract error details
+            error_msg = str(e)
+            line_num = None
+            column = None
+            
+            if hasattr(e, 'problem_mark') and e.problem_mark:
+                line_num = e.problem_mark.line + 1
+                column = e.problem_mark.column + 1
+            
+            # Get context around error
+            context = ""
+            if line_num:
+                context = get_yaml_error_context(content, line_num, column)
+            
+            # Provide helpful error message
+            error_details = f"""
+YAML Parsing Error in config file: {full_config_path}
+
+Error: {error_msg}
 
-    if 'use_amp' not in data.keys():
-        data['training']['use_amp'] = True
+{"Error Context:" + chr(10) + context if context else ""}
 
-    data['audio']['chunk_size'] = chunk_size
-    data['inference']['num_overlap'] = overlap
-    if data['inference']['batch_size'] == 1:
-        data['inference']['batch_size'] = 2
+Possible causes:
+1. Unquoted string containing a colon (e.g., URLs like https://...)
+2. Unquoted Windows path with backslashes (e.g., C:\\path\\to\\file)
+3. Malformed YAML structure
+4. File corruption from previous processing
 
-    print(f"Using custom overlap and chunk_size: overlap={overlap}, chunk_size={chunk_size}")
-    with open(full_config_path, 'w') as f:
-        yaml.dump(data, f, default_flow_style=False, sort_keys=False, Dumper=yaml.Dumper)
+Suggested fixes:
+1. Delete the config file and let it re-download: {full_config_path}
+2. Manually edit the file to quote problematic values
+3. Check if the source config URL provides valid YAML
+"""
+            # Restore backup
+            if os.path.exists(backup_path):
+                try:
+                    shutil.copy2(backup_path, full_config_path)
+                except Exception:
+                    pass
+            
+            raise yaml.YAMLError(error_details) from e
+        
+        # Validate required sections exist
+        if not isinstance(data, dict):
+            raise ValueError(f"Config file does not contain a valid YAML dictionary: {full_config_path}")
+        
+        # Apply modifications safely
+        if 'use_amp' not in data:
+            if 'training' not in data:
+                data['training'] = {}
+            data['training']['use_amp'] = True
 
-def download_file(url):
-    """Downloads a file from a URL."""
+        # Do NOT overwrite audio.chunk_size — keep the model's native value from the YAML.
+        if 'audio' not in data:
+            data['audio'] = {}
+
+        if 'inference' not in data:
+            data['inference'] = {}
+        data['inference']['num_overlap'] = overlap
+        if data['inference'].get('batch_size', 1) == 1:
+            data['inference']['batch_size'] = 2
+        
+        # Write updated config
+        with open(full_config_path, 'w', encoding='utf-8') as f:
+            yaml.dump(data, f, default_flow_style=False, sort_keys=False, Dumper=yaml.Dumper)
+        
+        # Remove backup on success
+        if os.path.exists(backup_path):
+            try:
+                os.remove(backup_path)
+            except Exception:
+                pass
+                
+    except Exception as e:
+        # Restore backup on any error
+        if os.path.exists(backup_path):
+            try:
+                shutil.copy2(backup_path, full_config_path)
+                os.remove(backup_path)
+            except Exception:
+                pass
+        raise
+
+
+def redownload_config(model_name):
+    """Re-download a corrupted config file for a custom model.
+    
+    Args:
+        model_name: Name of the custom model
+        
+    Returns:
+        tuple: (success: bool, message: str)
+    """
+    custom_models = load_custom_models()
+    if model_name not in custom_models:
+        return False, f"Model '{model_name}' not found in custom models"
+    
+    config = custom_models[model_name]
+    config_url = config.get('config_url')
+    config_filename = config.get('config_filename')
+    
+    if not config_url or not config_filename:
+        return False, f"Config URL or filename not found for model '{model_name}'"
+    
+    config_path = os.path.join(CHECKPOINT_DIR, config_filename)
+    
+    # Auto-fix URL before re-downloading
+    config_url = fix_huggingface_url(config_url)
+    
+    # Delete existing config
+    if os.path.exists(config_path):
+        try:
+            os.remove(config_path)
+        except Exception as e:
+            return False, f"Could not delete config file: {e}"
+    
+    # Re-download with validation
+    try:
+        download_file(config_url, target_filename=config_filename, validate_yaml=True)
+        return True, f"Config file re-downloaded successfully: {config_filename}"
+    except Exception as e:
+        return False, f"Failed to re-download config: {e}"
+
+def download_file(url, path=None, target_filename=None, validate_yaml=True):
+    """Downloads a file from a URL with progress reporting.
+    
+    Args:
+        url: The URL to download from.
+        path: The directory to save the file to. Defaults to CHECKPOINT_DIR.
+        target_filename: Optional custom filename to save as. If None, uses filename from URL.
+        validate_yaml: If True and file is .yaml/.yml, validate it's not HTML
+    """
     import requests
+    
+    # Auto-fix Hugging Face URLs
+    url = fix_huggingface_url(url)
+    
     encoded_url = quote(url, safe=':/')
-    path = CHECKPOINT_DIR
+    if path is None:
+        path = CHECKPOINT_DIR
     os.makedirs(path, exist_ok=True)
-    filename = os.path.basename(encoded_url)
+    # Use custom target filename if provided, otherwise extract from URL
+    filename = target_filename if target_filename else os.path.basename(encoded_url)
     file_path = os.path.join(path, filename)
     if os.path.exists(file_path):
         print(f"File '{filename}' already exists at '{path}'.")
         return
     try:
-        response = requests.get(url)
+        response = requests.get(url, stream=True)
         if response.status_code == 200:
-            with open(file_path, 'wb') as f:
-                f.write(response.content)
-            print(f"File '{filename}' downloaded successfully")
+            # Get total file size for progress reporting
+            total_size = int(response.headers.get('content-length', 0))
+            
+            # For YAML files, download to memory first and validate
+            is_yaml_file = filename.lower().endswith(('.yaml', '.yml'))
+            
+            if is_yaml_file and validate_yaml:
+                content = response.content
+                is_valid, error_msg = validate_yaml_content(content, file_path)
+                if not is_valid:
+                    print(f"ERROR: Downloaded file is not valid YAML!")
+                    print(error_msg)
+                    raise ValueError(f"Downloaded file is HTML, not YAML. URL may be incorrect: {url}")
+                
+                with open(file_path, 'wb') as f:
+                    f.write(content)
+            else:
+                # Download with progress reporting
+                downloaded_size = 0
+                last_percent = -1
+                print(f"[SESA_DOWNLOAD]START:{filename}", flush=True)
+                
+                with open(file_path, 'wb') as f:
+                    for chunk in response.iter_content(chunk_size=8192):
+                        f.write(chunk)
+                        downloaded_size += len(chunk)
+                        
+                        # Report download progress
+                        if total_size > 0:
+                            percent = int((downloaded_size / total_size) * 100)
+                            if percent != last_percent:
+                                last_percent = percent
+                                # Format: [SESA_DOWNLOAD]filename:percent
+                                print(f"[SESA_DOWNLOAD]{filename}:{percent}", flush=True)
+                
+                print(f"[SESA_DOWNLOAD]END:{filename}", flush=True)
         else:
             print(f"Error downloading '{filename}': Status code {response.status_code}")
     except Exception as e:
         print(f"Error downloading file '{filename}' from '{url}': {e}")
+        raise
 
 # Model konfigurasyonlarını kategorize bir sözlükte tut
 MODEL_CONFIGS = {
     "Vocal Models": {
+        # === NEW MODELS (en üstte) ===
+        'bs_roformer_voc_hyperacev2 (by unwa)': {
+            'model_type': 'bs_roformer_custom',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_hyperacev2_voc.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_roformer_voc_hyperacev2.ckpt'),
+            'download_urls': [
+                ('https://huggingface.co/pcunwa/BS-Roformer-HyperACE/resolve/main/v2_voc/config.yaml', 'config_hyperacev2_voc.yaml'),
+                'https://huggingface.co/pcunwa/BS-Roformer-HyperACE/resolve/main/v2_voc/bs_roformer_voc_hyperacev2.ckpt'
+            ],
+            'custom_model_url': 'https://huggingface.co/pcunwa/BS-Roformer-HyperACE/resolve/main/v2_voc/bs_roformer.py',
+            'needs_conf_edit': True
+        },
+        'BS-Roformer-Resurrection (by unwa)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'BS-Roformer-Resurrection-Config.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'BS-Roformer-Resurrection.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/pcunwa/BS-Roformer-Resurrection/resolve/main/BS-Roformer-Resurrection-Config.yaml',
+                'https://huggingface.co/pcunwa/BS-Roformer-Resurrection/resolve/main/BS-Roformer-Resurrection.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'bs_roformer_revive3e (by unwa)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_revive.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_roformer_revive3e.ckpt'),
+            'download_urls': [
+                ('https://huggingface.co/pcunwa/BS-Roformer-Revive/resolve/main/config.yaml', 'config_revive.yaml'),
+                'https://huggingface.co/pcunwa/BS-Roformer-Revive/resolve/main/bs_roformer_revive3e.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'bs_roformer_revive2 (by unwa)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_revive.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_roformer_revive2.ckpt'),
+            'download_urls': [
+                ('https://huggingface.co/pcunwa/BS-Roformer-Revive/resolve/main/config.yaml', 'config_revive.yaml'),
+                'https://huggingface.co/pcunwa/BS-Roformer-Revive/resolve/main/bs_roformer_revive2.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'bs_roformer_revive (by unwa)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_revive.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_roformer_revive.ckpt'),
+            'download_urls': [
+                ('https://huggingface.co/pcunwa/BS-Roformer-Revive/resolve/main/config.yaml', 'config_revive.yaml'),
+                'https://huggingface.co/pcunwa/BS-Roformer-Revive/resolve/main/bs_roformer_revive.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'karaoke_bs_roformer_anvuew (by anvuew)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'karaoke_bs_roformer_anvuew.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'karaoke_bs_roformer_anvuew.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/anvuew/karaoke_bs_roformer/resolve/main/karaoke_bs_roformer_anvuew.yaml',
+                'https://huggingface.co/anvuew/karaoke_bs_roformer/resolve/main/karaoke_bs_roformer_anvuew.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        # === EXISTING MODELS ===
         'VOCALS-big_beta6X (by Unwa)': {
             'model_type': 'mel_band_roformer',
             'config_path': os.path.join(CHECKPOINT_DIR, 'big_beta6x.yaml'),
@@ -272,9 +814,204 @@ MODEL_CONFIGS = {
                 'https://huggingface.co/Aname-Tommy/MelBandRoformers/blob/main/FullnessVocalModel.ckpt'
             ],
             'needs_conf_edit': True
+        },
+        'voc_fv4 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'voc_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'voc_fv4.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_fv4.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'voc_fv5 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'voc_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'voc_fv5.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_fv5.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'voc_fv6 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'voc_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'voc_fv6.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_fv6.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'voc_fv7 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'v7.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'voc_fv7.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/v7.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_fv7.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'vocfv7beta1 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'voc_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'vocfv7beta1.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/experimental/vocfv7beta1.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'vocfv7beta2 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'voc_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'vocfv7beta2.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/experimental/vocfv7beta2.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'vocfv7beta3 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'voc_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'vocfv7beta3.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/vocals/voc_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/experimental/vocfv7beta3.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'MelBandRoformerSYHFTV3Epsilon (by SYH99999)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_vocals_mel_band_roformer_ft.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'MelBandRoformerSYHFTV3Epsilon.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/SYH99999/MelBandRoformerSYHFT/resolve/main/config_vocals_mel_band_roformer_ft.yaml',
+                'https://huggingface.co/SYH99999/MelBandRoformerSYHFTV3Epsilon/resolve/main/MelBandRoformerSYHFTV3Epsilon.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'MelBandRoformerBigSYHFTV1 (by SYH99999)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_big_syhft.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'MelBandRoformerBigSYHFTV1.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/SYH99999/MelBandRoformerBigSYHFTV1Fast/resolve/main/config.yaml',
+                'https://huggingface.co/SYH99999/MelBandRoformerBigSYHFTV1Fast/resolve/main/MelBandRoformerBigSYHFTV1.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'model_chorus_bs_roformer_ep_146 (by Sucial)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_chorus_male_female_bs_roformer.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'model_chorus_bs_roformer_ep_146_sdr_23.8613.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/Sucial/Chorus_Male_Female_BS_Roformer/resolve/main/config_chorus_male_female_bs_roformer.yaml',
+                'https://huggingface.co/Sucial/Chorus_Male_Female_BS_Roformer/resolve/main/model_chorus_bs_roformer_ep_146_sdr_23.8613.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'model_chorus_bs_roformer_ep_267 (by Sucial)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_chorus_male_female_bs_roformer.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'model_chorus_bs_roformer_ep_267_sdr_24.1275.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/Sucial/Chorus_Male_Female_BS_Roformer/resolve/main/config_chorus_male_female_bs_roformer.yaml',
+                'https://huggingface.co/Sucial/Chorus_Male_Female_BS_Roformer/resolve/main/model_chorus_bs_roformer_ep_267_sdr_24.1275.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'BS-Rofo-SW-Fixed (by jarredou)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'BS-Rofo-SW-Fixed.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'BS-Rofo-SW-Fixed.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/jarredou/BS-ROFO-SW-Fixed/resolve/main/BS-Rofo-SW-Fixed.yaml',
+                'https://huggingface.co/jarredou/BS-ROFO-SW-Fixed/resolve/main/BS-Rofo-SW-Fixed.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'BS_ResurrectioN (by Gabox)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'BS-Roformer-Resurrection-Inst-Config.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'BS_ResurrectioN.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/pcunwa/BS-Roformer-Resurrection/resolve/main/BS-Roformer-Resurrection-Inst-Config.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/experimental/BS_ResurrectioN.ckpt'
+            ],
+            'needs_conf_edit': True
         }
     },
     "Instrumental Models": {
+        # === NEW MODELS (en üstte) ===
+        'Neo_InstVFX (by natanworkspace)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_neo_inst.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'Neo_InstVFX.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/natanworkspace/melband_roformer/resolve/main/config_neo_inst.yaml',
+                'https://huggingface.co/natanworkspace/melband_roformer/resolve/main/Neo_InstVFX.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'BS-Roformer-Resurrection-Inst (by unwa)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'BS-Roformer-Resurrection-Inst-Config.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'BS-Roformer-Resurrection-Inst.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/pcunwa/BS-Roformer-Resurrection/resolve/main/BS-Roformer-Resurrection-Inst-Config.yaml',
+                'https://huggingface.co/pcunwa/BS-Roformer-Resurrection/resolve/main/BS-Roformer-Resurrection-Inst.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'bs_roformer_inst_hyperacev2 (by unwa)': {
+            'model_type': 'bs_roformer_custom',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_hyperacev2_inst.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_roformer_inst_hyperacev2.ckpt'),
+            'download_urls': [
+                ('https://huggingface.co/pcunwa/BS-Roformer-HyperACE/resolve/main/v2_inst/config.yaml', 'config_hyperacev2_inst.yaml'),
+                'https://huggingface.co/pcunwa/BS-Roformer-HyperACE/resolve/main/v2_inst/bs_roformer_inst_hyperacev2.ckpt'
+            ],
+            'custom_model_url': 'https://huggingface.co/pcunwa/BS-Roformer-HyperACE/resolve/main/v2_inst/bs_roformer.py',
+            'needs_conf_edit': True
+        },
+        'BS-Roformer-Large-Inst (by unwa)': {
+            'model_type': 'bs_roformer_custom',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_bs_large_inst.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_large_v2_inst.ckpt'),
+            'download_urls': [
+                ('https://huggingface.co/pcunwa/BS-Roformer-Large-Inst/resolve/main/config.yaml', 'config_bs_large_inst.yaml'),
+                'https://huggingface.co/pcunwa/BS-Roformer-Large-Inst/resolve/main/bs_large_v2_inst.ckpt'
+            ],
+            'custom_model_url': 'https://huggingface.co/pcunwa/BS-Roformer-Large-Inst/resolve/main/bs_roformer.py',
+            'needs_conf_edit': True
+        },
+        'bs_roformer_fno (by unwa)': {
+            'model_type': 'bs_roformer_custom',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'bsrofo_fno.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_roformer_fno.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/pcunwa/BS-Roformer-Inst-FNO/resolve/main/bsrofo_fno.yaml',
+                'https://huggingface.co/pcunwa/BS-Roformer-Inst-FNO/resolve/main/bs_roformer_fno.ckpt'
+            ],
+            'custom_model_url': 'https://huggingface.co/listra92/MyModels/resolve/main/misc/bs_roformer.py',
+            'needs_conf_edit': True
+        },
+        'Rifforge_final_sdr_14.24 (by meskvlla33)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_rifforge_full_mesk.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'rifforge_full_sdr_14.2436.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/meskvlla33/rifforge/resolve/main/config_rifforge_full_mesk.yaml',
+                'https://huggingface.co/meskvlla33/rifforge/resolve/main/rifforge_full_sdr_14.2436.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        # === EXISTING MODELS ===
         'Inst_GaboxFv8 (by Gabox)': {
             'model_type': 'mel_band_roformer',
             'config_path': os.path.join(CHECKPOINT_DIR, 'inst_gabox.yaml'),
@@ -504,6 +1241,86 @@ MODEL_CONFIGS = {
                 'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/Inst_GaboxV7.ckpt'
             ],
             'needs_conf_edit': True
+        },
+        'inst_Fv4 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'inst_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'inst_Fv4.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/inst_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/inst_Fv4.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'INSTV7N (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'inst_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'INSTV7N.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/inst_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/INSTV7N.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'inst_fv7b (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'inst_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'inst_fv7b.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/inst_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/experimental/inst_fv7b.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'inst_fv7z (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'inst_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'Inst_GaboxFv7z.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/inst_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/Inst_GaboxFv7z.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'Inst_GaboxFv9 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'inst_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'Inst_GaboxFv9.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/inst_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/Inst_GaboxFv9.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'inst_gaboxFlowersV10 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'v10.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'inst_gaboxFlowersV10.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/v10.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/inst_gaboxFlowersV10.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'Inst_FV8b (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'inst_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'Inst_FV8b.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/inst_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/experimental/Inst_FV8b.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'Inst_Fv8 (by Gabox)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'inst_gabox.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'Inst_Fv8.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/instrumental/inst_gabox.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/experimental/Inst_Fv8.ckpt'
+            ],
+            'needs_conf_edit': True
         }
     },
     "4-Stem Models": {
@@ -660,6 +1477,56 @@ MODEL_CONFIGS = {
                 'https://huggingface.co/anvuew/dereverb_mel_band_roformer/resolve/main/dereverb_mel_band_roformer_mono_anvuew_sdr_20.4029.ckpt'
             ],
             'needs_conf_edit': True
+        },
+        'dereverb-echo_128_4_4 (by Sucial)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_dereverb-echo_128_4_4_mel_band_roformer.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'dereverb-echo_128_4_4_mel_band_roformer_sdr_dry_12.4235.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/config_dereverb-echo_128_4_4_mel_band_roformer.yaml',
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/dereverb-echo_128_4_4_mel_band_roformer_sdr_dry_12.4235.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'dereverb_echo_mbr_v2 (by Sucial)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_dereverb_echo_mbr_v2.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'dereverb_echo_mbr_v2_sdr_dry_13.4843.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/config_dereverb_echo_mbr_v2.yaml',
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/dereverb_echo_mbr_v2_sdr_dry_13.4843.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'de_big_reverb_mbr_ep_362 (by Sucial)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_dereverb_echo_mbr_v2.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'de_big_reverb_mbr_ep_362.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/config_dereverb_echo_mbr_v2.yaml',
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/de_big_reverb_mbr_ep_362.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'de_super_big_reverb_mbr_ep_346 (by Sucial)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_dereverb_echo_mbr_v2.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'de_super_big_reverb_mbr_ep_346.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/config_dereverb_echo_mbr_v2.yaml',
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/de_super_big_reverb_mbr_ep_346.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'dereverb_room (by anvuew)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'dereverb_room_anvuew.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'dereverb_room_anvuew_sdr_13.7432.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/anvuew/dereverb_room/resolve/main/dereverb_room_anvuew.yaml',
+                'https://huggingface.co/anvuew/dereverb_room/resolve/main/dereverb_room_anvuew_sdr_13.7432.ckpt'
+            ],
+            'needs_conf_edit': True
         }
     },
     "Karaoke": {
@@ -673,13 +1540,43 @@ MODEL_CONFIGS = {
             ],
             'needs_conf_edit': True
         },
-        'KaraokeGabox': {
+        'KaraokeGabox (by Gabox)': {
             'model_type': 'mel_band_roformer',
-            'config_path': os.path.join(CHECKPOINT_DIR, 'config_mel_band_roformer_karaoke.yaml'),
-            'start_check_point': os.path.join(CHECKPOINT_DIR, 'KaraokeGabox.ckpt'),
+            'config_path': os.path.join(CHECKPOINT_DIR, 'karaokegabox_1750911344.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'Karaoke_GaboxV1.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/karaoke/karaokegabox_1750911344.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/melbandroformers/karaoke/Karaoke_GaboxV1.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'bs_karaoke_gabox_IS (by Gabox)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'karaoke_bs_roformer.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_karaoke_gabox_IS.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/bsroformers/karaoke_bs_roformer.yaml',
+                'https://huggingface.co/GaboxR67/MelBandRoformers/resolve/main/bsroformers/bs_karaoke_gabox_IS.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'bs_roformer_karaoke_frazer_becruily': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_karaoke_frazer_becruily.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_roformer_karaoke_frazer_becruily.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/becruily/bs-roformer-karaoke/resolve/main/config_karaoke_frazer_becruily.yaml',
+                'https://huggingface.co/becruily/bs-roformer-karaoke/resolve/main/bs_roformer_karaoke_frazer_becruily.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'mel_band_roformer_karaoke_becruily': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_karaoke_becruily.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'mel_band_roformer_karaoke_becruily.ckpt'),
             'download_urls': [
-                'https://github.com/deton24/Colab-for-new-MDX_UVR_models/releases/download/v1.0.0/config_mel_band_roformer_karaoke.yaml',
-                'https://huggingface.co/GaboxR67/MelBandRoformers/blob/main/melbandroformers/experimental/KaraokeGabox.ckpt'
+                'https://huggingface.co/becruily/mel-band-roformer-karaoke/resolve/main/config_karaoke_becruily.yaml',
+                'https://huggingface.co/becruily/mel-band-roformer-karaoke/resolve/main/mel_band_roformer_karaoke_becruily.ckpt'
             ],
             'needs_conf_edit': True
         }
@@ -764,23 +1661,220 @@ MODEL_CONFIGS = {
                 'https://huggingface.co/SYH99999/MelBandRoformerSYHFTB1/resolve/main/model3.ckpt'
             ],
             'needs_conf_edit': True
+        },
+        'bs_hyperace (by unwa)': {
+            'model_type': 'bs_roformer_custom',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_hyperace.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_hyperace.ckpt'),
+            'download_urls': [
+                ('https://huggingface.co/pcunwa/BS-Roformer-HyperACE/resolve/main/config.yaml', 'config_hyperace.yaml'),
+                'https://huggingface.co/pcunwa/BS-Roformer-HyperACE/resolve/main/bs_hyperace.ckpt'
+            ],
+            'custom_model_url': 'https://huggingface.co/pcunwa/BS-Roformer-HyperACE/resolve/main/bs_roformer.py',
+            'needs_conf_edit': True
+        },
+        'becruily_deux (by becruily)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_deux_becruily.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'becruily_deux.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/becruily/mel-band-roformer-deux/resolve/main/config_deux_becruily.yaml',
+                'https://huggingface.co/becruily/mel-band-roformer-deux/resolve/main/becruily_deux.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'becruily_guitar (by becruily)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_guitar_becruily.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'becruily_guitar.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/becruily/mel-band-roformer-guitar/resolve/main/config_guitar_becruily.yaml',
+                'https://huggingface.co/becruily/mel-band-roformer-guitar/resolve/main/becruily_guitar.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'aspiration_mel_band_roformer (by Sucial)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_aspiration_mel_band_roformer.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'aspiration_mel_band_roformer_sdr_18.9845.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/Sucial/Aspiration_Mel_Band_Roformer/resolve/main/config_aspiration_mel_band_roformer.yaml',
+                'https://huggingface.co/Sucial/Aspiration_Mel_Band_Roformer/resolve/main/aspiration_mel_band_roformer_sdr_18.9845.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'dereverb_echo_mbr_v2 (by Sucial)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_dereverb_echo_mbr_v2.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'dereverb_echo_mbr_v2_sdr_dry_13.4843.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/config_dereverb_echo_mbr_v2.yaml',
+                'https://huggingface.co/Sucial/Dereverb-Echo_Mel_Band_Roformer/resolve/main/dereverb_echo_mbr_v2_sdr_dry_13.4843.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'mdx23c_similarity (by ZFTurbo)': {
+            'model_type': 'mdx23c',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_mdx23c_similarity.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'model_mdx23c_ep_271_l1_freq_72.2383.ckpt'),
+            'download_urls': [
+                'https://github.com/ZFTurbo/Music-Source-Separation-Training/releases/download/v1.0.10/config_mdx23c_similarity.yaml',
+                'https://github.com/ZFTurbo/Music-Source-Separation-Training/releases/download/v1.0.10/model_mdx23c_ep_271_l1_freq_72.2383.ckpt'
+            ],
+            'needs_conf_edit': False
+        },
+        'mel_band_roformer_Lead_Rhythm_Guitar (by listra92)': {
+            'model_type': 'mel_band_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_mel_band_roformer_Lead_Rhythm_Guitar.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'model_mel_band_roformer_ep_72_sdr_3.2232.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/listra92/MyModels/resolve/main/misc/config_mel_band_roformer_Lead_Rhythm_Guitar.yaml',
+                'https://huggingface.co/listra92/MyModels/resolve/main/misc/model_mel_band_roformer_ep_72_sdr_3.2232.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'last_bs_roformer_4stem (by Amane)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_last_bs.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'last_bs_roformer.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/listra92/MyModels/resolve/main/misc/config.yaml',
+                'https://huggingface.co/listra92/MyModels/resolve/main/misc/last_bs_roformer.ckpt'
+            ],
+            'needs_conf_edit': True
+        },
+        'bs_roformer_4stems_ft (by SYH99999)': {
+            'model_type': 'bs_roformer',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_bs_4stems_ft.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'bs_roformer_4stems_ft.pth'),
+            'download_urls': [
+                'https://huggingface.co/SYH99999/bs_roformer_4stems_ft/resolve/main/config.yaml',
+                'https://huggingface.co/SYH99999/bs_roformer_4stems_ft/resolve/main/bs_roformer_4stems_ft.pth'
+            ],
+            'needs_conf_edit': True
+        },
+        'CINEMATIC-BandIt_v2_Eng (by kwatcharasupat)': {
+            'model_type': 'bandit_v2',
+            'config_path': os.path.join(CHECKPOINT_DIR, 'config_dnr_bandit_v2_mus64.yaml'),
+            'start_check_point': os.path.join(CHECKPOINT_DIR, 'checkpoint-eng_state_dict.ckpt'),
+            'download_urls': [
+                'https://huggingface.co/jarredou/banditv2_state_dicts_only/resolve/main/config_dnr_bandit_v2_mus64.yaml',
+                'https://huggingface.co/jarredou/banditv2_state_dicts_only/resolve/main/checkpoint-eng_state_dict.ckpt'
+            ],
+            'needs_conf_edit': True
         }
     }
 }
 
 def get_model_config(clean_model=None, chunk_size=None, overlap=None):
-    """Returns model type, config path, and checkpoint path for a given model name, downloading files if needed."""
+    """Returns model type, config path, and checkpoint path for a given model name, downloading files if needed.
+    
+    download_urls can contain:
+        - Simple strings: 'url' - downloads with filename from URL
+        - Tuples: ('url', 'target_filename') - downloads with custom filename
+    
+    Also handles custom models loaded from custom_models.json
+    """
     if clean_model is None:
-        return {model_name for category in MODEL_CONFIGS.values() for model_name in category.keys()}
+        all_models = {model_name for category in MODEL_CONFIGS.values() for model_name in category.keys()}
+        # Add custom models
+        custom_models = load_custom_models()
+        all_models.update(custom_models.keys())
+        return all_models
     
+    # First check built-in models
     for category in MODEL_CONFIGS.values():
         if clean_model in category:
             config = category[clean_model]
-            for url in config['download_urls']:
-                download_file(url)
+            for url_entry in config['download_urls']:
+                # Handle both simple URL strings and (url, target_filename) tuples
+                if isinstance(url_entry, tuple):
+                    url, target_filename = url_entry
+                    download_file(url, target_filename=target_filename)
+                else:
+                    download_file(url_entry)
+            if config.get('custom_model_url'):
+                custom_path = os.path.join(BASE_DIR, 'models', 'bs_roformer', 'bs_roformer_custom')
+                os.makedirs(custom_path, exist_ok=True)
+                # Create __init__.py for Python import support
+                init_file = os.path.join(custom_path, '__init__.py')
+                if not os.path.exists(init_file):
+                    with open(init_file, 'w') as f:
+                        f.write('# Auto-generated for custom BSRoformer models\n')
+                download_file(config['custom_model_url'], path=custom_path)
             if config['needs_conf_edit'] and chunk_size is not None and overlap is not None:
                 conf_edit(config['config_path'], chunk_size, overlap)
             return config['model_type'], config['config_path'], config['start_check_point']
+    
+    # Then check custom models
+    custom_models = load_custom_models()
+    if clean_model in custom_models:
+        config = custom_models[clean_model]
+        checkpoint_path = os.path.join(CHECKPOINT_DIR, config['checkpoint_filename'])
+        config_path = os.path.join(CHECKPOINT_DIR, config['config_filename'])
+        
+        # Download checkpoint
+        download_file(config['checkpoint_url'], target_filename=config['checkpoint_filename'])
+        # Download config with custom filename
+        download_file(config['config_url'], target_filename=config['config_filename'])
+        
+        # Handle custom model URL if present
+        if config.get('custom_model_url'):
+            custom_path = os.path.join(BASE_DIR, 'models', 'bs_roformer', 'bs_roformer_custom')
+            os.makedirs(custom_path, exist_ok=True)
+            init_file = os.path.join(custom_path, '__init__.py')
+            if not os.path.exists(init_file):
+                with open(init_file, 'w') as f:
+                    f.write('# Auto-generated for custom BSRoformer models\n')
+            download_file(config['custom_model_url'], path=custom_path)
+        
+        # Apply config edits if needed
+        if config.get('needs_conf_edit', True) and chunk_size is not None and overlap is not None:
+            conf_edit(config_path, chunk_size, overlap, model_name=clean_model)
+        
+        return config['model_type'], config_path, checkpoint_path
+    
     return "", "", ""
 
-get_model_config.keys = lambda: {model_name for category in MODEL_CONFIGS.values() for model_name in category.keys()}
+def get_all_model_configs_with_custom():
+    """Returns MODEL_CONFIGS with Custom Models category added dynamically."""
+    all_configs = dict(MODEL_CONFIGS)
+    custom_models = load_custom_models()
+    if custom_models:
+        all_configs["Custom Models"] = {
+            name: {
+                'model_type': cfg['model_type'],
+                'config_path': os.path.join(CHECKPOINT_DIR, cfg['config_filename']),
+                'start_check_point': os.path.join(CHECKPOINT_DIR, cfg['checkpoint_filename']),
+                'download_urls': [cfg['checkpoint_url'], cfg['config_url']],
+                'custom_model_url': cfg.get('custom_model_url'),
+                'needs_conf_edit': cfg.get('needs_conf_edit', True)
+            }
+            for name, cfg in custom_models.items()
+        }
+    return all_configs
+
+get_model_config.keys = lambda: {model_name for category in MODEL_CONFIGS.values() for model_name in category.keys()}.union(load_custom_models().keys())
+
+
+def get_model_chunk_size(model_name):
+    """Read the native chunk_size from a model's local YAML config if already downloaded.
+    
+    Returns the int chunk_size on success, or None if the config is unavailable or
+    doesn't contain an audio.chunk_size entry.
+    """
+    all_models = get_all_model_configs_with_custom()
+    for category in all_models.values():
+        if model_name in category:
+            config_path = category[model_name].get('config_path', '')
+            if config_path and os.path.exists(config_path):
+                try:
+                    with open(config_path, 'r', encoding='utf-8') as f:
+                        data = yaml.safe_load(f)
+                    if isinstance(data, dict):
+                        chunk_size = data.get('audio', {}).get('chunk_size')
+                        if chunk_size:
+                            return int(chunk_size)
+                except Exception:
+                    pass
+    return None

models/bs_roformer/attend_sage.py

@@ -0,0 +1,145 @@
+from functools import wraps
+from packaging import version
+from collections import namedtuple
+
+import os
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+
+from einops import rearrange, reduce
+
+def _print_once(msg):
+    printed = False
+    @wraps(print)
+    def inner():
+        nonlocal printed
+        if not printed:
+            print(msg)
+            printed = True
+    return inner
+
+try:
+    from sageattention import sageattn
+    _has_sage_attention = True
+    # _print_sage_found = _print_once("SageAttention found. Will be used when flash=True.")
+    # _print_sage_found()
+except ImportError:
+    _has_sage_attention = False
+    _print_sage_not_found = _print_once("SageAttention not found. Will fall back to PyTorch SDPA (if available) or manual einsum.")
+    _print_sage_not_found()
+
+# helpers
+def exists(val):
+    return val is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+# main class
+class Attend(nn.Module):
+    def __init__(
+        self,
+        dropout = 0.,
+        flash = False, # If True, attempts to use SageAttention or PyTorch SDPA
+        scale = None
+    ):
+        super().__init__()
+        self.scale = scale # Store the scale if needed for einsum path
+        self.dropout = dropout # Store dropout if needed for einsum/SDPA path
+
+        # Determine which attention mechanism to *try* first
+        self.use_sage = flash and _has_sage_attention
+        self.use_pytorch_sdpa = False
+        self._sdpa_checked = False # Flag to check PyTorch version only once
+
+        if flash and not self.use_sage:
+            # Only consider PyTorch SDPA if Sage isn't available/chosen
+            if not self._sdpa_checked:
+                if version.parse(torch.__version__) >= version.parse('2.0.0'):
+                    self.use_pytorch_sdpa = True
+                    _print_sdpa_used = _print_once("Using PyTorch SDPA backend (FlashAttention-2, Memory-Efficient, or Math).")
+                    _print_sdpa_used()
+                else:
+                     _print_fallback_einsum = _print_once("Flash attention requested but Pytorch < 2.0 and SageAttention not found. Falling back to einsum.")
+                     _print_fallback_einsum()
+                self._sdpa_checked = True
+
+        # Dropout layer for manual einsum implementation ONLY
+        # SDPA and SageAttention handle dropout differently (or not at all in Sage's base API)
+        self.attn_dropout = nn.Dropout(dropout)
+
+    def forward(self, q, k, v):
+        """
+        einstein notation
+        b - batch
+        h - heads
+        n, i, j - sequence length (base sequence length, source, target)
+        d - feature dimension
+
+        Input tensors q, k, v expected in shape: (batch, heads, seq_len, dim_head) -> HND layout
+        """
+        q_len, k_len, device = q.shape[-2], k.shape[-2], q.device
+
+        # --- Priority 1: SageAttention ---
+        if self.use_sage:
+            # Assumes q, k, v are FP16/BF16 (handled by autocast upstream)
+            # Assumes scale is handled internally by sageattn
+            # Assumes dropout is NOT handled by sageattn kernel
+            # is_causal=False based on how Attend is called in mel_band_roformer
+            out = sageattn(q, k, v, tensor_layout='HND', is_causal=False)
+            return out
+            try:
+                return out
+                # print("Attempting SageAttention") # Optional: for debugging
+                out = sageattn(q, k, v, tensor_layout='HND', is_causal=False)
+                return out
+            except Exception as e:
+                print(f"SageAttention failed with error: {e}. Falling back.")
+                self.use_sage = False # Don't try Sage again if it failed once
+                # Decide fallback: Check if PyTorch SDPA is an option
+                if not self._sdpa_checked:
+                    if version.parse(torch.__version__) >= version.parse('2.0.0'):
+                        self.use_pytorch_sdpa = True
+                        _print_sdpa_fallback = _print_once("Falling back to PyTorch SDPA.")
+                        _print_sdpa_fallback()
+                    else:
+                        _print_einsum_fallback = _print_once("Falling back to einsum.")
+                        _print_einsum_fallback()
+                    self._sdpa_checked = True
+
+
+        # --- Priority 2: PyTorch SDPA ---
+        if self.use_pytorch_sdpa:
+             # Use PyTorch's Scaled Dot Product Attention (SDPA)
+             # It handles scaling and dropout internally.
+            try:
+                # print("Attempting PyTorch SDPA") # Optional: for debugging
+                # Let PyTorch choose the best backend (Flash V2, Mem Efficient, Math)
+                with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True):
+                    out = F.scaled_dot_product_attention(
+                        q, k, v,
+                        attn_mask=None, # Assuming no explicit mask needed here
+                        dropout_p = self.dropout if self.training else 0.,
+                        is_causal=False # Assuming not needed based on usage context
+                    )
+                return out
+            except Exception as e:
+                 print(f"PyTorch SDPA failed with error: {e}. Falling back to einsum.")
+                 self.use_pytorch_sdpa = False # Fallback to einsum on error
+
+
+        # Calculate scale
+        scale = default(self.scale, q.shape[-1] ** -0.5)
+
+        # similarity
+        sim = einsum(f"b h i d, b h j d -> b h i j", q, k) * scale
+
+        # attention
+        attn = sim.softmax(dim=-1)
+        attn = self.attn_dropout(attn) # Apply dropout ONLY in einsum path
+
+        # aggregate values
+        out = einsum(f"b h i j, b h j d -> b h i d", attn, v)
+
+        return out

models/bs_roformer/bs_roformer_experimental.py

@@ -0,0 +1,686 @@
+from functools import partial
+
+import torch
+from torch import nn, einsum, Tensor
+from torch.nn import Module, ModuleList
+import torch.nn.functional as F
+
+from models.bs_roformer.attend import Attend
+from torch.utils.checkpoint import checkpoint
+
+from beartype.typing import Tuple, Optional, List, Callable
+from beartype import beartype
+
+from rotary_embedding_torch import RotaryEmbedding
+
+from einops import rearrange, pack, unpack
+from einops.layers.torch import Rearrange
+
+from hyper_connections import get_init_and_expand_reduce_stream_functions
+
+# helper functions
+
+def exists(val):
+    return val is not None
+
+
+def default(v, d):
+    return v if exists(v) else d
+
+
+def pack_one(t, pattern):
+    return pack([t], pattern)
+
+
+def unpack_one(t, ps, pattern):
+    return unpack(t, ps, pattern)[0]
+
+
+# norm
+
+def l2norm(t):
+    return F.normalize(t, dim = -1, p = 2)
+
+
+class RMSNorm(Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.scale = dim ** 0.5
+        self.gamma = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        return F.normalize(x, dim=-1) * self.scale * self.gamma
+
+
+# attention
+
+class FeedForward(Module):
+    def __init__(
+            self,
+            dim,
+            mult=4,
+            dropout=0.
+    ):
+        super().__init__()
+        dim_inner = int(dim * mult)
+        self.net = nn.Sequential(
+            RMSNorm(dim),
+            nn.Linear(dim, dim_inner),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_inner, dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class Attention(Module):
+    def __init__(
+            self,
+            dim,
+            heads=8,
+            dim_head=64,
+            dropout=0.,
+            rotary_embed=None,
+            flash=True,
+            learned_value_residual_mix=False,
+    ):
+        super().__init__()
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+        dim_inner = heads * dim_head
+
+        self.rotary_embed = rotary_embed
+
+        self.attend = Attend(flash=flash, dropout=dropout)
+
+        self.norm = RMSNorm(dim)
+        self.to_qkv = nn.Linear(dim, dim_inner * 3, bias=False)
+
+        self.to_value_residual_mix = nn.Linear(dim, heads) if learned_value_residual_mix else None
+
+        self.to_gates = nn.Linear(dim, heads)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(dim_inner, dim, bias=False),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, value_residual=None):
+        x = self.norm(x)
+
+        q, k, v = rearrange(self.to_qkv(x), 'b n (qkv h d) -> qkv b h n d', qkv=3, h=self.heads)
+
+        orig_v = v
+
+        if exists(self.to_value_residual_mix):
+            mix = self.to_value_residual_mix(x)
+            mix = rearrange(mix, 'b n h -> b h n 1').sigmoid()
+
+            assert exists(value_residual)
+            v = v.lerp(value_residual, mix)
+
+        if exists(self.rotary_embed):
+            q = self.rotary_embed.rotate_queries_or_keys(q)
+            k = self.rotary_embed.rotate_queries_or_keys(k)
+
+        out = self.attend(q, k, v)
+
+        gates = self.to_gates(x)
+        out = out * rearrange(gates, 'b n h -> b h n 1').sigmoid()
+
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out), orig_v
+
+
+class LinearAttention(Module):
+    """
+    this flavor of linear attention proposed in https://arxiv.org/abs/2106.09681 by El-Nouby et al.
+    """
+
+    @beartype
+    def __init__(
+            self,
+            *,
+            dim,
+            dim_head=32,
+            heads=8,
+            scale=8,
+            flash=False,
+            dropout=0.
+    ):
+        super().__init__()
+        dim_inner = dim_head * heads
+        self.norm = RMSNorm(dim)
+
+        self.to_qkv = nn.Sequential(
+            nn.Linear(dim, dim_inner * 3, bias=False),
+            Rearrange('b n (qkv h d) -> qkv b h d n', qkv=3, h=heads)
+        )
+
+        self.temperature = nn.Parameter(torch.ones(heads, 1, 1))
+
+        self.attend = Attend(
+            scale=scale,
+            dropout=dropout,
+            flash=flash
+        )
+
+        self.to_out = nn.Sequential(
+            Rearrange('b h d n -> b n (h d)'),
+            nn.Linear(dim_inner, dim, bias=False)
+        )
+
+    def forward(
+            self,
+            x
+    ):
+        x = self.norm(x)
+
+        q, k, v = self.to_qkv(x)
+
+        q, k = map(l2norm, (q, k))
+        q = q * self.temperature.exp()
+
+        out = self.attend(q, k, v)
+
+        return self.to_out(out)
+
+
+class Transformer(Module):
+    def __init__(
+            self,
+            *,
+            dim,
+            depth,
+            dim_head=64,
+            heads=8,
+            attn_dropout=0.,
+            ff_dropout=0.,
+            ff_mult=4,
+            norm_output=True,
+            rotary_embed=None,
+            flash_attn=True,
+            linear_attn=False,
+            add_value_residual=False,
+            num_residual_streams=1,
+    ):
+        super().__init__()
+        self.layers = ModuleList([])
+
+        init_hyper_conn, *_ = get_init_and_expand_reduce_stream_functions(num_residual_streams, disable=num_residual_streams == 1)
+
+        for _ in range(depth):
+            if linear_attn:
+                attn = LinearAttention(dim=dim, dim_head=dim_head, heads=heads, dropout=attn_dropout, flash=flash_attn)
+            else:
+                if num_residual_streams != 1:
+                    attn = init_hyper_conn(dim=dim, branch=Attention(dim=dim, dim_head=dim_head, heads=heads, dropout=attn_dropout,
+                                 rotary_embed=rotary_embed, flash=flash_attn, learned_value_residual_mix=add_value_residual))
+                else:
+                    attn = Attention(
+                        dim=dim, dim_head=dim_head, heads=heads, dropout=attn_dropout,
+                        rotary_embed=rotary_embed, flash=flash_attn, learned_value_residual_mix=add_value_residual
+                    )
+            if num_residual_streams != 1:
+                ff = init_hyper_conn(dim=dim, branch=FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout))
+            else:
+                ff = FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout)
+
+            self.layers.append(ModuleList([
+                attn,
+                ff
+            ]))
+
+        self.norm = RMSNorm(dim) if norm_output else nn.Identity()
+
+    def forward(self, x, value_residual=None):
+
+        first_values = None
+        if value_residual is not None:
+            for attn, ff in self.layers:
+                x, next_values = attn(x, value_residual=value_residual)
+                first_values = default(first_values, next_values)
+                x = ff(x)
+        else:
+            # Compatibility with old weights
+            for attn, ff in self.layers:
+                attn_out, next_values = attn(x, value_residual=None)
+                first_values = default(first_values, next_values)
+                x = attn_out + x
+                x = ff(x) + x
+
+        return self.norm(x), first_values
+
+
+# bandsplit module
+
+class BandSplit(Module):
+    @beartype
+    def __init__(
+            self,
+            dim,
+            dim_inputs: Tuple[int, ...]
+    ):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_features = ModuleList([])
+
+        for dim_in in dim_inputs:
+            net = nn.Sequential(
+                RMSNorm(dim_in),
+                nn.Linear(dim_in, dim)
+            )
+
+            self.to_features.append(net)
+
+    def forward(self, x):
+        x = x.split(self.dim_inputs, dim=-1)
+
+        outs = []
+        for split_input, to_feature in zip(x, self.to_features):
+            split_output = to_feature(split_input)
+            outs.append(split_output)
+
+        return torch.stack(outs, dim=-2)
+
+
+def MLP(
+        dim_in,
+        dim_out,
+        dim_hidden=None,
+        depth=1,
+        activation=nn.Tanh
+):
+    dim_hidden = default(dim_hidden, dim_in)
+
+    net = []
+    dims = (dim_in, *((dim_hidden,) * (depth - 1)), dim_out)
+
+    for ind, (layer_dim_in, layer_dim_out) in enumerate(zip(dims[:-1], dims[1:])):
+        is_last = ind == (len(dims) - 2)
+
+        net.append(nn.Linear(layer_dim_in, layer_dim_out))
+
+        if is_last:
+            continue
+
+        net.append(activation())
+
+    return nn.Sequential(*net)
+
+
+class MaskEstimator(Module):
+    @beartype
+    def __init__(
+            self,
+            dim,
+            dim_inputs: Tuple[int, ...],
+            depth,
+            mlp_expansion_factor=4
+    ):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_freqs = ModuleList([])
+        dim_hidden = dim * mlp_expansion_factor
+
+        for dim_in in dim_inputs:
+            net = []
+
+            mlp = nn.Sequential(
+                MLP(dim, dim_in * 2, dim_hidden=dim_hidden, depth=depth),
+                nn.GLU(dim=-1)
+            )
+
+            self.to_freqs.append(mlp)
+
+    def forward(self, x):
+        x = x.unbind(dim=-2)
+
+        outs = []
+
+        for band_features, mlp in zip(x, self.to_freqs):
+            freq_out = mlp(band_features)
+            outs.append(freq_out)
+
+        return torch.cat(outs, dim=-1)
+
+
+# main class
+
+DEFAULT_FREQS_PER_BANDS = (
+    2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+    2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
+    2, 2, 2, 2,
+    4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4,
+    12, 12, 12, 12, 12, 12, 12, 12,
+    24, 24, 24, 24, 24, 24, 24, 24,
+    48, 48, 48, 48, 48, 48, 48, 48,
+    128, 129,
+)
+
+
+class BSRoformer(Module):
+
+    @beartype
+    def __init__(
+            self,
+            dim,
+            *,
+            depth,
+            stereo=False,
+            num_stems=1,
+            time_transformer_depth=2,
+            freq_transformer_depth=2,
+            linear_transformer_depth=0,
+            freqs_per_bands: Tuple[int, ...] = DEFAULT_FREQS_PER_BANDS,
+            # in the paper, they divide into ~60 bands, test with 1 for starters
+            dim_head=64,
+            heads=8,
+            attn_dropout=0.,
+            ff_dropout=0.,
+            flash_attn=True,
+            dim_freqs_in=1025,
+            stft_n_fft=2048,
+            stft_hop_length=512,
+            # 10ms at 44100Hz, from sections 4.1, 4.4 in the paper - @faroit recommends // 2 or // 4 for better reconstruction
+            stft_win_length=2048,
+            stft_normalized=False,
+            stft_window_fn: Optional[Callable] = None,
+            mask_estimator_depth=2,
+            multi_stft_resolution_loss_weight=1.,
+            multi_stft_resolutions_window_sizes: Tuple[int, ...] = (4096, 2048, 1024, 512, 256),
+            multi_stft_hop_size=147,
+            multi_stft_normalized=False,
+            multi_stft_window_fn: Callable = torch.hann_window,
+            mlp_expansion_factor=4,
+            use_torch_checkpoint=False,
+            skip_connection=False,
+            use_value_residual_learning=False,
+            num_residual_streams=1,  # set to 1. to disable hyper connections (Default in original is 4)
+    ):
+        super().__init__()
+
+        self.stereo = stereo
+        self.audio_channels = 2 if stereo else 1
+        self.num_stems = num_stems
+        self.use_torch_checkpoint = use_torch_checkpoint
+        self.skip_connection = skip_connection
+        self.num_residual_streams = num_residual_streams
+
+        _, self.expand_stream, self.reduce_stream = get_init_and_expand_reduce_stream_functions(num_residual_streams, disable=num_residual_streams == 1)
+
+        self.layers = ModuleList([])
+
+        transformer_kwargs = dict(
+            dim=dim,
+            heads=heads,
+            dim_head=dim_head,
+            attn_dropout=attn_dropout,
+            ff_dropout=ff_dropout,
+            flash_attn=flash_attn,
+            norm_output=False,
+            num_residual_streams=num_residual_streams,
+        )
+
+        time_rotary_embed = RotaryEmbedding(dim=dim_head)
+        freq_rotary_embed = RotaryEmbedding(dim=dim_head)
+
+        for layer_index in range(depth):
+            if use_value_residual_learning:
+                is_first = layer_index == 0
+            else:
+                is_first = True
+
+            tran_modules = []
+            if linear_transformer_depth > 0:
+                tran_modules.append(Transformer(depth=linear_transformer_depth, linear_attn=True, **transformer_kwargs))
+            tran_modules.append(
+                Transformer(depth=time_transformer_depth, rotary_embed=time_rotary_embed, add_value_residual=not is_first, **transformer_kwargs)
+            )
+            tran_modules.append(
+                Transformer(depth=freq_transformer_depth, rotary_embed=freq_rotary_embed, add_value_residual=not is_first, **transformer_kwargs)
+            )
+            self.layers.append(nn.ModuleList(tran_modules))
+
+        self.final_norm = RMSNorm(dim)
+
+        self.stft_kwargs = dict(
+            n_fft=stft_n_fft,
+            hop_length=stft_hop_length,
+            win_length=stft_win_length,
+            normalized=stft_normalized
+        )
+
+        self.stft_window_fn = partial(default(stft_window_fn, torch.hann_window), stft_win_length)
+
+        freqs = torch.stft(torch.randn(1, 4096), **self.stft_kwargs, window=torch.ones(stft_win_length), return_complex=True).shape[1]
+
+        assert len(freqs_per_bands) > 1
+        assert sum(
+            freqs_per_bands) == freqs, f'the number of freqs in the bands must equal {freqs} based on the STFT settings, but got {sum(freqs_per_bands)}'
+
+        freqs_per_bands_with_complex = tuple(2 * f * self.audio_channels for f in freqs_per_bands)
+
+        self.band_split = BandSplit(
+            dim=dim,
+            dim_inputs=freqs_per_bands_with_complex
+        )
+
+        self.mask_estimators = nn.ModuleList([])
+
+        for _ in range(num_stems):
+            mask_estimator = MaskEstimator(
+                dim=dim,
+                dim_inputs=freqs_per_bands_with_complex,
+                depth=mask_estimator_depth,
+                mlp_expansion_factor=mlp_expansion_factor,
+            )
+
+            self.mask_estimators.append(mask_estimator)
+
+        # for the multi-resolution stft loss
+
+        self.multi_stft_resolution_loss_weight = multi_stft_resolution_loss_weight
+        self.multi_stft_resolutions_window_sizes = multi_stft_resolutions_window_sizes
+        self.multi_stft_n_fft = stft_n_fft
+        self.multi_stft_window_fn = multi_stft_window_fn
+
+        self.multi_stft_kwargs = dict(
+            hop_length=multi_stft_hop_size,
+            normalized=multi_stft_normalized
+        )
+
+    def forward(
+            self,
+            raw_audio,
+            target=None,
+            return_loss_breakdown=False
+    ):
+        """
+        einops
+
+        b - batch
+        f - freq
+        t - time
+        s - audio channel (1 for mono, 2 for stereo)
+        n - number of 'stems'
+        c - complex (2)
+        d - feature dimension
+        """
+
+        device = raw_audio.device
+
+        # defining whether model is loaded on MPS (MacOS GPU accelerator)
+        x_is_mps = True if device.type == "mps" else False
+
+        if raw_audio.ndim == 2:
+            raw_audio = rearrange(raw_audio, 'b t -> b 1 t')
+
+        channels = raw_audio.shape[1]
+        assert (not self.stereo and channels == 1) or (self.stereo and channels == 2), 'stereo needs to be set to True if passing in audio signal that is stereo (channel dimension of 2). also need to be False if mono (channel dimension of 1)'
+
+        # to stft
+
+        raw_audio, batch_audio_channel_packed_shape = pack_one(raw_audio, '* t')
+
+        stft_window = self.stft_window_fn(device=device)
+
+        # RuntimeError: FFT operations are only supported on MacOS 14+
+        # Since it's tedious to define whether we're on correct MacOS version - simple try-catch is used
+        try:
+            stft_repr = torch.stft(raw_audio, **self.stft_kwargs, window=stft_window, return_complex=True)
+        except:
+            stft_repr = torch.stft(raw_audio.cpu() if x_is_mps else raw_audio, **self.stft_kwargs,
+                                   window=stft_window.cpu() if x_is_mps else stft_window, return_complex=True).to(
+                device)
+        stft_repr = torch.view_as_real(stft_repr)
+
+        stft_repr = unpack_one(stft_repr, batch_audio_channel_packed_shape, '* f t c')
+
+        # merge stereo / mono into the frequency, with frequency leading dimension, for band splitting
+        stft_repr = rearrange(stft_repr,'b s f t c -> b (f s) t c')
+
+        x = rearrange(stft_repr, 'b f t c -> b t (f c)')
+
+        if self.use_torch_checkpoint:
+            x = checkpoint(self.band_split, x, use_reentrant=False)
+        else:
+            x = self.band_split(x)
+
+        # value residuals
+
+        time_v_residual = None
+        freq_v_residual = None
+
+        # maybe expand residual streams
+        if self.num_residual_streams != 1:
+            x = self.expand_stream(x)
+
+        # axial / hierarchical attention
+
+        store = [None] * len(self.layers)
+        for i, transformer_block in enumerate(self.layers):
+
+            if len(transformer_block) == 3:
+                linear_transformer, time_transformer, freq_transformer = transformer_block
+
+                x, ft_ps = pack([x], 'b * d')
+                if self.use_torch_checkpoint:
+                    x = checkpoint(linear_transformer, x, use_reentrant=False)
+                else:
+                    x = linear_transformer(x)
+                x, = unpack(x, ft_ps, 'b * d')
+            else:
+                time_transformer, freq_transformer = transformer_block
+
+            if self.skip_connection:
+                # Sum all previous
+                for j in range(i):
+                    x = x + store[j]
+
+            x = rearrange(x, 'b t f d -> b f t d')
+            x, ps = pack([x], '* t d')
+
+            if self.use_torch_checkpoint:
+                x, next_time_v_residual = checkpoint(time_transformer, x, time_v_residual, use_reentrant=False)
+            else:
+                x, next_time_v_residual = time_transformer(x, value_residual=time_v_residual)
+            time_v_residual = default(time_v_residual, next_time_v_residual)
+
+            x, = unpack(x, ps, '* t d')
+            x = rearrange(x, 'b f t d -> b t f d')
+            x, ps = pack([x], '* f d')
+
+            if self.use_torch_checkpoint:
+                x, next_freq_v_residual = checkpoint(freq_transformer, x, freq_v_residual, use_reentrant=False)
+            else:
+                x, next_freq_v_residual = freq_transformer(x, value_residual=freq_v_residual)
+            freq_v_residual = default(freq_v_residual, next_freq_v_residual)
+
+            x, = unpack(x, ps, '* f d')
+
+            if self.skip_connection:
+                store[i] = x
+
+        # maybe reduce residual streams
+        if self.num_residual_streams != 1:
+            x = self.reduce_stream(x)
+
+        x = self.final_norm(x)
+
+        num_stems = len(self.mask_estimators)
+
+        if self.use_torch_checkpoint:
+            mask = torch.stack([checkpoint(fn, x, use_reentrant=False) for fn in self.mask_estimators], dim=1)
+        else:
+            mask = torch.stack([fn(x) for fn in self.mask_estimators], dim=1)
+        mask = rearrange(mask, 'b n t (f c) -> b n f t c', c=2)
+
+        # modulate frequency representation
+
+        stft_repr = rearrange(stft_repr, 'b f t c -> b 1 f t c')
+
+        # complex number multiplication
+
+        stft_repr = torch.view_as_complex(stft_repr)
+        mask = torch.view_as_complex(mask)
+
+        stft_repr = stft_repr * mask
+
+        # istft
+
+        stft_repr = rearrange(stft_repr, 'b n (f s) t -> (b n s) f t', s=self.audio_channels)
+
+        # same as torch.stft() fix for MacOS MPS above
+        try:
+            recon_audio = torch.istft(stft_repr, **self.stft_kwargs, window=stft_window, return_complex=False, length=raw_audio.shape[-1])
+        except:
+            recon_audio = torch.istft(stft_repr.cpu() if x_is_mps else stft_repr, **self.stft_kwargs, window=stft_window.cpu() if x_is_mps else stft_window, return_complex=False, length=raw_audio.shape[-1]).to(device)
+
+        recon_audio = rearrange(recon_audio, '(b n s) t -> b n s t', s=self.audio_channels, n=num_stems)
+
+        if num_stems == 1:
+            recon_audio = rearrange(recon_audio, 'b 1 s t -> b s t')
+
+        # if a target is passed in, calculate loss for learning
+
+        if not exists(target):
+            return recon_audio
+
+        if self.num_stems > 1:
+            assert target.ndim == 4 and target.shape[1] == self.num_stems
+
+        if target.ndim == 2:
+            target = rearrange(target, '... t -> ... 1 t')
+
+        target = target[..., :recon_audio.shape[-1]]  # protect against lost length on istft
+
+        loss = F.l1_loss(recon_audio, target)
+
+        multi_stft_resolution_loss = 0.
+
+        for window_size in self.multi_stft_resolutions_window_sizes:
+            res_stft_kwargs = dict(
+                n_fft=max(window_size, self.multi_stft_n_fft),  # not sure what n_fft is across multi resolution stft
+                win_length=window_size,
+                return_complex=True,
+                window=self.multi_stft_window_fn(window_size, device=device),
+                **self.multi_stft_kwargs,
+            )
+
+            recon_Y = torch.stft(rearrange(recon_audio, '... s t -> (... s) t'), **res_stft_kwargs)
+            target_Y = torch.stft(rearrange(target, '... s t -> (... s) t'), **res_stft_kwargs)
+
+            multi_stft_resolution_loss = multi_stft_resolution_loss + F.l1_loss(recon_Y, target_Y)
+
+        weighted_multi_resolution_loss = multi_stft_resolution_loss * self.multi_stft_resolution_loss_weight
+
+        total_loss = loss + weighted_multi_resolution_loss
+
+        if not return_loss_breakdown:
+            return total_loss
+
+        return total_loss, (loss, multi_stft_resolution_loss)

models/bs_roformer/mel_band_roformer_experimental.py

@@ -0,0 +1,733 @@
+from functools import partial
+
+import torch
+from torch import nn, einsum, Tensor
+from torch.nn import Module, ModuleList
+import torch.nn.functional as F
+
+from models.bs_roformer.attend import Attend
+from torch.utils.checkpoint import checkpoint
+
+from beartype.typing import Tuple, Optional, List, Callable
+from beartype import beartype
+
+from rotary_embedding_torch import RotaryEmbedding
+
+from einops import rearrange, pack, unpack, reduce, repeat
+from einops.layers.torch import Rearrange
+
+from hyper_connections import get_init_and_expand_reduce_stream_functions
+
+from librosa import filters
+
+
+# helper functions
+
+def exists(val):
+    return val is not None
+
+
+def default(v, d):
+    return v if exists(v) else d
+
+
+def pack_one(t, pattern):
+    return pack([t], pattern)
+
+
+def unpack_one(t, ps, pattern):
+    return unpack(t, ps, pattern)[0]
+
+
+def pad_at_dim(t, pad, dim=-1, value=0.):
+    dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
+    zeros = ((0, 0) * dims_from_right)
+    return F.pad(t, (*zeros, *pad), value=value)
+
+
+def l2norm(t):
+    return F.normalize(t, dim=-1, p=2)
+
+
+# norm
+
+class RMSNorm(Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.scale = dim ** 0.5
+        self.gamma = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        return F.normalize(x, dim=-1) * self.scale * self.gamma
+
+
+# attention
+
+class FeedForward(Module):
+    def __init__(
+            self,
+            dim,
+            mult=4,
+            dropout=0.
+    ):
+        super().__init__()
+        dim_inner = int(dim * mult)
+        self.net = nn.Sequential(
+            RMSNorm(dim),
+            nn.Linear(dim, dim_inner),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_inner, dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class Attention(Module):
+    def __init__(
+            self,
+            dim,
+            heads=8,
+            dim_head=64,
+            dropout=0.,
+            rotary_embed=None,
+            flash=True,
+            learned_value_residual_mix=False,
+    ):
+        super().__init__()
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+        dim_inner = heads * dim_head
+
+        self.rotary_embed = rotary_embed
+
+        self.attend = Attend(flash=flash, dropout=dropout)
+
+        self.norm = RMSNorm(dim)
+        self.to_qkv = nn.Linear(dim, dim_inner * 3, bias=False)
+
+        self.to_value_residual_mix = nn.Linear(dim, heads) if learned_value_residual_mix else None
+
+        self.to_gates = nn.Linear(dim, heads)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(dim_inner, dim, bias=False),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, value_residual=None):
+        x = self.norm(x)
+
+        q, k, v = rearrange(self.to_qkv(x), 'b n (qkv h d) -> qkv b h n d', qkv=3, h=self.heads)
+
+        orig_v = v
+
+        if exists(self.to_value_residual_mix):
+            mix = self.to_value_residual_mix(x)
+            mix = rearrange(mix, 'b n h -> b h n 1').sigmoid()
+
+            assert exists(value_residual)
+            v = v.lerp(value_residual, mix)
+
+        if exists(self.rotary_embed):
+            q = self.rotary_embed.rotate_queries_or_keys(q)
+            k = self.rotary_embed.rotate_queries_or_keys(k)
+
+        out = self.attend(q, k, v)
+
+        gates = self.to_gates(x)
+        out = out * rearrange(gates, 'b n h -> b h n 1').sigmoid()
+
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out), orig_v
+
+
+class LinearAttention(Module):
+    """
+    this flavor of linear attention proposed in https://arxiv.org/abs/2106.09681 by El-Nouby et al.
+    """
+
+    @beartype
+    def __init__(
+            self,
+            *,
+            dim,
+            dim_head=32,
+            heads=8,
+            scale=8,
+            flash=False,
+            dropout=0.
+    ):
+        super().__init__()
+        dim_inner = dim_head * heads
+        self.norm = RMSNorm(dim)
+
+        self.to_qkv = nn.Sequential(
+            nn.Linear(dim, dim_inner * 3, bias=False),
+            Rearrange('b n (qkv h d) -> qkv b h d n', qkv=3, h=heads)
+        )
+
+        self.temperature = nn.Parameter(torch.zeros(heads, 1, 1))
+
+        self.attend = Attend(
+            scale=scale,
+            dropout=dropout,
+            flash=flash
+        )
+
+        self.to_out = nn.Sequential(
+            Rearrange('b h d n -> b n (h d)'),
+            nn.Linear(dim_inner, dim, bias=False)
+        )
+
+    def forward(
+            self,
+            x
+    ):
+        x = self.norm(x)
+
+        q, k, v = self.to_qkv(x)
+
+        q, k = map(l2norm, (q, k))
+        q = q * self.temperature.exp()
+
+        out = self.attend(q, k, v)
+
+        return self.to_out(out)
+
+
+class Transformer(Module):
+    def __init__(
+            self,
+            *,
+            dim,
+            depth,
+            dim_head=64,
+            heads=8,
+            attn_dropout=0.,
+            ff_dropout=0.,
+            ff_mult=4,
+            norm_output=True,
+            rotary_embed=None,
+            flash_attn=True,
+            linear_attn=False,
+            add_value_residual=False,
+            num_residual_streams=1,
+    ):
+        super().__init__()
+        self.layers = ModuleList([])
+
+        init_hyper_conn, *_ = get_init_and_expand_reduce_stream_functions(num_residual_streams, disable=num_residual_streams == 1)
+
+        for _ in range(depth):
+            if linear_attn:
+                attn = LinearAttention(dim=dim, dim_head=dim_head, heads=heads, dropout=attn_dropout, flash=flash_attn)
+            else:
+                if num_residual_streams != 1:
+                    attn = init_hyper_conn(dim=dim, branch=Attention(dim=dim, dim_head=dim_head, heads=heads,
+                                                                     dropout=attn_dropout,
+                                                                     rotary_embed=rotary_embed, flash=flash_attn,
+                                                                     learned_value_residual_mix=add_value_residual))
+                else:
+                    attn = Attention(
+                        dim=dim, dim_head=dim_head, heads=heads, dropout=attn_dropout,
+                        rotary_embed=rotary_embed, flash=flash_attn, learned_value_residual_mix=add_value_residual
+                    )
+
+            if num_residual_streams != 1:
+                ff = init_hyper_conn(dim=dim, branch=FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout))
+            else:
+                ff = FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout)
+            self.layers.append(ModuleList([
+                attn,
+                ff,
+            ]))
+
+        self.norm = RMSNorm(dim) if norm_output else nn.Identity()
+
+    def forward(self, x, value_residual=None):
+
+        first_values = None
+        if value_residual is not None:
+            for attn, ff in self.layers:
+                x, next_values = attn(x, value_residual=value_residual)
+                first_values = default(first_values, next_values)
+                x = ff(x)
+        else:
+            # Compatibility with old weights
+            for attn, ff in self.layers:
+                attn_out, next_values = attn(x, value_residual=None)
+                first_values = default(first_values, next_values)
+                x = attn_out + x
+                x = ff(x) + x
+
+        return self.norm(x), first_values
+
+
+# bandsplit module
+
+class BandSplit(Module):
+    @beartype
+    def __init__(
+            self,
+            dim,
+            dim_inputs: Tuple[int, ...]
+    ):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_features = ModuleList([])
+
+        for dim_in in dim_inputs:
+            net = nn.Sequential(
+                RMSNorm(dim_in),
+                nn.Linear(dim_in, dim)
+            )
+
+            self.to_features.append(net)
+
+    def forward(self, x):
+        x = x.split(self.dim_inputs, dim=-1)
+
+        outs = []
+        for split_input, to_feature in zip(x, self.to_features):
+            split_output = to_feature(split_input)
+            outs.append(split_output)
+
+        return torch.stack(outs, dim=-2)
+
+
+def MLP(
+        dim_in,
+        dim_out,
+        dim_hidden=None,
+        depth=1,
+        activation=nn.Tanh
+):
+    dim_hidden = default(dim_hidden, dim_in)
+
+    net = []
+    dims = (dim_in, *((dim_hidden,) * depth), dim_out)
+
+    for ind, (layer_dim_in, layer_dim_out) in enumerate(zip(dims[:-1], dims[1:])):
+        is_last = ind == (len(dims) - 2)
+
+        net.append(nn.Linear(layer_dim_in, layer_dim_out))
+
+        if is_last:
+            continue
+
+        net.append(activation())
+
+    return nn.Sequential(*net)
+
+
+class MaskEstimator(Module):
+    @beartype
+    def __init__(
+            self,
+            dim,
+            dim_inputs: Tuple[int, ...],
+            depth,
+            mlp_expansion_factor=4
+    ):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_freqs = ModuleList([])
+        dim_hidden = dim * mlp_expansion_factor
+
+        for dim_in in dim_inputs:
+            net = []
+
+            mlp = nn.Sequential(
+                MLP(dim, dim_in * 2, dim_hidden=dim_hidden, depth=depth),
+                nn.GLU(dim=-1)
+            )
+
+            self.to_freqs.append(mlp)
+
+    def forward(self, x):
+        x = x.unbind(dim=-2)
+
+        outs = []
+
+        for band_features, mlp in zip(x, self.to_freqs):
+            freq_out = mlp(band_features)
+            outs.append(freq_out)
+
+        return torch.cat(outs, dim=-1)
+
+
+# main class
+
+class MelBandRoformer(Module):
+
+    @beartype
+    def __init__(
+            self,
+            dim,
+            *,
+            depth,
+            stereo=False,
+            num_stems=1,
+            time_transformer_depth=2,
+            freq_transformer_depth=2,
+            linear_transformer_depth=0,
+            num_bands=60,
+            dim_head=64,
+            heads=8,
+            attn_dropout=0.1,
+            ff_dropout=0.1,
+            flash_attn=True,
+            dim_freqs_in=1025,
+            sample_rate=44100,  # needed for mel filter bank from librosa
+            stft_n_fft=2048,
+            stft_hop_length=512,
+            # 10ms at 44100Hz, from sections 4.1, 4.4 in the paper - @faroit recommends // 2 or // 4 for better reconstruction
+            stft_win_length=2048,
+            stft_normalized=False,
+            stft_window_fn: Optional[Callable] = None,
+            mask_estimator_depth=1,
+            multi_stft_resolution_loss_weight=1.,
+            multi_stft_resolutions_window_sizes: Tuple[int, ...] = (4096, 2048, 1024, 512, 256),
+            multi_stft_hop_size=147,
+            multi_stft_normalized=False,
+            multi_stft_window_fn: Callable = torch.hann_window,
+            match_input_audio_length=False,  # if True, pad output tensor to match length of input tensor
+            mlp_expansion_factor=4,
+            use_torch_checkpoint=False,
+            skip_connection=False,
+            use_value_residual_learning=False,
+            num_residual_streams=1,  # set to 1. to disable hyper connections (Default in original is 4)
+    ):
+        super().__init__()
+
+        self.stereo = stereo
+        self.audio_channels = 2 if stereo else 1
+        self.num_stems = num_stems
+        self.use_torch_checkpoint = use_torch_checkpoint
+        self.skip_connection = skip_connection
+        self.num_residual_streams = num_residual_streams
+
+        _, self.expand_stream, self.reduce_stream = get_init_and_expand_reduce_stream_functions(num_residual_streams, disable=num_residual_streams == 1)
+
+        self.layers = ModuleList([])
+
+        transformer_kwargs = dict(
+            dim=dim,
+            heads=heads,
+            dim_head=dim_head,
+            attn_dropout=attn_dropout,
+            ff_dropout=ff_dropout,
+            flash_attn=flash_attn,
+            num_residual_streams=num_residual_streams,
+        )
+
+        time_rotary_embed = RotaryEmbedding(dim=dim_head)
+        freq_rotary_embed = RotaryEmbedding(dim=dim_head)
+
+        for layer_index in range(depth):
+            if use_value_residual_learning:
+                is_first = layer_index == 0
+            else:
+                is_first = True
+
+            tran_modules = []
+            if linear_transformer_depth > 0:
+                tran_modules.append(Transformer(depth=linear_transformer_depth, linear_attn=True, **transformer_kwargs))
+            tran_modules.append(
+                Transformer(depth=time_transformer_depth, rotary_embed=time_rotary_embed, add_value_residual=not is_first, **transformer_kwargs)
+            )
+            tran_modules.append(
+                Transformer(depth=freq_transformer_depth, rotary_embed=freq_rotary_embed, add_value_residual=not is_first, **transformer_kwargs)
+            )
+            self.layers.append(nn.ModuleList(tran_modules))
+
+        self.stft_window_fn = partial(default(stft_window_fn, torch.hann_window), stft_win_length)
+
+        self.stft_kwargs = dict(
+            n_fft=stft_n_fft,
+            hop_length=stft_hop_length,
+            win_length=stft_win_length,
+            normalized=stft_normalized
+        )
+
+        freqs = torch.stft(torch.randn(1, 4096), **self.stft_kwargs, window=torch.ones(stft_n_fft), return_complex=True).shape[1]
+
+        # create mel filter bank
+        # with librosa.filters.mel as in section 2 of paper
+
+        mel_filter_bank_numpy = filters.mel(sr=sample_rate, n_fft=stft_n_fft, n_mels=num_bands)
+
+        mel_filter_bank = torch.from_numpy(mel_filter_bank_numpy)
+
+        # for some reason, it doesn't include the first freq? just force a value for now
+
+        mel_filter_bank[0][0] = 1.
+
+        # In some systems/envs we get 0.0 instead of ~1.9e-18 in the last position,
+        # so let's force a positive value
+
+        mel_filter_bank[-1, -1] = 1.
+
+        # binary as in paper (then estimated masks are averaged for overlapping regions)
+
+        freqs_per_band = mel_filter_bank > 0
+        assert freqs_per_band.any(dim=0).all(), 'all frequencies need to be covered by all bands for now'
+
+        repeated_freq_indices = repeat(torch.arange(freqs), 'f -> b f', b=num_bands)
+        freq_indices = repeated_freq_indices[freqs_per_band]
+
+        if stereo:
+            freq_indices = repeat(freq_indices, 'f -> f s', s=2)
+            freq_indices = freq_indices * 2 + torch.arange(2)
+            freq_indices = rearrange(freq_indices, 'f s -> (f s)')
+
+        self.register_buffer('freq_indices', freq_indices, persistent=False)
+        self.register_buffer('freqs_per_band', freqs_per_band, persistent=False)
+
+        num_freqs_per_band = reduce(freqs_per_band, 'b f -> b', 'sum')
+        num_bands_per_freq = reduce(freqs_per_band, 'b f -> f', 'sum')
+
+        self.register_buffer('num_freqs_per_band', num_freqs_per_band, persistent=False)
+        self.register_buffer('num_bands_per_freq', num_bands_per_freq, persistent=False)
+
+        # band split and mask estimator
+
+        freqs_per_bands_with_complex = tuple(2 * f * self.audio_channels for f in num_freqs_per_band.tolist())
+
+        self.band_split = BandSplit(
+            dim=dim,
+            dim_inputs=freqs_per_bands_with_complex
+        )
+
+        self.mask_estimators = nn.ModuleList([])
+
+        for _ in range(num_stems):
+            mask_estimator = MaskEstimator(
+                dim=dim,
+                dim_inputs=freqs_per_bands_with_complex,
+                depth=mask_estimator_depth,
+                mlp_expansion_factor=mlp_expansion_factor,
+            )
+
+            self.mask_estimators.append(mask_estimator)
+
+        # for the multi-resolution stft loss
+
+        self.multi_stft_resolution_loss_weight = multi_stft_resolution_loss_weight
+        self.multi_stft_resolutions_window_sizes = multi_stft_resolutions_window_sizes
+        self.multi_stft_n_fft = stft_n_fft
+        self.multi_stft_window_fn = multi_stft_window_fn
+
+        self.multi_stft_kwargs = dict(
+            hop_length=multi_stft_hop_size,
+            normalized=multi_stft_normalized
+        )
+
+        self.match_input_audio_length = match_input_audio_length
+
+    def forward(
+            self,
+            raw_audio,
+            target=None,
+            return_loss_breakdown=False
+    ):
+        """
+        einops
+
+        b - batch
+        f - freq
+        t - time
+        s - audio channel (1 for mono, 2 for stereo)
+        n - number of 'stems'
+        c - complex (2)
+        d - feature dimension
+        """
+
+        device = raw_audio.device
+
+        if raw_audio.ndim == 2:
+            raw_audio = rearrange(raw_audio, 'b t -> b 1 t')
+
+        batch, channels, raw_audio_length = raw_audio.shape
+
+        istft_length = raw_audio_length if self.match_input_audio_length else None
+
+        assert (not self.stereo and channels == 1) or (
+                    self.stereo and channels == 2), 'stereo needs to be set to True if passing in audio signal that is stereo (channel dimension of 2). also need to be False if mono (channel dimension of 1)'
+
+        # to stft
+
+        raw_audio, batch_audio_channel_packed_shape = pack_one(raw_audio, '* t')
+
+        stft_window = self.stft_window_fn(device=device)
+
+        stft_repr = torch.stft(raw_audio, **self.stft_kwargs, window=stft_window, return_complex=True)
+        stft_repr = torch.view_as_real(stft_repr)
+
+        stft_repr = unpack_one(stft_repr, batch_audio_channel_packed_shape, '* f t c')
+
+        # merge stereo / mono into the frequency, with frequency leading dimension, for band splitting
+        stft_repr = rearrange(stft_repr,'b s f t c -> b (f s) t c')
+
+        # index out all frequencies for all frequency ranges across bands ascending in one go
+
+        batch_arange = torch.arange(batch, device=device)[..., None]
+
+        # account for stereo
+
+        x = stft_repr[batch_arange, self.freq_indices]
+
+        # fold the complex (real and imag) into the frequencies dimension
+
+        x = rearrange(x, 'b f t c -> b t (f c)')
+
+        if self.use_torch_checkpoint:
+            x = checkpoint(self.band_split, x, use_reentrant=False)
+        else:
+            x = self.band_split(x)
+
+        # value residuals
+        time_v_residual = None
+        freq_v_residual = None
+
+        # maybe expand residual streams
+        if self.num_residual_streams != 1:
+            x = self.expand_stream(x)
+
+        # axial / hierarchical attention
+
+        store = [None] * len(self.layers)
+        for i, transformer_block in enumerate(self.layers):
+
+            if len(transformer_block) == 3:
+                linear_transformer, time_transformer, freq_transformer = transformer_block
+
+                x, ft_ps = pack([x], 'b * d')
+                if self.use_torch_checkpoint:
+                    x = checkpoint(linear_transformer, x, use_reentrant=False)
+                else:
+                    x = linear_transformer(x)
+                x, = unpack(x, ft_ps, 'b * d')
+            else:
+                time_transformer, freq_transformer = transformer_block
+
+            if self.skip_connection:
+                # Sum all previous
+                for j in range(i):
+                    x = x + store[j]
+
+            x = rearrange(x, 'b t f d -> b f t d')
+            x, ps = pack([x], '* t d')
+
+            if self.use_torch_checkpoint:
+                x, next_time_v_residual = checkpoint(time_transformer, x, time_v_residual, use_reentrant=False)
+            else:
+                x, next_time_v_residual = time_transformer(x, time_v_residual)
+            time_v_residual = default(time_v_residual, next_time_v_residual)
+
+            x, = unpack(x, ps, '* t d')
+            x = rearrange(x, 'b f t d -> b t f d')
+            x, ps = pack([x], '* f d')
+
+            if self.use_torch_checkpoint:
+                x, next_freq_v_residual = checkpoint(freq_transformer, x, freq_v_residual, use_reentrant=False)
+            else:
+                x, next_freq_v_residual = freq_transformer(x, value_residual=freq_v_residual)
+            freq_v_residual = default(freq_v_residual, next_freq_v_residual)
+
+            x, = unpack(x, ps, '* f d')
+
+            if self.skip_connection:
+                store[i] = x
+
+        # maybe reduce residual streams
+        if self.num_residual_streams != 1:
+            x = self.reduce_stream(x)
+
+        num_stems = len(self.mask_estimators)
+        if self.use_torch_checkpoint:
+            masks = torch.stack([checkpoint(fn, x, use_reentrant=False) for fn in self.mask_estimators], dim=1)
+        else:
+            masks = torch.stack([fn(x) for fn in self.mask_estimators], dim=1)
+        masks = rearrange(masks, 'b n t (f c) -> b n f t c', c=2)
+
+        # modulate frequency representation
+
+        stft_repr = rearrange(stft_repr, 'b f t c -> b 1 f t c')
+
+        # complex number multiplication
+
+        stft_repr = torch.view_as_complex(stft_repr)
+        masks = torch.view_as_complex(masks)
+
+        masks = masks.type(stft_repr.dtype)
+
+        # need to average the estimated mask for the overlapped frequencies
+
+        scatter_indices = repeat(self.freq_indices, 'f -> b n f t', b=batch, n=num_stems, t=stft_repr.shape[-1])
+
+        stft_repr_expanded_stems = repeat(stft_repr, 'b 1 ... -> b n ...', n=num_stems)
+        masks_summed = torch.zeros_like(stft_repr_expanded_stems).scatter_add_(2, scatter_indices, masks)
+
+        denom = repeat(self.num_bands_per_freq, 'f -> (f r) 1', r=channels)
+
+        masks_averaged = masks_summed / denom.clamp(min=1e-8)
+
+        # modulate stft repr with estimated mask
+
+        stft_repr = stft_repr * masks_averaged
+
+        # istft
+
+        stft_repr = rearrange(stft_repr, 'b n (f s) t -> (b n s) f t', s=self.audio_channels)
+
+        recon_audio = torch.istft(stft_repr, **self.stft_kwargs, window=stft_window, return_complex=False,
+                                  length=istft_length)
+
+        recon_audio = rearrange(recon_audio, '(b n s) t -> b n s t', b=batch, s=self.audio_channels, n=num_stems)
+
+        if num_stems == 1:
+            recon_audio = rearrange(recon_audio, 'b 1 s t -> b s t')
+
+        # if a target is passed in, calculate loss for learning
+
+        if not exists(target):
+            return recon_audio
+
+        if self.num_stems > 1:
+            assert target.ndim == 4 and target.shape[1] == self.num_stems
+
+        if target.ndim == 2:
+            target = rearrange(target, '... t -> ... 1 t')
+
+        target = target[..., :recon_audio.shape[-1]]  # protect against lost length on istft
+
+        loss = F.l1_loss(recon_audio, target)
+
+        multi_stft_resolution_loss = 0.
+
+        for window_size in self.multi_stft_resolutions_window_sizes:
+            res_stft_kwargs = dict(
+                n_fft=max(window_size, self.multi_stft_n_fft),  # not sure what n_fft is across multi resolution stft
+                win_length=window_size,
+                return_complex=True,
+                window=self.multi_stft_window_fn(window_size, device=device),
+                **self.multi_stft_kwargs,
+            )
+
+            recon_Y = torch.stft(rearrange(recon_audio, '... s t -> (... s) t'), **res_stft_kwargs)
+            target_Y = torch.stft(rearrange(target, '... s t -> (... s) t'), **res_stft_kwargs)
+
+            multi_stft_resolution_loss = multi_stft_resolution_loss + F.l1_loss(recon_Y, target_Y)
+
+        weighted_multi_resolution_loss = multi_stft_resolution_loss * self.multi_stft_resolution_loss_weight
+
+        total_loss = loss + weighted_multi_resolution_loss
+
+        if not return_loss_breakdown:
+            return total_loss
+
+        return total_loss, (loss, multi_stft_resolution_loss)

models/conformer_model.py

@@ -0,0 +1,188 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional
+from conformer import Conformer
+
+class NeuralModel(nn.Module):
+    """
+    Принимает |X| STFT: (B, C, F, T_spec) и предсказывает комплексные маски
+    в свернутом виде: (B, 2 * (sources*channels), F, T_spec)
+    где 2 — это [real, imag].
+    """
+    def __init__(
+        self,
+        in_channels: int = 2,
+        sources: int = 2,
+        freq_bins: int = 2049,
+        embed_dim: int = 512,
+        depth: int = 8,
+        dim_head: int = 64,
+        heads: int = 8,
+        ff_mult: int = 4,
+        conv_expansion_factor: int = 2,
+        conv_kernel_size: int = 31,
+        attn_dropout: float = 0.1,
+        ff_dropout: float = 0.1,
+        conv_dropout: float = 0.1,
+    ):
+        super().__init__()
+        self.freq_bins = freq_bins
+        self.in_channels = in_channels
+        self.sources = sources
+        self.out_masks = sources * in_channels
+        self.embed_dim = embed_dim
+
+        self.input_proj_stft = nn.Linear(freq_bins * in_channels, embed_dim)
+        self.model = Conformer(
+            dim=embed_dim,
+            depth=depth,
+            dim_head=dim_head,
+            heads=heads,
+            ff_mult=ff_mult,
+            conv_expansion_factor=conv_expansion_factor,
+            conv_kernel_size=conv_kernel_size,
+            attn_dropout=attn_dropout,
+            ff_dropout=ff_dropout,
+            conv_dropout=conv_dropout,
+        )
+        # 2 = [real, imag]
+        self.output_proj = nn.Linear(embed_dim, freq_bins * self.out_masks * 2)
+
+    def forward(self, x_stft_mag: torch.Tensor) -> torch.Tensor:
+        """
+        x_stft_mag: (B, C, F, T_spec)
+        returns: (B, 2 * (sources*channels), F, T_spec)  — real/imag масок
+        """
+        assert x_stft_mag.dim() == 4, f"Expected (B,C,F,T), got {tuple(x_stft_mag.shape)}"
+        B, C, F, T_spec = x_stft_mag.shape
+        # (B, T_spec, C*F)
+        x_stft_mag = x_stft_mag.permute(0, 3, 1, 2).contiguous().view(B, T_spec, C * F)
+
+        x = self.input_proj_stft(x_stft_mag)     # (B, T_spec, E)
+        x = self.model(x)                        # (B, T_spec, E)
+        x = torch.tanh(x)                        # стабилизируем
+        x = self.output_proj(x)                  # (B, T_spec, F * out_masks * 2)
+
+        # back to (B, 2*out_masks, F, T_spec)
+        x = x.reshape(B, T_spec, self.out_masks * 2, F).permute(0, 2, 3, 1).contiguous()
+        return x
+
+
+class ConformerMSS(nn.Module):
+    """
+    Совместимо с твоим train:
+      forward(x: (B, C, T)) -> y_hat: (B, S, C, T)
+    где S = число источников (sources).
+    Внутри: STFT -> NeuralModel -> комплексные маски -> iSTFT.
+    """
+    def __init__(
+        self,
+        core: NeuralModel,
+        n_fft: int = 4096,
+        hop_length: int = 1024,
+        win_length: Optional[int] = None,
+        center: bool = True,
+    ):
+        super().__init__()
+        self.core = core
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+        self.win_length = win_length if win_length is not None else n_fft
+        self.center = center
+
+        window = torch.hann_window(self.win_length)
+        # окно — буфер, чтобы таскалось на .to(device)
+        self.register_buffer("window", window, persistent=False)
+
+        # sanity-check: freq_bins у core должен совпадать с n_fft//2 + 1
+        expected_bins = n_fft // 2 + 1
+        assert core.freq_bins == expected_bins, (
+            f"NeuralModel.freq_bins={core.freq_bins} != n_fft//2+1={expected_bins}. "
+            f"Поставь freq_bins={expected_bins} при создании core."
+        )
+
+    def _stft(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        x: (B, C, T) -> spec: complex (B, C, F, TT)
+        """
+        assert x.dim() == 3, f"Expected (B,C,T), got {tuple(x.shape)}"
+        B, C, T = x.shape
+        x_bc_t = x.reshape(B * C, T)
+        spec = torch.stft(
+            x_bc_t,
+            n_fft=self.n_fft,
+            hop_length=self.hop_length,
+            win_length=self.win_length,
+            window=self.window.to(x.device),
+            center=self.center,
+            return_complex=True,
+        )  # (B*C, F, TT)
+        F, TT = spec.shape[-2], spec.shape[-1]
+        spec = spec.reshape(B, C, F, TT)
+        return spec
+
+    def _istft(self, spec: torch.Tensor, length: int) -> torch.Tensor:
+        """
+        spec: complex (B, C, F, TT) -> audio: (B, C, T)
+        """
+        B, C, F, TT = spec.shape
+        spec_bc = spec.reshape(B * C, F, TT)
+        y_bc_t = torch.istft(
+            spec_bc,
+            n_fft=self.n_fft,
+            hop_length=self.hop_length,
+            win_length=self.win_length,
+            window=self.window.to(spec.device),
+            center=self.center,
+            length=length,
+        )
+        return y_bc_t.reshape(B, C, -1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        x: (B, C, T)  (микс в волне)
+        returns y_hat: (B, S, C, T) — предсказанные источники в волне
+        """
+        B, C, T = x.shape
+        # 1) STFT
+        mix_spec = self._stft(x)                     # (B, C, F, TT)
+        mix_mag = mix_spec.abs()                     # (B, C, F, TT)
+
+        # 2) Прогон через core -> real/imag масок
+        mask_ri = self.core(mix_mag)                 # (B, 2*(S*C), F, TT2)
+        _, two_sc, F, TT2 = mask_ri.shape
+
+        S = self.core.sources
+        assert two_sc == 2 * (S * C), (
+            f"core вернул {two_sc} каналов масок, ожидалось {2*(S*C)} "
+            f"(2*[real/imag]*[sources*channels]). Проверь in_channels/sources."
+        )
+
+        # 3) Синхронизация по времени (если вдруг TT != TT2)
+        TT = mix_spec.shape[-1]
+        TT_min = min(TT, TT2)
+        if TT != TT_min:
+            mix_spec = mix_spec[..., :TT_min]
+        if TT2 != TT_min:
+            mask_ri = mask_ri[..., :TT_min]
+        TT = TT_min
+        # теперь у обоих время = TT
+
+        # 4) Преобразуем к (B, 2, S, C, F, TT)
+        mask_ri = mask_ri.view(B, 2, S, C, F, TT).contiguous()
+        mask_real = mask_ri[:, 0]                    # (B, S, C, F, TT)
+        mask_imag = mask_ri[:, 1]                    # (B, S, C, F, TT)
+        masks_c = torch.complex(mask_real, mask_imag)
+
+        # 5) Применяем маски к комплексному спектру микса
+        mix_spec_bc = mix_spec.unsqueeze(1)          # (B, 1, C, F, TT)
+        est_specs = masks_c * mix_spec_bc            # (B, S, C, F, TT)
+
+        # 6) iSTFT по каждому источнику
+        outs = []
+        for s in range(S):
+            y_s = self._istft(est_specs[:, s], length=T)  # (B, C, T)
+            outs.append(y_s)
+        y_hat = torch.stack(outs, dim=1)             # (B, S, C, T)
+        return y_hat

models/mel_band_conformer.py

@@ -0,0 +1,424 @@
+from functools import partial
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from conformer import Conformer
+from torch.nn import Module, ModuleList
+from librosa import filters
+from beartype.typing import Tuple, Optional, List, Callable
+from beartype import beartype
+from einops import rearrange, pack, unpack, reduce, repeat
+
+# helper functions
+
+def exists(val):
+    return val is not None
+
+
+def default(v, d):
+    return v if exists(v) else d
+
+
+class RMSNorm(Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.scale = dim ** 0.5
+        self.gamma = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        return F.normalize(x, dim=-1) * self.scale * self.gamma
+
+
+# attention
+
+def MLP(
+        dim_in,
+        dim_out,
+        dim_hidden=None,
+        depth=1,
+        activation=nn.Tanh
+):
+    dim_hidden = default(dim_hidden, dim_in)
+
+    net = []
+    dims = (dim_in, *((dim_hidden,) * depth), dim_out)
+
+    for ind, (layer_dim_in, layer_dim_out) in enumerate(zip(dims[:-1], dims[1:])):
+        is_last = ind == (len(dims) - 2)
+
+        net.append(nn.Linear(layer_dim_in, layer_dim_out))
+
+        if is_last:
+            continue
+
+        net.append(activation())
+
+    return nn.Sequential(*net)
+
+
+class MaskEstimator(Module):
+    @beartype
+    def __init__(
+            self,
+            dim,
+            dim_inputs: Tuple[int, ...],
+            depth,
+            mlp_expansion_factor=4
+    ):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_freqs = ModuleList([])
+        dim_hidden = dim * mlp_expansion_factor
+
+        for dim_in in dim_inputs:
+            net = []
+
+            mlp = nn.Sequential(
+                MLP(dim, dim_in * 2, dim_hidden=dim_hidden, depth=depth),
+                nn.GLU(dim=-1)
+            )
+
+            self.to_freqs.append(mlp)
+
+    def forward(self, x):
+        # split along band dimension and run per-band MLP
+        x = x.unbind(dim=-2)
+
+        outs = []
+
+        for band_features, mlp in zip(x, self.to_freqs):
+            freq_out = mlp(band_features)
+            outs.append(freq_out)
+
+        return torch.cat(outs, dim=-1)
+
+
+class BandSplit(Module):
+    @beartype
+    def __init__(
+            self,
+            dim,
+            dim_inputs: Tuple[int, ...]
+    ):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_features = ModuleList([])
+
+        for dim_in in dim_inputs:
+            net = nn.Sequential(
+                RMSNorm(dim_in),
+                nn.Linear(dim_in, dim)
+            )
+
+            self.to_features.append(net)
+
+    def forward(self, x):
+        # split input into predefined frequency-band chunks
+        x = x.split(self.dim_inputs, dim=-1)
+
+        outs = []
+        for split_input, to_feature in zip(x, self.to_features):
+            split_output = to_feature(split_input)
+            outs.append(split_output)
+
+        # stack back as (bands) axis
+        return torch.stack(outs, dim=-2)
+
+
+class MelBandConformer(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        *,
+        depth: int,
+        stereo: bool = False,
+        num_stems: int = 1,
+        time_conformer_depth: int = 2,
+        freq_conformer_depth: int = 2,
+        num_bands: int = 60,
+        dim_head: int = 64,
+        heads: int = 8,
+        # Conformer params
+        ff_mult: int = 4,
+        conv_expansion_factor: int = 2,
+        conv_kernel_size: int = 31,
+        attn_dropout: float = 0.0,
+        ff_dropout: float = 0.0,
+        conv_dropout: float = 0.0,
+        # STFT
+        dim_freqs_in: int = 1025,
+        sample_rate: int = 44100,
+        stft_n_fft: int = 2048,
+        stft_hop_length: int = 512,
+        stft_win_length: int = 2048,
+        stft_normalized: bool = False,
+        stft_window_fn: Optional[Callable] = None,
+        # Loss
+        mask_estimator_depth: int = 1,
+        multi_stft_resolution_loss_weight: float = 1.0,
+        multi_stft_resolutions_window_sizes: Tuple[int, ...] = (4096, 2048, 1024, 512, 256),
+        multi_stft_hop_size: int = 147,
+        multi_stft_normalized: bool = False,
+        multi_stft_window_fn: Callable = torch.hann_window,
+        match_input_audio_length: bool = False,
+
+        use_torch_checkpoint: bool = False,
+        skip_connection: bool = False,
+    ):
+        super().__init__()
+
+        self.stereo = stereo
+        self.audio_channels = 2 if stereo else 1
+        self.num_stems = num_stems
+        self.use_torch_checkpoint = use_torch_checkpoint
+        self.skip_connection = skip_connection
+
+        self.layers = nn.ModuleList([])
+
+        # Layers per block: [ time-Conformer, freq-Conformer ]
+        conformer_kwargs = dict(
+            dim=dim,
+            dim_head=dim_head,
+            heads=heads,
+            ff_mult=ff_mult,
+            conv_expansion_factor=conv_expansion_factor,
+            conv_kernel_size=conv_kernel_size,
+            attn_dropout=attn_dropout,
+            ff_dropout=ff_dropout,
+            conv_dropout=conv_dropout,
+        )
+
+        for _ in range(depth):
+            time_block = Conformer(depth=time_conformer_depth, **conformer_kwargs)
+            freq_block = Conformer(depth=freq_conformer_depth, **conformer_kwargs)
+            self.layers.append(nn.ModuleList([time_block, freq_block]))
+
+        self.stft_window_fn = partial(stft_window_fn or torch.hann_window, stft_win_length)
+
+        self.stft_kwargs = dict(
+            n_fft=stft_n_fft,
+            hop_length=stft_hop_length,
+            win_length=stft_win_length,
+            normalized=stft_normalized
+        )
+
+        # number of frequency bins produced by STFT (ignoring complex axis)
+        freqs = torch.stft(
+            torch.randn(1, 4096),
+            **self.stft_kwargs,
+            window=torch.ones(stft_n_fft),
+            return_complex=True
+        ).shape[1]
+
+        # build mel filter bank to define band grouping
+        mel_filter_bank_numpy = filters.mel(sr=sample_rate, n_fft=stft_n_fft, n_mels=num_bands)
+        mel_filter_bank = torch.from_numpy(mel_filter_bank_numpy)
+        # ensure coverage at the boundaries
+        mel_filter_bank[0][0] = 1.0
+        mel_filter_bank[-1, -1] = 1.0
+
+        freqs_per_band = mel_filter_bank > 0
+        assert freqs_per_band.any(dim=0).all(), 'all frequency bins must be covered by bands'
+
+        repeated_freq_indices = repeat(torch.arange(freqs), 'f -> b f', b=num_bands)
+        freq_indices = repeated_freq_indices[freqs_per_band]
+
+        if stereo:
+            # duplicate indices for stereo by interleaving channels along the freq axis
+            freq_indices = repeat(freq_indices, 'f -> f s', s=2)
+            freq_indices = freq_indices * 2 + torch.arange(2)
+            freq_indices = rearrange(freq_indices, 'f s -> (f s)')
+
+        self.register_buffer('freq_indices', freq_indices, persistent=False)
+        self.register_buffer('freqs_per_band', freqs_per_band, persistent=False)
+
+        num_freqs_per_band = reduce(freqs_per_band, 'b f -> b', 'sum')
+        num_bands_per_freq = reduce(freqs_per_band, 'b f -> f', 'sum')
+
+        self.register_buffer('num_freqs_per_band', num_freqs_per_band, persistent=False)
+        self.register_buffer('num_bands_per_freq', num_bands_per_freq, persistent=False)
+
+        # BandSplit and MaskEstimator — same structure as your original
+        freqs_per_bands_with_complex = tuple(2 * f * self.audio_channels for f in num_freqs_per_band.tolist())
+
+        self.band_split = BandSplit(
+            dim=dim,
+            dim_inputs=freqs_per_bands_with_complex
+        )
+
+        self.mask_estimators = nn.ModuleList([
+            MaskEstimator(
+                dim=dim,
+                dim_inputs=freqs_per_bands_with_complex,
+                depth=mask_estimator_depth,
+                mlp_expansion_factor=4,  # could be exposed as a parameter
+            )
+            for _ in range(num_stems)
+        ])
+
+        # multi-resolution STFT loss setup
+        self.multi_stft_resolution_loss_weight = multi_stft_resolution_loss_weight
+        self.multi_stft_resolutions_window_sizes = multi_stft_resolutions_window_sizes
+        self.multi_stft_n_fft = stft_n_fft
+        self.multi_stft_window_fn = multi_stft_window_fn
+
+        self.multi_stft_kwargs = dict(
+            hop_length=multi_stft_hop_size,
+            normalized=multi_stft_normalized
+        )
+
+        self.match_input_audio_length = match_input_audio_length
+
+    def forward(
+        self,
+        raw_audio: torch.Tensor,
+        target: Optional[torch.Tensor] = None,
+        return_loss_breakdown: bool = False
+    ):
+        """
+        b - batch
+        f - freq
+        t - time
+        s - audio channel (1 mono / 2 stereo)
+        n - stems
+        c - complex (2)
+        d - feature dim
+        """
+        device = raw_audio.device
+
+        if raw_audio.ndim == 2:
+            raw_audio = rearrange(raw_audio, 'b t -> b 1 t')
+
+        batch, channels, raw_audio_length = raw_audio.shape
+        istft_length = raw_audio_length if self.match_input_audio_length else None
+
+        assert (not self.stereo and channels == 1) or (self.stereo and channels == 2), \
+            'set stereo=True for stereo input (C=2), stereo=False for mono (C=1)'
+
+        # --- STFT ---
+        raw_audio_flat, packed_shape = raw_audio.reshape(-1, raw_audio.shape[-1]), raw_audio.shape[:2]
+        stft_window = self.stft_window_fn(device=device)
+
+        stft_repr = torch.stft(raw_audio_flat, **self.stft_kwargs, window=stft_window, return_complex=True)
+        stft_repr = torch.view_as_real(stft_repr)                  # (B*C, F, T, 2)
+        stft_repr = stft_repr.view(*packed_shape, *stft_repr.shape[1:])  # (b, s, f, t, c)
+
+        # fold channel into frequency axis (as in your setup)
+        stft_repr_fs = rearrange(stft_repr, 'b s f t c -> b (f s) t c')
+
+        # index frequencies by mel bands
+        b_idx = torch.arange(batch, device=device)[..., None]
+        x = stft_repr_fs[b_idx, self.freq_indices]                 # (b, sum(freqs_in_bands), t, c)
+        x = rearrange(x, 'b f t c -> b t (f c)')                   # flatten complex axis into features
+
+        # --- BandSplit -> (b, t, bands, dim) ---
+        if self.use_torch_checkpoint:
+            x = torch.utils.checkpoint.checkpoint(self.band_split, x, use_reentrant=False)
+        else:
+            x = self.band_split(x)
+
+        # --- Axial Conformer (time, then freq) ---
+        store = [None] * len(self.layers)
+
+        for i, (time_conf, freq_conf) in enumerate(self.layers):
+            # Time axis: (b, t, bands, d) -> ((b*bands), t, d)
+            bsz, tlen, bands, d = x.shape
+            x_time = rearrange(x, 'b t f d -> (b f) t d')
+
+            if self.use_torch_checkpoint:
+                x_time = torch.utils.checkpoint.checkpoint(time_conf, x_time, use_reentrant=False)
+            else:
+                x_time = time_conf(x_time)
+
+            x = rearrange(x_time, '(b f) t d -> b t f d', b=bsz, f=bands)
+
+            # Freq axis: (b, t, f, d) -> ((b*t), f, d)
+            bsz, tlen, bands, d = x.shape
+            x_freq = rearrange(x, 'b t f d -> (b t) f d')
+
+            if self.use_torch_checkpoint:
+                x_freq = torch.utils.checkpoint.checkpoint(freq_conf, x_freq, use_reentrant=False)
+            else:
+                x_freq = freq_conf(x_freq)
+
+            x = rearrange(x_freq, '(b t) f d -> b t f d', b=bsz, t=tlen)
+
+            if self.skip_connection:
+                store[i] = x if store[i] is None else store[i] + x
+
+        # --- Mask estimation ---
+        # (b, t, f_bands, d) -> per-stem MLP over bands
+        if self.use_torch_checkpoint:
+            masks = torch.stack([torch.utils.checkpoint.checkpoint(fn, x, use_reentrant=False)
+                                 for fn in self.mask_estimators], dim=1)
+        else:
+            masks = torch.stack([fn(x) for fn in self.mask_estimators], dim=1)
+        masks = rearrange(masks, 'b n t (f c) -> b n f t c', c=2)
+
+        # --- Complex modulation ---
+        stft_repr_c = rearrange(stft_repr, 'b s f t c -> b 1 (f s) t c')
+        stft_repr_c = torch.view_as_complex(stft_repr_c)          # (b, 1, F*S, T)
+        masks_c = torch.view_as_complex(masks)                     # (b, n, F*S, T)
+
+        masks_c = masks_c.type(stft_repr_c.dtype)
+
+        scatter_idx = repeat(self.freq_indices, 'f -> b n f t', b=batch, n=self.num_stems, t=stft_repr_c.shape[-1])
+        stft_repr_expanded = repeat(stft_repr_c, 'b 1 ... -> b n ...', n=self.num_stems)
+
+        masks_summed = torch.zeros_like(stft_repr_expanded).scatter_add_(2, scatter_idx, masks_c)
+        denom = repeat(self.num_bands_per_freq, 'f -> (f r) 1', r=self.audio_channels)
+
+        masks_averaged = masks_summed / denom.clamp(min=1e-8)
+        stft_mod = stft_repr_c * masks_averaged
+
+        # --- iSTFT ---
+        stft_mod = rearrange(stft_mod, 'b n (f s) t -> (b n s) f t', s=self.audio_channels)
+
+        recon_audio = torch.istft(
+            stft_mod,
+            **self.stft_kwargs,
+            window=stft_window,
+            return_complex=False,
+            length=istft_length
+        )
+        recon_audio = rearrange(recon_audio, '(b n s) t -> b n s t', b=batch, s=self.audio_channels, n=self.num_stems)
+
+        if self.num_stems == 1:
+            recon_audio = rearrange(recon_audio, 'b 1 s t -> b s t')
+
+        # Loss
+        if target is None:
+            return recon_audio
+
+        if self.num_stems > 1:
+            assert target.ndim == 4 and target.shape[1] == self.num_stems
+
+        if target.ndim == 2:
+            target = rearrange(target, '... t -> ... 1 t')
+
+        target = target[..., :recon_audio.shape[-1]]
+
+        loss = F.l1_loss(recon_audio, target)
+
+        multi_stft_resolution_loss = 0.0
+        for window_size in self.multi_stft_resolutions_window_sizes:
+            res_stft_kwargs = dict(
+                n_fft=max(window_size, self.multi_stft_n_fft),
+                win_length=window_size,
+                return_complex=True,
+                window=self.multi_stft_window_fn(window_size, device=device),
+                **self.multi_stft_kwargs,
+            )
+
+            recon_Y = torch.stft(rearrange(recon_audio, '... s t -> (... s) t'), **res_stft_kwargs)
+            target_Y = torch.stft(rearrange(target, '... s t -> (... s) t'), **res_stft_kwargs)
+
+            multi_stft_resolution_loss += F.l1_loss(recon_Y, target_Y)
+
+        total_loss = loss + self.multi_stft_resolution_loss_weight * multi_stft_resolution_loss
+
+        if not return_loss_breakdown:
+            return total_loss
+
+        return total_loss, (loss, multi_stft_resolution_loss)

models/scnet/scnet_masked.py

@@ -0,0 +1,419 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from collections import deque
+from .separation import SeparationNet
+import typing as tp
+import math
+
+
+class Swish(nn.Module):
+    def forward(self, x):
+        return x * x.sigmoid()
+
+
+class ConvolutionModule(nn.Module):
+    """
+    Convolution Module in SD block.
+
+    Args:
+        channels (int): input/output channels.
+        depth (int): number of layers in the residual branch. Each layer has its own
+        compress (float): amount of channel compression.
+        kernel (int): kernel size for the convolutions.
+        """
+
+    def __init__(self, channels, depth=2, compress=4, kernel=3):
+        super().__init__()
+        assert kernel % 2 == 1
+        self.depth = abs(depth)
+        hidden_size = int(channels / compress)
+        norm = lambda d: nn.GroupNorm(1, d)
+        self.layers = nn.ModuleList([])
+        for _ in range(self.depth):
+            padding = (kernel // 2)
+            mods = [
+                norm(channels),
+                nn.Conv1d(channels, hidden_size * 2, kernel, padding=padding),
+                nn.GLU(1),
+                nn.Conv1d(hidden_size, hidden_size, kernel, padding=padding, groups=hidden_size),
+                norm(hidden_size),
+                Swish(),
+                nn.Conv1d(hidden_size, channels, 1),
+            ]
+            layer = nn.Sequential(*mods)
+            self.layers.append(layer)
+
+    def forward(self, x):
+        for layer in self.layers:
+            x = x + layer(x)
+        return x
+
+
+class FusionLayer(nn.Module):
+    """
+    A FusionLayer within the decoder.
+
+    Args:
+    - channels (int): Number of input channels.
+    - kernel_size (int, optional): Kernel size for the convolutional layer, defaults to 3.
+    - stride (int, optional): Stride for the convolutional layer, defaults to 1.
+    - padding (int, optional): Padding for the convolutional layer, defaults to 1.
+    """
+
+    def __init__(self, channels, kernel_size=3, stride=1, padding=1):
+        super(FusionLayer, self).__init__()
+        self.conv = nn.Conv2d(channels * 2, channels * 2, kernel_size, stride=stride, padding=padding)
+
+    def forward(self, x, skip=None):
+        if skip is not None:
+            x += skip
+        x = x.repeat(1, 2, 1, 1)
+        x = self.conv(x)
+        x = F.glu(x, dim=1)
+        return x
+
+
+class SDlayer(nn.Module):
+    """
+    Implements a Sparse Down-sample Layer for processing different frequency bands separately.
+
+    Args:
+    - channels_in (int): Input channel count.
+    - channels_out (int): Output channel count.
+    - band_configs (dict): A dictionary containing configuration for each frequency band.
+                           Keys are 'low', 'mid', 'high' for each band, and values are
+                           dictionaries with keys 'SR', 'stride', and 'kernel' for proportion,
+                           stride, and kernel size, respectively.
+    """
+
+    def __init__(self, channels_in, channels_out, band_configs):
+        super(SDlayer, self).__init__()
+
+        # Initializing convolutional layers for each band
+        self.convs = nn.ModuleList()
+        self.strides = []
+        self.kernels = []
+        for config in band_configs.values():
+            self.convs.append(
+                nn.Conv2d(channels_in, channels_out, (config['kernel'], 1), (config['stride'], 1), (0, 0)))
+            self.strides.append(config['stride'])
+            self.kernels.append(config['kernel'])
+
+        # Saving rate proportions for determining splits
+        self.SR_low = band_configs['low']['SR']
+        self.SR_mid = band_configs['mid']['SR']
+
+    def forward(self, x):
+        B, C, Fr, T = x.shape
+        # Define splitting points based on sampling rates
+        splits = [
+            (0, math.ceil(Fr * self.SR_low)),
+            (math.ceil(Fr * self.SR_low), math.ceil(Fr * (self.SR_low + self.SR_mid))),
+            (math.ceil(Fr * (self.SR_low + self.SR_mid)), Fr)
+        ]
+
+        # Processing each band with the corresponding convolution
+        outputs = []
+        original_lengths = []
+        for conv, stride, kernel, (start, end) in zip(self.convs, self.strides, self.kernels, splits):
+            extracted = x[:, :, start:end, :]
+            original_lengths.append(end - start)
+            current_length = extracted.shape[2]
+
+            # padding
+            if stride == 1:
+                total_padding = kernel - stride
+            else:
+                total_padding = (stride - current_length % stride) % stride
+            pad_left = total_padding // 2
+            pad_right = total_padding - pad_left
+
+            padded = F.pad(extracted, (0, 0, pad_left, pad_right))
+
+            output = conv(padded)
+            outputs.append(output)
+
+        return outputs, original_lengths
+
+
+class SUlayer(nn.Module):
+    """
+    Implements a Sparse Up-sample Layer in decoder.
+
+    Args:
+    - channels_in: The number of input channels.
+    - channels_out: The number of output channels.
+    - convtr_configs: Dictionary containing the configurations for transposed convolutions.
+    """
+
+    def __init__(self, channels_in, channels_out, band_configs):
+        super(SUlayer, self).__init__()
+
+        # Initializing convolutional layers for each band
+        self.convtrs = nn.ModuleList([
+            nn.ConvTranspose2d(channels_in, channels_out, [config['kernel'], 1], [config['stride'], 1])
+            for _, config in band_configs.items()
+        ])
+
+    def forward(self, x, lengths, origin_lengths):
+        B, C, Fr, T = x.shape
+        # Define splitting points based on input lengths
+        splits = [
+            (0, lengths[0]),
+            (lengths[0], lengths[0] + lengths[1]),
+            (lengths[0] + lengths[1], None)
+        ]
+        # Processing each band with the corresponding convolution
+        outputs = []
+        for idx, (convtr, (start, end)) in enumerate(zip(self.convtrs, splits)):
+            out = convtr(x[:, :, start:end, :])
+            # Calculate the distance to trim the output symmetrically to original length
+            current_Fr_length = out.shape[2]
+            dist = abs(origin_lengths[idx] - current_Fr_length) // 2
+
+            # Trim the output to the original length symmetrically
+            trimmed_out = out[:, :, dist:dist + origin_lengths[idx], :]
+
+            outputs.append(trimmed_out)
+
+        # Concatenate trimmed outputs along the frequency dimension to return the final tensor
+        x = torch.cat(outputs, dim=2)
+
+        return x
+
+
+class SDblock(nn.Module):
+    """
+    Implements a simplified Sparse Down-sample block in encoder.
+
+    Args:
+    - channels_in (int): Number of input channels.
+    - channels_out (int): Number of output channels.
+    - band_config (dict): Configuration for the SDlayer specifying band splits and convolutions.
+    - conv_config (dict): Configuration for convolution modules applied to each band.
+    - depths (list of int): List specifying the convolution depths for low, mid, and high frequency bands.
+    """
+
+    def __init__(self, channels_in, channels_out, band_configs={}, conv_config={}, depths=[3, 2, 1], kernel_size=3):
+        super(SDblock, self).__init__()
+        self.SDlayer = SDlayer(channels_in, channels_out, band_configs)
+
+        # Dynamically create convolution modules for each band based on depths
+        self.conv_modules = nn.ModuleList([
+            ConvolutionModule(channels_out, depth, **conv_config) for depth in depths
+        ])
+        # Set the kernel_size to an odd number.
+        self.globalconv = nn.Conv2d(channels_out, channels_out, kernel_size, 1, (kernel_size - 1) // 2)
+
+    def forward(self, x):
+        bands, original_lengths = self.SDlayer(x)
+        # B, C, f, T = band.shape
+        bands = [
+            F.gelu(
+                conv(band.permute(0, 2, 1, 3).reshape(-1, band.shape[1], band.shape[3]))
+                .view(band.shape[0], band.shape[2], band.shape[1], band.shape[3])
+                .permute(0, 2, 1, 3)
+            )
+            for conv, band in zip(self.conv_modules, bands)
+
+        ]
+        lengths = [band.size(-2) for band in bands]
+        full_band = torch.cat(bands, dim=2)
+        skip = full_band
+
+        output = self.globalconv(full_band)
+
+        return output, skip, lengths, original_lengths
+
+
+class SCNet(nn.Module):
+    """
+    The implementation of SCNet: Sparse Compression Network for Music Source Separation. Paper: https://arxiv.org/abs/2401.13276.pdf
+
+    Args:
+    - sources (List[str]): List of sources to be separated.
+    - audio_channels (int): Number of audio channels.
+    - nfft (int): Number of FFTs to determine the frequency dimension of the input.
+    - hop_size (int): Hop size for the STFT.
+    - win_size (int): Window size for STFT.
+    - normalized (bool): Whether to normalize the STFT.
+    - dims (List[int]): List of channel dimensions for each block.
+    - band_SR (List[float]): The proportion of each frequency band.
+    - band_stride (List[int]): The down-sampling ratio of each frequency band.
+    - band_kernel (List[int]): The kernel sizes for down-sampling convolution in each frequency band
+    - conv_depths (List[int]): List specifying the number of convolution modules in each SD block.
+    - compress (int): Compression factor for convolution module.
+    - conv_kernel (int): Kernel size for convolution layer in convolution module.
+    - num_dplayer (int): Number of dual-path layers.
+    - expand (int): Expansion factor in the dual-path RNN, default is 1.
+
+    """
+
+    def __init__(self,
+                 sources=['drums', 'bass', 'other', 'vocals'],
+                 audio_channels=2,
+                 # Main structure
+                 dims=[4, 32, 64, 128],  # dims = [4, 64, 128, 256] in SCNet-large
+                 # STFT
+                 nfft=4096,
+                 hop_size=1024,
+                 win_size=4096,
+                 normalized=True,
+                 # SD/SU layer
+                 band_SR=[0.175, 0.392, 0.433],
+                 band_stride=[1, 4, 16],
+                 band_kernel=[3, 4, 16],
+                 # Convolution Module
+                 conv_depths=[3, 2, 1],
+                 compress=4,
+                 conv_kernel=3,
+                 # Dual-path RNN
+                 num_dplayer=6,
+                 expand=1,
+                 ):
+        super().__init__()
+        self.sources = sources
+        self.audio_channels = audio_channels
+        self.dims = dims
+        band_keys = ['low', 'mid', 'high']
+        self.band_configs = {band_keys[i]: {'SR': band_SR[i], 'stride': band_stride[i], 'kernel': band_kernel[i]} for i
+                             in range(len(band_keys))}
+        self.hop_length = hop_size
+        self.conv_config = {
+            'compress': compress,
+            'kernel': conv_kernel,
+        }
+
+        self.embed_dim = dims[0]
+        self.max_f = nfft // 2 + 1
+        self.pos_embed_f = nn.Parameter(torch.zeros(1, self.embed_dim, self.max_f, 1))
+        nn.init.trunc_normal_(self.pos_embed_f, std=.02)
+
+        window = torch.hann_window(window_length=nfft, periodic=True)
+        self.register_buffer('window', window, persistent=False)
+
+        self.stft_config = {
+            'n_fft': nfft,
+            'hop_length': hop_size,
+            'win_length': win_size,
+            'center': True,
+            'normalized': normalized
+        }
+
+        self.encoder = nn.ModuleList()
+        self.decoder = nn.ModuleList()
+
+        for index in range(len(dims) - 1):
+            enc = SDblock(
+                channels_in=dims[index],
+                channels_out=dims[index + 1],
+                band_configs=self.band_configs,
+                conv_config=self.conv_config,
+                depths=conv_depths
+            )
+            self.encoder.append(enc)
+
+            dec = nn.Sequential(
+                FusionLayer(channels=dims[index + 1]),
+                SUlayer(
+                    channels_in=dims[index + 1],
+                    channels_out=dims[index] if index != 0 else dims[index] * len(sources),
+                    band_configs=self.band_configs,
+                )
+            )
+            self.decoder.insert(0, dec)
+
+        self.separation_net = SeparationNet(
+            channels=dims[-1],
+            expand=expand,
+            num_layers=num_dplayer,
+        )
+
+        self.mask_layer = nn.Sequential(
+            nn.Conv2d(
+                4 * len(self.sources),
+                64,
+                kernel_size=3,
+                padding="same"
+            ),
+            nn.GELU(),
+            nn.Conv2d(
+                64,
+                4 * len(self.sources),
+                kernel_size=1,
+                padding="same",
+            ),
+            nn.Tanh()
+        )
+
+    def forward(self, x):
+        # B, C, L = x.shape
+        B = x.shape[0]
+        # In the initial padding, ensure that the number of frames after the STFT (the length of the T dimension) is even,
+        # so that the RFFT operation can be used in the separation network.
+        padding = self.hop_length - x.shape[-1] % self.hop_length
+        if (x.shape[-1] + padding) // self.hop_length % 2 == 0:
+            padding += self.hop_length
+        x = F.pad(x, (0, padding))
+
+        # STFT
+        L = x.shape[-1]
+        x = x.reshape(-1, L)
+        stft_opts = {**self.stft_config, 'window': self.window.to(x.device)}
+        x = torch.stft(x, **stft_opts, return_complex=True)
+        x = torch.view_as_real(x)
+        x = x.permute(0, 3, 1, 2).reshape(x.shape[0] // self.audio_channels, x.shape[3] * self.audio_channels, x.shape[1], x.shape[2])
+
+        B, C, Fr, T = x.shape
+
+        assert C == self.embed_dim, f"Input channel dimension {C} after STFT/reshape doesn't match self.embed_dim {self.embed_dim}"
+        mixture = x.repeat(1, len(self.sources), 1, 1)
+
+        if Fr > self.max_f:
+             print(f"Warning: Input frequency dim {Fr} > max_f {self.max_f}. Positional embedding will be truncated/repeated.")
+             repeats = math.ceil(Fr / self.max_f)
+             pos_f = self.pos_embed_f.repeat(1, 1, repeats, 1)[:, :, :Fr, :]
+        else:
+             pos_f = self.pos_embed_f[:, :, :Fr, :]
+        x = x + pos_f
+
+        save_skip = deque()
+        save_lengths = deque()
+        save_original_lengths = deque()
+        # encoder
+        for sd_layer in self.encoder:
+            x, skip, lengths, original_lengths = sd_layer(x)
+            save_skip.append(skip)
+            save_lengths.append(lengths)
+            save_original_lengths.append(original_lengths)
+
+        # separation
+        x = self.separation_net(x)
+
+        # decoder
+        for fusion_layer, su_layer in self.decoder:
+            x = fusion_layer(x, save_skip.pop())
+            x = su_layer(x, save_lengths.pop(), save_original_lengths.pop())
+
+        mask = self.mask_layer(x)
+
+        # output
+        n = self.dims[0]
+
+        mixture = mixture.view(B, n, -1, Fr, T)
+        mixture = mixture.reshape(-1, 2, Fr, T).permute(0, 2, 3, 1)
+        mixture = torch.view_as_complex(mixture.contiguous())
+
+        mask = mask.view(B, n, -1, Fr, T)
+        mask = mask.reshape(-1, 2, Fr, T).permute(0, 2, 3, 1)
+        mask = torch.view_as_complex(mask.contiguous())
+
+        x = mixture * mask
+
+        x = torch.istft(x, **stft_opts)
+        x = x.reshape(B, len(self.sources), self.audio_channels, -1)
+
+        x = x[:, :, :, :-padding]
+
+        return x

models/scnet/scnet_tran.py

@@ -0,0 +1,668 @@
+import torch
+from collections import deque
+import typing as tp
+import math
+import torch
+import torch.nn as nn
+from torch.nn.modules.rnn import LSTM
+from torch.nn import Module, ModuleList
+from einops import rearrange, pack, unpack, reduce, repeat
+from einops.layers.torch import Rearrange
+import torch.nn.functional as F
+from models.bs_roformer.attend import Attend
+from rotary_embedding_torch import RotaryEmbedding
+
+
+# helper functions
+
+def exists(val):
+    return val is not None
+
+
+def default(v, d):
+    return v if exists(v) else d
+
+
+def pack_one(t, pattern):
+    return pack([t], pattern)
+
+
+def unpack_one(t, ps, pattern):
+    return unpack(t, ps, pattern)[0]
+
+
+def pad_at_dim(t, pad, dim=-1, value=0.):
+    dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
+    zeros = ((0, 0) * dims_from_right)
+    return F.pad(t, (*zeros, *pad), value=value)
+
+
+def l2norm(t):
+    return F.normalize(t, dim=-1, p=2)
+
+
+# norm
+
+class RMSNorm(Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.scale = dim ** 0.5
+        self.gamma = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        return F.normalize(x, dim=-1) * self.scale * self.gamma
+
+
+# attention
+
+class FeedForward(Module):
+    def __init__(
+            self,
+            dim,
+            mult=4,
+            dropout=0.
+    ):
+        super().__init__()
+        dim_inner = int(dim * mult)
+        self.net = nn.Sequential(
+            RMSNorm(dim),
+            nn.Linear(dim, dim_inner),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_inner, dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class Attention(Module):
+    def __init__(
+            self,
+            dim,
+            heads=8,
+            dim_head=64,
+            dropout=0.,
+            rotary_embed=None,
+            flash=True
+    ):
+        super().__init__()
+        self.heads = heads
+        self.scale = dim_head ** -0.5
+        dim_inner = heads * dim_head
+
+        self.rotary_embed = rotary_embed
+
+        self.attend = Attend(flash=flash, dropout=dropout)
+
+        self.norm = RMSNorm(dim)
+        self.to_qkv = nn.Linear(dim, dim_inner * 3, bias=False)
+
+        self.to_gates = nn.Linear(dim, heads)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(dim_inner, dim, bias=False),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x):
+        x = self.norm(x)
+
+        q, k, v = rearrange(self.to_qkv(x), 'b n (qkv h d) -> qkv b h n d', qkv=3, h=self.heads)
+
+        if exists(self.rotary_embed):
+            q = self.rotary_embed.rotate_queries_or_keys(q)
+            k = self.rotary_embed.rotate_queries_or_keys(k)
+
+        out = self.attend(q, k, v)
+
+        gates = self.to_gates(x)
+        out = out * rearrange(gates, 'b n h -> b h n 1').sigmoid()
+
+        out = rearrange(out, 'b h n d -> b n (h d)')
+        return self.to_out(out)
+
+
+class Transformer(Module):
+    def __init__(
+            self,
+            *,
+            dim,
+            depth,
+            dim_head=64,
+            heads=8,
+            attn_dropout=0.,
+            ff_dropout=0.,
+            ff_mult=4,
+            norm_output=True,
+            rotary_embed=None,
+            flash_attn=True,
+            linear_attn=False
+    ):
+        super().__init__()
+        self.layers = ModuleList([])
+
+        for _ in range(depth):
+            attn = Attention(dim=dim, dim_head=dim_head, heads=heads, dropout=attn_dropout, rotary_embed=rotary_embed, flash=flash_attn)
+
+            self.layers.append(ModuleList([
+                attn,
+                FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout)
+            ]))
+
+        self.norm = RMSNorm(dim) if norm_output else nn.Identity()
+
+    def forward(self, x):
+
+        for attn, ff in self.layers:
+            x = attn(x) + x
+            x = ff(x) + x
+
+        return self.norm(x)
+
+
+class FeatureConversion(nn.Module):
+    """
+    Integrates into the adjacent Dual-Path layer.
+
+    Args:
+        channels (int): Number of input channels.
+        inverse (bool): If True, uses ifft; otherwise, uses rfft.
+    """
+
+    def __init__(self, channels, inverse):
+        super().__init__()
+        self.inverse = inverse
+        self.channels = channels
+
+    def forward(self, x):
+        # B, C, F, T = x.shape
+        if self.inverse:
+            x = x.float()
+            x_r = x[:, :self.channels // 2, :, :]
+            x_i = x[:, self.channels // 2:, :, :]
+            x = torch.complex(x_r, x_i)
+            x = torch.fft.irfft(x, dim=3, norm="ortho")
+        else:
+            x = x.float()
+            x = torch.fft.rfft(x, dim=3, norm="ortho")
+            x_real = x.real
+            x_imag = x.imag
+            x = torch.cat([x_real, x_imag], dim=1)
+        return x
+
+
+class DualPathTran(nn.Module):
+    """
+    Dual-Path Transformer in Separation Network.
+
+    Args:
+        d_model (int): The number of expected features in the input (input_size).
+        expand (int): Expansion factor used to calculate the hidden_size of LSTM.
+        bidirectional (bool): If True, becomes a bidirectional LSTM.
+    """
+
+    def __init__(self, d_model, time_rotary_embed, freq_rotary_embed, tran_params):
+        super(DualPathTran, self).__init__()
+
+        self.d_model = d_model
+
+        transformer_kwargs = dict(
+            dim=d_model,
+            heads=tran_params['heads'],
+            dim_head=tran_params['dim_head'],
+            attn_dropout=tran_params['attn_dropout'],
+            ff_dropout=tran_params['ff_dropout'],
+            flash_attn=tran_params['flash_attn']
+        )
+        self.norm_layers = nn.ModuleList([nn.GroupNorm(1, d_model) for _ in range(2)])
+        self.time_layer = Transformer(depth=tran_params['depth'], rotary_embed=time_rotary_embed, **transformer_kwargs)
+        self.freq_layer = Transformer(depth=tran_params['depth'], rotary_embed=freq_rotary_embed, **transformer_kwargs)
+
+    def forward(self, x):
+        B, C, F, T = x.shape
+
+        # Process dual-path rnn
+        original_x = x
+        # Frequency-path
+        x = self.norm_layers[0](x)
+        x = x.transpose(1, 3).contiguous().view(B * T, F, C)
+        # print('XXX', x.shape)
+        x = self.freq_layer(x)
+        x = x.view(B, T, F, C).transpose(1, 3)
+        x = x + original_x
+
+        original_x = x
+        # Time-path
+        x = self.norm_layers[1](x)
+        x = x.transpose(1, 2).contiguous().view(B * F, C, T).transpose(1, 2)
+        # print('RRR', x.shape)
+        x = self.time_layer(x)
+        x = x.transpose(1, 2).contiguous().view(B, F, C, T).transpose(1, 2)
+        x = x + original_x
+
+        return x
+
+
+class SeparationNetTran(nn.Module):
+    """
+    Implements a simplified Sparse Down-sample block in an encoder architecture.
+
+    Args:
+    - channels (int): Number input channels.
+    - expand (int): Expansion factor used to calculate the hidden_size of LSTM.
+    - num_layers (int): Number of dual-path layers.
+    """
+
+    def __init__(self, channels, expand=1, num_layers=6, tran_params=None):
+        super(SeparationNetTran, self).__init__()
+
+        self.num_layers = num_layers
+
+        time_rotary_embed = RotaryEmbedding(dim=tran_params['rotary_embedding_dim'])
+        freq_rotary_embed = RotaryEmbedding(dim=tran_params['rotary_embedding_dim'])
+
+        modules = []
+        for i in range(num_layers):
+            m = DualPathTran(channels * (2 if i % 2 == 1 else 1), time_rotary_embed, freq_rotary_embed, tran_params)
+            modules.append(m)
+        self.dp_modules = nn.ModuleList(modules)
+
+        self.feature_conversion = nn.ModuleList([
+            FeatureConversion(channels * 2, inverse=False if i % 2 == 0 else True) for i in range(num_layers)
+        ])
+
+    def forward(self, x):
+        for i in range(self.num_layers):
+            x = self.dp_modules[i](x)
+            x = self.feature_conversion[i](x)
+        return x
+
+
+class Swish(nn.Module):
+    def forward(self, x):
+        return x * x.sigmoid()
+
+
+class ConvolutionModule(nn.Module):
+    """
+    Convolution Module in SD block.
+
+    Args:
+        channels (int): input/output channels.
+        depth (int): number of layers in the residual branch. Each layer has its own
+        compress (float): amount of channel compression.
+        kernel (int): kernel size for the convolutions.
+        """
+
+    def __init__(self, channels, depth=2, compress=4, kernel=3):
+        super().__init__()
+        assert kernel % 2 == 1
+        self.depth = abs(depth)
+        hidden_size = int(channels / compress)
+        norm = lambda d: nn.GroupNorm(1, d)
+        self.layers = nn.ModuleList([])
+        for _ in range(self.depth):
+            padding = (kernel // 2)
+            mods = [
+                norm(channels),
+                nn.Conv1d(channels, hidden_size * 2, kernel, padding=padding),
+                nn.GLU(1),
+                nn.Conv1d(hidden_size, hidden_size, kernel, padding=padding, groups=hidden_size),
+                norm(hidden_size),
+                Swish(),
+                nn.Conv1d(hidden_size, channels, 1),
+            ]
+            layer = nn.Sequential(*mods)
+            self.layers.append(layer)
+
+    def forward(self, x):
+        for layer in self.layers:
+            x = x + layer(x)
+        return x
+
+
+class FusionLayer(nn.Module):
+    """
+    A FusionLayer within the decoder.
+
+    Args:
+    - channels (int): Number of input channels.
+    - kernel_size (int, optional): Kernel size for the convolutional layer, defaults to 3.
+    - stride (int, optional): Stride for the convolutional layer, defaults to 1.
+    - padding (int, optional): Padding for the convolutional layer, defaults to 1.
+    """
+
+    def __init__(self, channels, kernel_size=3, stride=1, padding=1):
+        super(FusionLayer, self).__init__()
+        self.conv = nn.Conv2d(channels * 2, channels * 2, kernel_size, stride=stride, padding=padding)
+
+    def forward(self, x, skip=None):
+        if skip is not None:
+            x += skip
+        x = x.repeat(1, 2, 1, 1)
+        x = self.conv(x)
+        x = F.glu(x, dim=1)
+        return x
+
+
+class SDlayer(nn.Module):
+    """
+    Implements a Sparse Down-sample Layer for processing different frequency bands separately.
+
+    Args:
+    - channels_in (int): Input channel count.
+    - channels_out (int): Output channel count.
+    - band_configs (dict): A dictionary containing configuration for each frequency band.
+                           Keys are 'low', 'mid', 'high' for each band, and values are
+                           dictionaries with keys 'SR', 'stride', and 'kernel' for proportion,
+                           stride, and kernel size, respectively.
+    """
+
+    def __init__(self, channels_in, channels_out, band_configs):
+        super(SDlayer, self).__init__()
+
+        # Initializing convolutional layers for each band
+        self.convs = nn.ModuleList()
+        self.strides = []
+        self.kernels = []
+        for config in band_configs.values():
+            self.convs.append(
+                nn.Conv2d(channels_in, channels_out, (config['kernel'], 1), (config['stride'], 1), (0, 0)))
+            self.strides.append(config['stride'])
+            self.kernels.append(config['kernel'])
+
+        # Saving rate proportions for determining splits
+        self.SR_low = band_configs['low']['SR']
+        self.SR_mid = band_configs['mid']['SR']
+
+    def forward(self, x):
+        B, C, Fr, T = x.shape
+        # Define splitting points based on sampling rates
+        splits = [
+            (0, math.ceil(Fr * self.SR_low)),
+            (math.ceil(Fr * self.SR_low), math.ceil(Fr * (self.SR_low + self.SR_mid))),
+            (math.ceil(Fr * (self.SR_low + self.SR_mid)), Fr)
+        ]
+
+        # Processing each band with the corresponding convolution
+        outputs = []
+        original_lengths = []
+        for conv, stride, kernel, (start, end) in zip(self.convs, self.strides, self.kernels, splits):
+            extracted = x[:, :, start:end, :]
+            original_lengths.append(end - start)
+            current_length = extracted.shape[2]
+
+            # padding
+            if stride == 1:
+                total_padding = kernel - stride
+            else:
+                total_padding = (stride - current_length % stride) % stride
+            pad_left = total_padding // 2
+            pad_right = total_padding - pad_left
+
+            padded = F.pad(extracted, (0, 0, pad_left, pad_right))
+
+            output = conv(padded)
+            outputs.append(output)
+
+        return outputs, original_lengths
+
+
+class SUlayer(nn.Module):
+    """
+    Implements a Sparse Up-sample Layer in decoder.
+
+    Args:
+    - channels_in: The number of input channels.
+    - channels_out: The number of output channels.
+    - convtr_configs: Dictionary containing the configurations for transposed convolutions.
+    """
+
+    def __init__(self, channels_in, channels_out, band_configs):
+        super(SUlayer, self).__init__()
+
+        # Initializing convolutional layers for each band
+        self.convtrs = nn.ModuleList([
+            nn.ConvTranspose2d(channels_in, channels_out, [config['kernel'], 1], [config['stride'], 1])
+            for _, config in band_configs.items()
+        ])
+
+    def forward(self, x, lengths, origin_lengths):
+        B, C, Fr, T = x.shape
+        # Define splitting points based on input lengths
+        splits = [
+            (0, lengths[0]),
+            (lengths[0], lengths[0] + lengths[1]),
+            (lengths[0] + lengths[1], None)
+        ]
+        # Processing each band with the corresponding convolution
+        outputs = []
+        for idx, (convtr, (start, end)) in enumerate(zip(self.convtrs, splits)):
+            out = convtr(x[:, :, start:end, :])
+            # Calculate the distance to trim the output symmetrically to original length
+            current_Fr_length = out.shape[2]
+            dist = abs(origin_lengths[idx] - current_Fr_length) // 2
+
+            # Trim the output to the original length symmetrically
+            trimmed_out = out[:, :, dist:dist + origin_lengths[idx], :]
+
+            outputs.append(trimmed_out)
+
+        # Concatenate trimmed outputs along the frequency dimension to return the final tensor
+        x = torch.cat(outputs, dim=2)
+
+        return x
+
+
+class SDblock(nn.Module):
+    """
+    Implements a simplified Sparse Down-sample block in encoder.
+
+    Args:
+    - channels_in (int): Number of input channels.
+    - channels_out (int): Number of output channels.
+    - band_config (dict): Configuration for the SDlayer specifying band splits and convolutions.
+    - conv_config (dict): Configuration for convolution modules applied to each band.
+    - depths (list of int): List specifying the convolution depths for low, mid, and high frequency bands.
+    """
+
+    def __init__(self, channels_in, channels_out, band_configs={}, conv_config={}, depths=[3, 2, 1], kernel_size=3):
+        super(SDblock, self).__init__()
+        self.SDlayer = SDlayer(channels_in, channels_out, band_configs)
+
+        # Dynamically create convolution modules for each band based on depths
+        self.conv_modules = nn.ModuleList([
+            ConvolutionModule(channels_out, depth, **conv_config) for depth in depths
+        ])
+        # Set the kernel_size to an odd number.
+        self.globalconv = nn.Conv2d(channels_out, channels_out, kernel_size, 1, (kernel_size - 1) // 2)
+
+    def forward(self, x):
+        bands, original_lengths = self.SDlayer(x)
+        # B, C, f, T = band.shape
+        bands = [
+            F.gelu(
+                conv(band.permute(0, 2, 1, 3).reshape(-1, band.shape[1], band.shape[3]))
+                .view(band.shape[0], band.shape[2], band.shape[1], band.shape[3])
+                .permute(0, 2, 1, 3)
+            )
+            for conv, band in zip(self.conv_modules, bands)
+
+        ]
+        lengths = [band.size(-2) for band in bands]
+        full_band = torch.cat(bands, dim=2)
+        skip = full_band
+
+        output = self.globalconv(full_band)
+
+        return output, skip, lengths, original_lengths
+
+
+class SCNet_Tran(nn.Module):
+    """
+    The implementation of SCNet: Sparse Compression Network for Music Source Separation. Paper: https://arxiv.org/abs/2401.13276.pdf
+    LSTM layers replaced with transformer layers
+
+    Args:
+    - sources (List[str]): List of sources to be separated.
+    - audio_channels (int): Number of audio channels.
+    - nfft (int): Number of FFTs to determine the frequency dimension of the input.
+    - hop_size (int): Hop size for the STFT.
+    - win_size (int): Window size for STFT.
+    - normalized (bool): Whether to normalize the STFT.
+    - dims (List[int]): List of channel dimensions for each block.
+    - band_SR (List[float]): The proportion of each frequency band.
+    - band_stride (List[int]): The down-sampling ratio of each frequency band.
+    - band_kernel (List[int]): The kernel sizes for down-sampling convolution in each frequency band
+    - conv_depths (List[int]): List specifying the number of convolution modules in each SD block.
+    - compress (int): Compression factor for convolution module.
+    - conv_kernel (int): Kernel size for convolution layer in convolution module.
+    - num_dplayer (int): Number of dual-path layers.
+    - expand (int): Expansion factor in the dual-path RNN, default is 1.
+
+    """
+
+    def __init__(
+            self,
+            sources=('drums', 'bass', 'other', 'vocals'),
+            audio_channels=2,
+            # Main structure
+            dims=(4, 32, 64, 128),  # dims = [4, 64, 128, 256] in SCNet-large
+            # STFT
+            nfft=4096,
+            hop_size=1024,
+            win_size=4096,
+            normalized=True,
+            # SD/SU layer
+            band_SR=(0.175, 0.392, 0.433),
+            band_stride=(1, 4, 16),
+            band_kernel=(3, 4, 16),
+            # Convolution Module
+            conv_depths=(3, 2, 1),
+            compress=4,
+            conv_kernel=3,
+            # Dual-path RNN
+            num_dplayer=6,
+            expand=1,
+            tran_rotary_embedding_dim=64,
+            tran_depth=1,
+            tran_heads=8,
+            tran_dim_head=64,
+            tran_attn_dropout=0.0,
+            tran_ff_dropout=0.0,
+            tran_flash_attn=False,
+    ):
+        super().__init__()
+        self.sources = sources
+        self.audio_channels = audio_channels
+        self.dims = dims
+        band_keys = ['low', 'mid', 'high']
+        self.band_configs = {band_keys[i]: {'SR': band_SR[i], 'stride': band_stride[i], 'kernel': band_kernel[i]} for i
+                             in range(len(band_keys))}
+        self.hop_length = hop_size
+        self.conv_config = {
+            'compress': compress,
+            'kernel': conv_kernel,
+        }
+        self.tran_params = {
+            'rotary_embedding_dim': tran_rotary_embedding_dim,
+            'depth': tran_depth,
+            'heads': tran_heads,
+            'dim_head': tran_dim_head,
+            'attn_dropout': tran_attn_dropout,
+            'ff_dropout': tran_ff_dropout,
+            'flash_attn': tran_flash_attn,
+        }
+
+        self.stft_config = {
+            'n_fft': nfft,
+            'hop_length': hop_size,
+            'win_length': win_size,
+            'center': True,
+            'normalized': normalized
+        }
+
+        self.first_conv = nn.Conv2d(dims[0], dims[0], 1, 1, 0, bias=False)
+
+        self.encoder = nn.ModuleList()
+        self.decoder = nn.ModuleList()
+
+        for index in range(len(dims) - 1):
+            enc = SDblock(
+                channels_in=dims[index],
+                channels_out=dims[index + 1],
+                band_configs=self.band_configs,
+                conv_config=self.conv_config,
+                depths=conv_depths
+            )
+            self.encoder.append(enc)
+
+            dec = nn.Sequential(
+                FusionLayer(channels=dims[index + 1]),
+                SUlayer(
+                    channels_in=dims[index + 1],
+                    channels_out=dims[index] if index != 0 else dims[index] * len(sources),
+                    band_configs=self.band_configs,
+                )
+            )
+            self.decoder.insert(0, dec)
+
+        self.separation_net = SeparationNetTran(
+            channels=dims[-1],
+            expand=expand,
+            num_layers=num_dplayer,
+            tran_params=self.tran_params
+        )
+
+    def forward(self, x):
+        # B, C, L = x.shape
+        B = x.shape[0]
+        # In the initial padding, ensure that the number of frames after the STFT (the length of the T dimension) is even,
+        # so that the RFFT operation can be used in the separation network.
+        padding = self.hop_length - x.shape[-1] % self.hop_length
+        if (x.shape[-1] + padding) // self.hop_length % 2 == 0:
+            padding += self.hop_length
+        x = F.pad(x, (0, padding))
+
+        # STFT
+        L = x.shape[-1]
+        x = x.reshape(-1, L)
+        x = torch.stft(x, **self.stft_config, return_complex=True)
+        x = torch.view_as_real(x)
+        x = x.permute(0, 3, 1, 2).reshape(x.shape[0] // self.audio_channels, x.shape[3] * self.audio_channels,
+                                          x.shape[1], x.shape[2])
+
+        B, C, Fr, T = x.shape
+
+        save_skip = deque()
+        save_lengths = deque()
+        save_original_lengths = deque()
+        # encoder
+        for sd_layer in self.encoder:
+            x, skip, lengths, original_lengths = sd_layer(x)
+            save_skip.append(skip)
+            save_lengths.append(lengths)
+            save_original_lengths.append(original_lengths)
+
+        # separation
+        x = self.separation_net(x)
+
+        # decoder
+        for fusion_layer, su_layer in self.decoder:
+            x = fusion_layer(x, save_skip.pop())
+            x = su_layer(x, save_lengths.pop(), save_original_lengths.pop())
+
+        # output
+        n = self.dims[0]
+        x = x.view(B, n, -1, Fr, T)
+
+        x = x.reshape(-1, 2, Fr, T).permute(0, 2, 3, 1)
+        x = torch.view_as_complex(x.contiguous())
+        x = torch.istft(x, **self.stft_config)
+        x = x.reshape(B, len(self.sources), self.audio_channels, -1)
+
+        x = x[:, :, :, :-padding]
+
+        return x

phase_fixer.py

@@ -0,0 +1,139 @@
+import torch
+import torchaudio
+import os
+import gc
+
+def frequency_blend_phases(phase1, phase2, freq_bins, low_cutoff=500, high_cutoff=5000, base_factor=0.25, scale_factor=1.85):
+    if phase1.shape != phase2.shape:
+        raise ValueError("phase1 and phase2 must have the same shape.")
+    if len(freq_bins) != phase1.shape[0]:
+        raise ValueError("freq_bins must have the same length as the number of frequency bins in phase1 and phase2.")
+    if low_cutoff >= high_cutoff:
+        raise ValueError("low_cutoff must be less than high_cutoff.")
+
+    blended_phase = torch.zeros_like(phase1)
+    blend_factors = torch.zeros_like(freq_bins)
+
+    blend_factors[freq_bins < low_cutoff] = base_factor
+    blend_factors[freq_bins > high_cutoff] = base_factor + scale_factor
+
+    in_range_mask = (freq_bins >= low_cutoff) & (freq_bins <= high_cutoff)
+    blend_factors[in_range_mask] = base_factor + scale_factor * (
+        (freq_bins[in_range_mask] - low_cutoff) / (high_cutoff - low_cutoff)
+    )
+
+    for i in range(phase1.shape[0]):
+        blended_phase[i, :] = (1 - blend_factors[i]) * phase1[i, :] + blend_factors[i] * phase2[i, :]
+
+    blended_phase = torch.remainder(blended_phase + torch.pi, 2 * torch.pi) - torch.pi
+
+    return blended_phase
+
+def transfer_magnitude_phase(source_file, target_file, output_folder, transfer_magnitude=False, transfer_phase=True, 
+                              low_cutoff=500, high_cutoff=9000, scale_factor=1.4, output_format='flac'):
+    target_name, target_ext = os.path.splitext(os.path.basename(target_file))
+    
+    target_name = target_name.replace("_other", "").replace("_vocals", "").replace("_instrumental", "")
+    target_name = target_name.replace("_Other", "").replace("_Vocals", "").replace("_Instrumental", "").strip()
+    
+    ext = '.flac' if output_format == 'flac' else '.wav'
+    output_file = os.path.join(output_folder, f"{target_name} (Fixed Instrumental){ext}")
+
+    print(f"Phase Fixing: {os.path.basename(target_file)}...")
+    source_waveform, source_sr = torchaudio.load(source_file)
+    target_waveform, target_sr = torchaudio.load(target_file)
+
+    if source_sr != target_sr:
+        raise ValueError("Sample rates of source and target audio files must match.")
+
+    n_fft = 2048
+    hop_length = 512
+    window = torch.hann_window(n_fft)
+
+    source_stfts = torch.stft(source_waveform, n_fft=n_fft, hop_length=hop_length, window=window, return_complex=True, pad_mode="reflect")
+    target_stfts = torch.stft(target_waveform, n_fft=n_fft, hop_length=hop_length, window=window, return_complex=True, pad_mode="reflect")
+
+    freqs = torch.linspace(0, source_sr // 2, steps=n_fft // 2 + 1)
+
+    modified_stfts = []
+    for source_stft, target_stft in zip(source_stfts, target_stfts):
+        source_mag, source_phs = torch.abs(source_stft), torch.angle(source_stft)
+        target_mag, target_phs = torch.abs(target_stft), torch.angle(target_stft)
+
+        modified_stft = target_stft.clone()
+        if transfer_magnitude:
+            modified_stft = source_mag * torch.exp(1j * torch.angle(modified_stft))
+
+        if transfer_phase:
+            blended_phase = frequency_blend_phases(target_phs, source_phs, freqs, low_cutoff, high_cutoff, scale_factor=scale_factor)
+            modified_stft = torch.abs(modified_stft) * torch.exp(1j * blended_phase)
+
+        modified_stfts.append(modified_stft)
+
+    modified_waveform = torch.istft(
+        torch.stack(modified_stfts),
+        n_fft=n_fft,
+        hop_length=hop_length,
+        window=window,
+        length=source_waveform.size(1)
+    )
+
+    if output_format == 'flac':
+        torchaudio.save(output_file, modified_waveform, target_sr, format="flac", bits_per_sample=16)
+    else:
+        torchaudio.save(output_file, modified_waveform, target_sr)
+    
+    print(f"Saved: {output_file}")
+    return output_file
+
+def process_phase_fix(source_file, target_file, output_folder, low_cutoff=500, high_cutoff=9000, 
+                      scale_factor=1.4, output_format='flac'):
+    os.makedirs(output_folder, exist_ok=True)
+    
+    try:
+        output_file = transfer_magnitude_phase(
+            source_file=source_file,
+            target_file=target_file,
+            output_folder=output_folder,
+            transfer_magnitude=False,
+            transfer_phase=True,
+            low_cutoff=low_cutoff,
+            high_cutoff=high_cutoff,
+            scale_factor=scale_factor,
+            output_format=output_format
+        )
+        gc.collect()
+        torch.cuda.empty_cache() if torch.cuda.is_available() else None
+        return output_file, "Phase fix completed successfully!"
+    except Exception as e:
+        return None, f"Error during phase fix: {str(e)}"
+
+SOURCE_MODELS = [
+    'VOCALS-MelBand-Roformer (by Becruily)',
+    'VOCALS-Mel-Roformer big beta 4 (by unwa)',
+    'VOCALS-Melband-Roformer BigBeta5e (by unwa)',
+    'VOCALS-big_beta6 (by Unwa)',
+    'VOCALS-big_beta6X (by Unwa)',
+    'VOCALS-MelBand-Roformer (by KimberleyJSN)',
+    'VOCALS-MelBand-Roformer Kim FT (by Unwa)',
+    'VOCALS-MelBand-Roformer Kim FT 2 (by Unwa)',
+    'VOCALS-MelBand-Roformer Kim FT 2 Blendless (by unwa)',
+    'VOCALS-Mel-Roformer FT 3 Preview (by unwa)',
+    'VOCALS-BS-Roformer_1296 (by viperx)',
+    'VOCALS-BS-Roformer_1297 (by viperx)',
+    'VOCALS-BS-RoformerLargev1 (by unwa)',
+    'bs_roformer_revive (by unwa)'
+]
+
+TARGET_MODELS = [
+    'INST-MelBand-Roformer (by Becruily)',
+    'INST-Mel-Roformer v1 (by unwa)',
+    'INST-Mel-Roformer v2 (by unwa)',
+    'inst_v1e (by unwa)',
+    'INST-Mel-Roformer v1e+ (by unwa)',
+    'Inst_GaboxV7 (by Gabox)',
+    'INST-VOC-Mel-Roformer a.k.a. duality (by unwa)',
+    'INST-VOC-Mel-Roformer a.k.a. duality v2 (by unwa)',
+    'inst_gabox (by Gabox)',
+    'inst_gaboxFlowersV10 (by Gabox)'
+]

processing.py

@@ -13,9 +13,16 @@ sys.path.append(current_dir)
 
 from datetime import datetime
 from helpers import INPUT_DIR, OLD_OUTPUT_DIR, ENSEMBLE_DIR, AUTO_ENSEMBLE_TEMP, move_old_files, clear_directory, BASE_DIR, clean_model, extract_model_name_from_checkpoint, sanitize_filename, find_clear_segment, save_segment, run_matchering, clamp_percentage
-from model import get_model_config
+from model import get_model_config, get_model_chunk_size
 from apollo_processing import process_with_apollo  # Import Apollo processing
 import torch
+
+# PyTorch optimized backend (always available)
+try:
+    from pytorch_backend import PyTorchBackend
+    PYTORCH_OPTIMIZED_AVAILABLE = True
+except ImportError:
+    PYTORCH_OPTIMIZED_AVAILABLE = False
 import yaml
 import gradio as gr
 import threading
@@ -29,15 +36,25 @@ import locale
 import re
 import psutil
 import concurrent.futures
-from google.oauth2.credentials import Credentials
+from tqdm import tqdm
+
+# Google OAuth imports (optional - for Colab/Google Drive support)
+try:
+    from google.oauth2.credentials import Credentials
+    GOOGLE_OAUTH_AVAILABLE = True
+except ImportError:
+    GOOGLE_OAUTH_AVAILABLE = False
+    Credentials = None
+
 import tempfile
-import traceback
 from urllib.parse import urlparse, quote
 try:
     from google.colab import drive
+    # Verify we're actually in a working Colab environment
     IS_COLAB = True
 except ImportError:
     IS_COLAB = False
+    drive = None
 import matchering as mg
 
 import warnings
@@ -61,14 +78,24 @@ def setup_directories():
     """Create necessary directories and check Google Drive access."""
     if IS_COLAB:
         try:
-            if not os.path.exists('/content/drive/MyDrive'):
+            # Check if Google Drive is already mounted
+            if os.path.exists('/content/drive/MyDrive'):
+                pass  # Already mounted, no action needed
+            else:
                 print("Mounting Google Drive...")
-                from google.colab import drive
-                drive.mount('/content/drive', force_remount=True)
-            if not os.path.exists('/content/drive/MyDrive'):
-                raise RuntimeError("Google Drive mount failed. Please mount manually with 'from google.colab import drive; drive.mount('/content/drive', force_remount=True)'.")
+                try:
+                    from google.colab import drive
+                    drive.mount('/content/drive', force_remount=True)
+                except AttributeError as ae:
+                    # Handle 'NoneType' object has no attribute 'kernel' error
+                    print(f"Warning: Google Drive mount skipped (Colab kernel issue): {str(ae)}")
+                    print("Continuing with local storage...")
+                except Exception as mount_error:
+                    print(f"Warning: Google Drive mount failed: {str(mount_error)}")
+                    print("Continuing with local storage...")
         except Exception as e:
-            raise RuntimeError(f"Failed to mount Google Drive: {str(e)}")
+            print(f"Warning: Google Drive setup error: {str(e)}")
+            print("Continuing without Google Drive...")
     os.makedirs(OUTPUT_DIR, exist_ok=True)
     os.makedirs(INPUT_DIR, exist_ok=True)
     os.makedirs(OLD_OUTPUT_DIR, exist_ok=True)
@@ -85,25 +112,47 @@ def refresh_auto_output():
     except Exception as e:
         return None, f"Error refreshing output: {str(e)}"
 
-import logging
-
-# Loglama ayarları
-logging.basicConfig(
-    level=logging.DEBUG,
-    filename='utils.log',
-    filemode='a',
-    format='%(asctime)s - %(levelname)s - %(message)s'
-)
-
-def update_progress_html(progress_label, progress_percent):
-    """Gradio arayüzü için ilerleme çubuğu HTML'si oluşturur."""
-    progress_percent = min(max(round(progress_percent), 0), 100)  # %1 hassasiyet
+def update_progress_html(progress_label, progress_percent, download_info=None):
+    """Generate progress HTML with smooth animations and optional download percentage.
+    
+    Args:
+        progress_label: Text label to show above the progress bar
+        progress_percent: Overall progress percentage (0-100)
+        download_info: Optional dict with 'filename' and 'percent' for download progress
+    """
+    progress_percent = clamp_percentage(progress_percent)
+    
+    # Determine if processing is active for pulse animation
+    is_active = 0 < progress_percent < 100
+    pulse_style = "animation: progress-pulse 1.5s ease-in-out infinite;" if is_active else ""
+    
+    # Build download sub-bar if downloading
+    download_html = ""
+    if download_info and isinstance(download_info, dict):
+        dl_filename = download_info.get('filename', '')
+        dl_percent = clamp_percentage(download_info.get('percent', 0))
+        download_html = f"""
+        <div style="margin-top: 8px; padding: 8px; background: rgba(0,0,0,0.3); border-radius: 5px;">
+            <div style="font-size: 0.85rem; color: #a0a0a0; margin-bottom: 4px;">{dl_filename} - %{int(dl_percent)}</div>
+            <div style="width: 100%; background-color: #333; border-radius: 4px; overflow: hidden;">
+                <div style="width: {dl_percent}%; height: 14px; background: linear-gradient(90deg, #4ade80, #22d3ee); transition: width 0.3s ease-out; border-radius: 4px;"></div>
+            </div>
+        </div>
+        """
+    
     return f"""
+    <style>
+        @keyframes progress-pulse {{
+            0%, 100% {{ opacity: 1; }}
+            50% {{ opacity: 0.85; }}
+        }}
+    </style>
     <div id="custom-progress" style="margin-top: 10px;">
         <div style="font-size: 1rem; color: #C0C0C0; margin-bottom: 5px;" id="progress-label">{progress_label}</div>
         <div style="width: 100%; background-color: #444; border-radius: 5px; overflow: hidden;">
-            <div id="progress-bar" style="width: {progress_percent}%; height: 20px; background-color: #6e8efb; transition: width 0.3s; max-width: 100%;"></div>
+            <div id="progress-bar" style="width: {progress_percent}%; height: 20px; background: linear-gradient(90deg, #6e8efb, #a855f7); transition: width 0.5s ease-out; max-width: 100%; {pulse_style}"></div>
         </div>
+        {download_html}
     </div>
     """
 
@@ -112,18 +161,38 @@ def extract_model_name_from_checkpoint(checkpoint_path):
         return "Unknown"
     base_name = os.path.basename(checkpoint_path)
     model_name = os.path.splitext(base_name)[0]
-    print(f"Original checkpoint path: {checkpoint_path}, extracted model_name: {model_name}")
     return model_name.strip()
 
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
 def run_command_and_process_files(
     model_type,
     config_path,
     start_check_point,
     INPUT_DIR,
     OUTPUT_DIR,
-    extract_instrumental=False,
-    use_tta=False,
-    demud_phaseremix_inst=False,
+    extract_instrumental,
+    use_tta,
+    demud_phaseremix_inst,
     progress=None,
     use_apollo=True,
     apollo_normal_model="Apollo Universal Model",
@@ -133,52 +202,113 @@ def run_command_and_process_files(
     apollo_overlap=2,
     apollo_method="normal_method",
     apollo_midside_model=None,
-    output_format="wav"
+    output_format="wav",
+    optimize_mode='channels_last',
+    enable_amp=True,
+    enable_tf32=True,
+    enable_cudnn_benchmark=True
 ):
+    """
+    Run inference.py with specified parameters and process output files.
+    This is a generator function that yields progress updates for real-time UI feedback.
+    """
     try:
-        logging.info(f"Starting run_command_and_process_files: model_type={model_type}, config_path={config_path}, inference_chunk_size={inference_chunk_size}, inference_overlap={inference_overlap}")
-
-        # Doğrulama
-        for path, name in [
-            (config_path, "Configuration file"),
-            (start_check_point, "Checkpoint file"),
-            (INPUT_DIR, "Input directory")
-        ]:
-            if not path:
-                raise ValueError(f"{name} is empty")
-            if not os.path.exists(path):
-                raise FileNotFoundError(f"{name} not found: {path}")
+        # Create directories and check Google Drive access
+        setup_directories()
+
+        if not config_path:
+            raise ValueError(f"Configuration path is empty: model_type: {model_type}")
+        if not os.path.exists(config_path):
+            raise FileNotFoundError(f"Configuration file not found: {config_path}")
+        if not start_check_point or not os.path.exists(start_check_point):
+            raise FileNotFoundError(f"Checkpoint file not found: {start_check_point}")
+
+        # Validate inference parameters
+        try:
+            inference_chunk_size = int(inference_chunk_size)
+            inference_overlap = int(inference_overlap)
+        except (TypeError, ValueError) as e:
+            print(f"Invalid inference_chunk_size or inference_overlap: {e}. Defaulting to: inference_chunk_size=352800, inference_overlap=2")
+            inference_chunk_size = 352800
+            inference_overlap = 2
+
+        # Validate Apollo parameters
+        try:
+            apollo_chunk_size = int(apollo_chunk_size)
+            apollo_overlap = int(apollo_overlap)
+        except (TypeError, ValueError) as e:
+            print(f"Invalid apollo_chunk_size or apollo_overlap: {e}. Defaulting to: apollo_chunk_size=19, apollo_overlap=2")
+            apollo_chunk_size = 19
+            apollo_overlap = 2
+
+        # Initial progress yield
+        yield {"progress": 0, "status": "Starting audio separation...", "outputs": None}
+
+        # Always use optimized PyTorch backend
+        python_exe = "python"
+        
+        if PYTORCH_OPTIMIZED_AVAILABLE:
+            from inference_pytorch import INFERENCE_PATH as PYTORCH_INFERENCE_PATH
+            inference_script = PYTORCH_INFERENCE_PATH if os.path.exists(PYTORCH_INFERENCE_PATH) else INFERENCE_PATH
+            print(f"Using PyTorch backend (mode: {optimize_mode})")
+            print(f"   AMP: {enable_amp} | TF32: {enable_tf32} | cuDNN: {enable_cudnn_benchmark}")
+        else:
+            inference_script = INFERENCE_PATH
+            print("Warning: PyTorch optimized backend not available, using standard inference")
+        
+
+
+
+
+
 
-        os.makedirs(OUTPUT_DIR, exist_ok=True)
 
-        # Komut oluştur
-        INFERENCE_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), "inference.py")
         cmd_parts = [
-            "python", INFERENCE_PATH,
-            "--model_type", str(model_type),
-            "--config_path", str(config_path),
-            "--start_check_point", str(start_check_point),
-            "--input_folder", str(INPUT_DIR),
-            "--store_dir", str(OUTPUT_DIR),
+            python_exe, inference_script,
+            "--model_type", model_type,
+            "--config_path", config_path,
+            "--start_check_point", start_check_point,
+            "--input_folder", INPUT_DIR,
+            "--store_dir", OUTPUT_DIR,
             "--chunk_size", str(inference_chunk_size),
             "--overlap", str(inference_overlap),
             "--export_format", f"{output_format} FLOAT"
         ]
-        for flag, value in [
-            ("--extract_instrumental", extract_instrumental),
-            ("--use_tta", use_tta),
-            ("--demud_phaseremix_inst", demud_phaseremix_inst)
-        ]:
-            if value:
-                cmd_parts.append(flag)
-
-        logging.info(f"Executing command: {' '.join(cmd_parts)}")
-        start_time = time.time()
-
-        # Subprocess başlat
+        
+
+
+
+
+
+
+
+
+                
+        # Add optimized backend arguments (always enabled)
+        if PYTORCH_OPTIMIZED_AVAILABLE:
+            cmd_parts.extend([
+                "--optimize_mode", optimize_mode
+            ])
+            if enable_amp:
+                cmd_parts.append("--enable_amp")
+            if enable_tf32:
+                cmd_parts.append("--enable_tf32")
+            if enable_cudnn_benchmark:
+                cmd_parts.append("--enable_cudnn_benchmark")
+        
+        if extract_instrumental:
+            cmd_parts.append("--extract_instrumental")
+        if use_tta:
+            cmd_parts.append("--use_tta")
+        if demud_phaseremix_inst:
+            cmd_parts.append("--demud_phaseremix_inst")
+
+        print(f"Running command: {' '.join(cmd_parts)}")
+        
+        # Use subprocess.Popen for real-time progress capture
         process = subprocess.Popen(
             cmd_parts,
-            cwd=os.path.dirname(os.path.abspath(__file__)),
+            cwd=BASE_DIR,
             stdout=subprocess.PIPE,
             stderr=subprocess.PIPE,
             text=True,
@@ -186,156 +316,188 @@ def run_command_and_process_files(
             universal_newlines=True
         )
 
-        # İlerleme takibi
-        mixture_paths = sorted(glob.glob(os.path.join(INPUT_DIR, '*.*')))
-        total_files = len(mixture_paths)
-        processed_files = 0
-        base_progress_per_file = 80 / total_files if total_files > 0 else 80  # 0-80% ayrıştırma
-
         stderr_output = ""
-        stdout_output = ""
-        while process.poll() is None:
-            line = process.stdout.readline().strip()
-            if line:
-                stdout_output += line + "\n"
-                if i18n("loaded_audio").lower() in line.lower():
-                    processed_files += 1
-                    progress_value = round(processed_files * base_progress_per_file)
-                    if progress is not None and callable(getattr(progress, '__call__', None)):
-                        progress(progress_value / 100, desc=i18n("running_separation").format(processed_files, total_files))
-                        update_progress_html(i18n("running_separation").format(processed_files, total_files), progress_value)
-                logging.debug(line)
-
-            err_line = process.stderr.readline().strip()
-            if err_line:
-                stderr_output += err_line + "\n"
-                logging.error(err_line)
-
-        # Kalan çıktıları topla
-        stdout, stderr = process.communicate()
-        stdout_output += stdout
-        stderr_output += stderr
-        logging.debug(f"Subprocess stdout: {stdout_output}")
-        if stderr_output:
-            logging.error(f"Subprocess stderr: {stderr_output}")
-
+        last_yield_percent = -1
+        downloading_file = None
+        
+        # Read stdout line-by-line for real-time progress updates
+        for line in process.stdout:
+            line_stripped = line.strip()
+            
+            # Check for download progress [SESA_DOWNLOAD]
+            if line_stripped.startswith("[SESA_DOWNLOAD]"):
+                try:
+                    dl_info = line_stripped.replace("[SESA_DOWNLOAD]", "")
+                    if dl_info.startswith("START:"):
+                        downloading_file = dl_info.replace("START:", "")
+                        yield {"progress": 0, "status": i18n("downloading_model_file").format(downloading_file), "outputs": None}
+                    elif dl_info.startswith("END:"):
+                        downloading_file = None
+                    elif ":" in dl_info:
+                        parts = dl_info.rsplit(":", 1)
+                        if len(parts) == 2:
+                            filename, percent_str = parts
+                            download_percent = int(percent_str)
+                            yield {"progress": 0, "status": i18n("downloading_file_progress").format(filename, download_percent), "outputs": None}
+                except (ValueError, TypeError):
+                    pass
+            # Check for [SESA_PROGRESS] prefix from inference script
+            elif line_stripped.startswith("[SESA_PROGRESS]"):
+                try:
+                    percentage_str = line_stripped.replace("[SESA_PROGRESS]", "").strip()
+                    percentage = float(percentage_str) if percentage_str else 0
+                    percentage = min(max(percentage, 0), 100)
+                    
+                    # Scale progress to 0-80% range (saving 80-100% for Apollo)
+                    scaled_progress = int(percentage * 0.8)
+                    
+                    # Yield on every percent change for smooth updates
+                    if int(percentage) != last_yield_percent:
+                        last_yield_percent = int(percentage)
+                        yield {"progress": scaled_progress, "status": f"Separating audio... {int(percentage)}%", "outputs": None}
+                except (ValueError, TypeError):
+                    pass
+            else:
+                # Only print important non-progress lines (errors, warnings, key info)
+                if line_stripped and not line_stripped.startswith(("  ", "    ")):
+                    print(line_stripped)
+        
+        # Capture stderr (only print errors)
+        for line in process.stderr:
+            stderr_output += line
+            line_s = line.strip()
+            if line_s and ("error" in line_s.lower() or "warning" in line_s.lower() or "traceback" in line_s.lower()):
+                print(f"Warning: {line_s}")
+        
+        process.wait()
+        
         if process.returncode != 0:
-            raise RuntimeError(f"Subprocess failed with code {process.returncode}: {stderr_output}")
-
-        elapsed_time = time.time() - start_time
-        logging.info(f"Subprocess completed in {elapsed_time:.2f} seconds")
+            raise subprocess.CalledProcessError(process.returncode, cmd_parts, stderr=stderr_output)
+        
+        yield {"progress": 80, "status": "Separation complete, processing outputs...", "outputs": None}
 
-        # Çıktıları işle
+        # Check if output files were created
         filename_model = extract_model_name_from_checkpoint(start_check_point)
-        output_files = sorted(os.listdir(OUTPUT_DIR))
+        output_files = os.listdir(OUTPUT_DIR)
         if not output_files:
-            raise FileNotFoundError(i18n("no_output_files").format(OUTPUT_DIR))
-
-        # Dosya yeniden adlandırma: 80-90%
-        total_output_files = len(output_files)
-        renamed_files = 0
-        for filename in output_files:
-            file_path = os.path.join(OUTPUT_DIR, filename)
-            if not any(filename.lower().endswith(ext) for ext in ['.mp3', '.wav', '.flac', '.aac', '.ogg', '.m4a']):
-                continue
-            base, ext = os.path.splitext(filename)
-            detected_type = None
-            for type_key in ['vocals', 'instrumental', 'phaseremix', 'drum', 'bass', 'other', 'effects', 'speech', 'music', 'dry', 'male', 'female', 'bleed', 'karaoke']:
-                if type_key.lower() in base.lower():
-                    detected_type = type_key
-                    break
-            type_suffix = detected_type.capitalize() if detected_type else 'Processed'
-            clean_base = sanitize_filename(base.split('_')[0]).rsplit('.', 1)[0]
-            new_filename = f"{clean_base}_{type_suffix}_{filename_model}{ext}"
-            new_file_path = os.path.join(OUTPUT_DIR, new_filename)
-            try:
-                os.rename(file_path, new_file_path)
-                renamed_files += 1
-                progress_value = round(80 + (renamed_files / total_output_files) * 10)
-                if progress is not None and callable(getattr(progress, '__call__', None)):
-                    progress(progress_value / 100, desc=i18n("renaming_files").format(renamed_files, total_output_files))
-                    update_progress_html(i18n("renaming_files").format(renamed_files, total_output_files), progress_value)
-            except Exception as e:
-                logging.error(f"Could not rename {file_path} to {new_file_path}: {e}")
+            raise FileNotFoundError("No output files created in OUTPUT_DIR")
+
+        def rename_files_with_model(folder, filename_model):
+            timestamp = datetime.now().strftime("%d-%m-%Y_%H-%M")
+            for filename in sorted(os.listdir(folder)):
+                file_path = os.path.join(folder, filename)
+                if not any(filename.lower().endswith(ext) for ext in ['.mp3', '.wav', '.flac', '.aac', '.ogg', '.m4a']):
+                    continue
+                base, ext = os.path.splitext(filename)
+                detected_type = None
+                for type_key in ['vocals', 'instrumental', 'instrument', 'phaseremix', 'drum', 'bass', 'other', 'effects', 'speech', 'music', 'dry', 'male', 'female', 'bleed', 'karaoke', 'mid', 'side']:
+                    if type_key in base.lower():
+                        detected_type = type_key
+                        break
+                # Normalize 'instrument' to 'Instrumental' for consistency
+                type_suffix = 'Instrumental' if detected_type == 'instrument' else (detected_type.capitalize() if detected_type else "Processed")
+                clean_base = sanitize_filename(base.split('_')[0]).rsplit('.', 1)[0]
+                new_filename = f"{timestamp}_{clean_base}_{type_suffix}_{filename_model}{ext}"
+                new_file_path = os.path.join(folder, new_filename)
+                try:
+                    os.rename(file_path, new_file_path)
+                except Exception as e:
+                    print(f"Could not rename file: {os.path.basename(file_path)} -> {os.path.basename(new_file_path)}: {str(e)}")
 
-        output_files = sorted(os.listdir(OUTPUT_DIR))
+        rename_files_with_model(OUTPUT_DIR, filename_model)
+
+        output_files = os.listdir(OUTPUT_DIR)
         if not output_files:
-            raise FileNotFoundError(i18n("no_output_files_after_rename").format(OUTPUT_DIR))
+            raise FileNotFoundError("No output files in OUTPUT_DIR after renaming")
 
-        # Çıktıları eşleştir
-        def find_file(keyword):
+        def find_file(keywords):
+            """Find file matching any of the keywords (can be single keyword or list)."""
+            if isinstance(keywords, str):
+                keywords = [keywords]
             matching_files = [
-                os.path.join(OUTPUT_DIR, f) for f in output_files
-                if keyword.lower() in f.lower()
+                os.path.join(OUTPUT_DIR, f) for f in output_files 
+                if any(kw in f.lower() for kw in keywords)
             ]
             return matching_files[0] if matching_files else None
 
         output_list = [
-            find_file('vocals'), find_file('instrumental'), find_file('phaseremix'),
+            find_file('vocals'), find_file(['instrumental', 'instrument']), find_file('phaseremix'),
             find_file('drum'), find_file('bass'), find_file('other'), find_file('effects'),
             find_file('speech'), find_file('music'), find_file('dry'), find_file('male'),
-            find_file('female'), find_file('bleed'), find_file('karaoke')
+            find_file('female'), find_file('bleed'), find_file('karaoke'),
+            find_file('mid'), find_file('side')
         ]
 
-        # Normalizasyon: 90-95%
         normalized_outputs = []
-        for i, output_file in enumerate(output_list):
+        for output_file in output_list:
             if output_file and os.path.exists(output_file):
                 normalized_file = os.path.join(OUTPUT_DIR, f"{sanitize_filename(os.path.splitext(os.path.basename(output_file))[0])}.{output_format}")
-                try:
-                    if output_file.endswith(f".{output_format}") and output_file != normalized_file:
-                        shutil.copy(output_file, normalized_file)
-                    elif output_file != normalized_file:
-                        audio, sr = librosa.load(output_file, sr=None, mono=False)
-                        sf.write(normalized_file, audio.T if audio.ndim > 1 else audio, sr)
-                    else:
-                        normalized_file = output_file
-                    normalized_outputs.append(normalized_file)
-
-                    progress_value = round(90 + (i + 1) / len(output_list) * 5)
-                    if progress is not None and callable(getattr(progress, '__call__', None)):
-                        progress(progress_value / 100, desc=i18n("normalizing_output").format(i + 1, len(output_list)))
-                        update_progress_html(i18n("normalizing_output").format(i + 1, len(output_list), progress_value))
-                except Exception as e:
-                    logging.error(f"Normalization failed for {output_file}: {e}")
-                    normalized_outputs.append(None)
+                if output_file.endswith(f".{output_format}") and output_file != normalized_file:
+                    shutil.copy(output_file, normalized_file)
+                elif output_file != normalized_file:
+                    audio, sr = librosa.load(output_file, sr=None, mono=False)
+                    sf.write(normalized_file, audio.T if audio.ndim > 1 else audio, sr)
+                else:
+                    normalized_file = output_file
+                normalized_outputs.append(normalized_file)
             else:
-                normalized_outputs.append(None)
+                normalized_outputs.append(output_file)
 
-        # Apollo işlemi: 95-100%
+        # Apollo processing
         if use_apollo:
-            try:
-                from apollo_processing import process_with_apollo
-                normalized_outputs = process_with_apollo(
-                    output_files=normalized_outputs,
-                    output_dir=OUTPUT_DIR,
-                    apollo_chunk_size=apollo_chunk_size,
-                    apollo_overlap=apollo_overlap,
-                    apollo_method=apollo_method,
-                    apollo_normal_model=apollo_normal_model,
-                    apollo_midside_model=apollo_midside_model,
-                    output_format=output_format,
-                    progress=lambda p, desc: progress((95 + p * 5) / 100, desc=desc) if progress else None,
-                    total_progress_start=95,
-                    total_progress_end=100
-                )
-            except ImportError:
-                logging.warning("apollo_processing module not found, skipping Apollo processing")
-            except Exception as e:
-                logging.error(f"Apollo processing failed: {e}")
-
-        # Tamamlandı
-        if progress is not None and callable(getattr(progress, '__call__', None)):
-            progress(1.0, desc=i18n("separation_complete"))
-            update_progress_html(i18n("separation_complete"), 100)
+            yield {"progress": 80, "status": "Enhancing with Apollo...", "outputs": None}
+            normalized_outputs = process_with_apollo(
+                output_files=normalized_outputs,
+                output_dir=OUTPUT_DIR,
+                apollo_chunk_size=apollo_chunk_size,
+                apollo_overlap=apollo_overlap,
+                apollo_method=apollo_method,
+                apollo_normal_model=apollo_normal_model,
+                apollo_midside_model=apollo_midside_model,
+                output_format=output_format,
+                progress=progress,
+                total_progress_start=80,
+                total_progress_end=100
+            )
 
-        logging.info(f"Processing completed successfully. Outputs: {normalized_outputs}")
-        return tuple(normalized_outputs)
+        # Final yield with outputs
+        yield {"progress": 100, "status": "Separation complete", "outputs": tuple(normalized_outputs)}
 
+    except subprocess.CalledProcessError as e:
+        print(f"Subprocess failed, code: {e.returncode}: {e.stderr}")
+        yield {"progress": 0, "status": f"Error: {e.stderr}", "outputs": (None,) * 16}
     except Exception as e:
-        logging.error(f"run_command_and_process_files error: {str(e)}")
-        traceback.print_exc(file=sys.stderr)
-        return (None,) * 14
+        print(f"run_command_and_process_files error: {str(e)}")
+        import traceback
+        traceback.print_exc()
+        yield {"progress": 0, "status": f"Error: {str(e)}", "outputs": (None,) * 16}
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
 
 def process_audio(
     input_audio_file,
@@ -343,6 +505,10 @@ def process_audio(
     chunk_size,
     overlap,
     export_format,
+    optimize_mode,
+    enable_amp,
+    enable_tf32,
+    enable_cudnn_benchmark,
     use_tta,
     demud_phaseremix_inst,
     extract_instrumental,
@@ -352,12 +518,16 @@ def process_audio(
     apollo_method,
     apollo_normal_model,
     apollo_midside_model,
-    use_matchering=False,
-    matchering_passes=1,
+    use_matchering,
+    matchering_passes,
     progress=gr.Progress(track_tqdm=True),
     *args,
     **kwargs
 ):
+    """
+    Process audio with the selected model. This is a generator function that yields
+    progress updates for real-time UI feedback.
+    """
     try:
         # Check Google Drive connection
         setup_directories()
@@ -365,29 +535,29 @@ def process_audio(
         if input_audio_file is not None:
             audio_path = input_audio_file.name if hasattr(input_audio_file, 'name') else input_audio_file
         else:
-            return (
-                None, None, None, None, None, None, None, None, None, None, None, None, None, None,
+            yield (
+                None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None,
                 "No audio file provided",
                 update_progress_html("No input provided", 0)
             )
+            return
 
         os.makedirs(OUTPUT_DIR, exist_ok=True)
         os.makedirs(OLD_OUTPUT_DIR, exist_ok=True)
         move_old_files(OUTPUT_DIR)
 
-        print(f"process_audio: model parameter received: {model}")
         # Clean model name, remove ⭐ and other unwanted characters
         clean_model_name = clean_model(model) if not model.startswith("/") else extract_model_name_from_checkpoint(model)
-        print(f"Processing audio: {audio_path}, model: {clean_model_name}")
-
-        print(f"Raw UI inputs - chunk_size: {chunk_size}, overlap: {overlap}, apollo_chunk_size: {apollo_chunk_size}, apollo_overlap: {apollo_overlap}, apollo_method: {apollo_method}")
+        print(f"Processing: {os.path.basename(audio_path)} | Model: {clean_model_name}")
 
         # Validate inference parameters
+        _use_yaml_chunk = (chunk_size == "yaml")
         try:
-            inference_chunk_size = int(chunk_size)
+            inference_chunk_size = 352800 if _use_yaml_chunk else int(chunk_size)
         except (TypeError, ValueError):
             print(f"Invalid chunk_size: {chunk_size}. Defaulting to: 352800.")
             inference_chunk_size = 352800
+            _use_yaml_chunk = True  # fallback: read from YAML
 
         try:
             inference_overlap = int(overlap)
@@ -416,21 +586,32 @@ def process_audio(
         else:
             print(f"Invalid apollo_method: {apollo_method}. Defaulting to: normal_method.")
             apollo_method = "normal_method"
-        print(f"Parsed apollo_method: {apollo_method}")
-
-        print(f"Corrected values - inference_chunk_size: {inference_chunk_size}, inference_overlap: {inference_overlap}, apollo_chunk_size: {apollo_chunk_size}, apollo_overlap: {apollo_overlap}")
-
         # Copy input file to INPUT_DIR
         input_filename = os.path.basename(audio_path)
         dest_path = os.path.join(INPUT_DIR, input_filename)
         shutil.copy(audio_path, dest_path)
-        print(f"Input file copied: {dest_path}")
 
-        # Get model configuration with cleaned model name
+        # Yield status for model loading
+        yield (
+            None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None,
+            f"Loading model: {clean_model_name}...",
+            update_progress_html(f"Loading model: {clean_model_name}", 0)
+        )
+        
+        # Get model configuration with cleaned model name (downloads if needed)
         model_type, config_path, start_check_point = get_model_config(clean_model_name, inference_chunk_size, inference_overlap)
-        print(f"Model configuration: model_type={model_type}, config_path={config_path}, start_check_point={start_check_point}")
 
-        outputs = run_command_and_process_files(
+        # Read the model's native chunk_size from its YAML (now guaranteed to be downloaded)
+        native_chunk = get_model_chunk_size(clean_model_name)
+        if _use_yaml_chunk and native_chunk:
+            print(f"Using model's native chunk_size from YAML: {native_chunk}")
+            inference_chunk_size = native_chunk
+        elif not _use_yaml_chunk:
+            print(f"Using user-selected chunk_size: {inference_chunk_size}")
+
+        # Iterate over the generator and yield progress updates
+        outputs = None
+        for update in run_command_and_process_files(
             model_type=model_type,
             config_path=config_path,
             start_check_point=start_check_point,
@@ -448,17 +629,32 @@ def process_audio(
             apollo_overlap=apollo_overlap,
             apollo_method=apollo_method,
             apollo_midside_model=apollo_midside_model,
-            output_format=export_format.split()[0].lower()
-        )
+            output_format=export_format.split()[0].lower(),
+            optimize_mode=optimize_mode,
+            enable_amp=enable_amp,
+            enable_tf32=enable_tf32,
+            enable_cudnn_benchmark=enable_cudnn_benchmark
+        ):
+            if update.get("outputs") is not None:
+                outputs = update["outputs"]
+            # Yield progress update to Gradio
+            yield (
+                None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None,
+                update["status"],
+                update_progress_html(update["status"], update["progress"])
+            )
 
         if outputs is None or all(output is None for output in outputs):
             raise ValueError("run_command_and_process_files returned None or all None outputs")
 
         # Apply Matchering (if enabled)
         if use_matchering:
-            # Progress update for Matchering
-            if progress is not None and callable(getattr(progress, '__call__', None)):
-                progress(90, desc="Applying Matchering")
+            # Yield progress update for Matchering
+            yield (
+                None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None,
+                "Applying Matchering...",
+                update_progress_html("Applying Matchering...", 90)
+            )
 
             # Find clean segment from original audio
             segment_start, segment_end, segment_audio = find_clear_segment(audio_path)
@@ -488,12 +684,11 @@ def process_audio(
 
             outputs = tuple(mastered_outputs)
 
-        if progress is not None and callable(getattr(progress, '__call__', None)):
-            progress(100, desc="Processing complete")
-
-        return (
+        # Final yield with all outputs
+        yield (
             outputs[0], outputs[1], outputs[2], outputs[3], outputs[4], outputs[5], outputs[6],
             outputs[7], outputs[8], outputs[9], outputs[10], outputs[11], outputs[12], outputs[13],
+            outputs[14], outputs[15],
             "Audio processing completed",
             update_progress_html("Audio processing completed", 100)
         )
@@ -502,8 +697,8 @@ def process_audio(
         print(f"process_audio error: {str(e)}")
         import traceback
         traceback.print_exc()
-        return (
-            None, None, None, None, None, None, None, None, None, None, None, None, None, None,
+        yield (
+            None, None, None, None, None, None, None, None, None, None, None, None, None, None, None, None,
             f"Error occurred: {str(e)}",
             update_progress_html("Error occurred", 0)
         )
@@ -534,23 +729,63 @@ def ensemble_audio_fn(files, method, weights, progress=gr.Progress()):
             ensemble_args += ["--weights", *weights_list]
         
         progress(0, desc="Starting ensemble process", total=100)
-        result = subprocess.run(
+        
+        # Run ensemble subprocess with real-time output capture
+        process = subprocess.Popen(
             ["python", "ensemble.py"] + ensemble_args,
-            capture_output=True,
-            text=True
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            text=True,
+            bufsize=1,
+            universal_newlines=True
         )
         
-        start_time = time.time()
-        total_estimated_time = 10.0  # Adjust based on actual ensemble duration
-        elapsed_time = 0
-        while elapsed_time < total_estimated_time:
-            elapsed_time = time.time() - start_time
-            progress_value = (elapsed_time / total_estimated_time) * 100
-            progress_value = clamp_percentage(progress_value)
-            progress(progress_value, desc=f"Ensembling progress: {progress_value}%")
-            time.sleep(0.1)
+        stdout_output = ""
+        stderr_output = ""
+        
+        # Read output in real-time and capture actual progress
+        for line in process.stdout:
+            stdout_output += line
+            line_stripped = line.strip()
+            
+            # Capture real progress percentage from ensemble.py with new format
+            if line_stripped.startswith("[SESA_PROGRESS]"):
+                try:
+                    percent_str = line_stripped.replace("[SESA_PROGRESS]", "").strip()
+                    percent = int(float(percent_str)) if percent_str else 0
+                    percent = min(max(percent, 0), 100)
+                    progress(percent, desc=f"Ensemble progress: {percent}%")
+                except (ValueError, TypeError):
+                    pass
+            # Legacy format support
+            elif line_stripped.startswith("Progress:"):
+                try:
+                    percent = int(line_stripped.split(":")[1].strip().replace("%", ""))
+                    percent = min(max(percent, 0), 100)
+                    progress(percent, desc=f"Ensemble progress: {percent}%")
+                except (ValueError, IndexError):
+                    pass
+            elif "loading" in line.lower():
+                print(f"Ensemble: {line_stripped}")
+                progress(5, desc="Loading audio files for ensemble...")
+            elif "processing ensemble" in line.lower():
+                print(f"Ensemble: {line_stripped}")
+                progress(10, desc="Starting ensemble processing...")
+            elif "saving" in line.lower():
+                print(f"Ensemble: {line_stripped}")
+                progress(95, desc="Saving ensemble output...")
+            elif line_stripped and not line_stripped.startswith("[SESA_PROGRESS]") and not line_stripped.startswith("Progress:"):
+                # Only print non-progress messages
+                print(f"Ensemble: {line_stripped}")
+        
+        for line in process.stderr:
+            stderr_output += line
+            print(f"Ensemble stderr: {line.strip()}")
         
-        progress(100, desc="Finalizing ensemble output")
+        process.wait()
+        result = type('Result', (), {'stdout': stdout_output, 'stderr': stderr_output, 'returncode': process.returncode})()
+        
+        progress(100, desc="Ensemble complete")
         log = f"Success: {result.stdout}" if not result.stderr else f"Error: {result.stderr}"
         return output_path, log
 
@@ -602,18 +837,15 @@ def auto_ensemble_process(
         input_filename = os.path.basename(audio_path)
         dest_path = os.path.join(INPUT_DIR, input_filename)
         shutil.copy(audio_path, dest_path)
-        print(f"Input file copied: {dest_path}")
 
         # Parse apollo method
         if auto_apollo_method in ["2", 2]:
             auto_apollo_method = "mid_side_method"
         elif auto_apollo_method in ["1", 1]:
             auto_apollo_method = "normal_method"
-        print(f"Parsed auto_apollo_method: {auto_apollo_method}")
 
         corrected_auto_chunk_size = int(auto_apollo_chunk_size)
         corrected_auto_overlap = int(auto_apollo_overlap)
-        print(f"Corrected values - auto_apollo_chunk_size: {corrected_auto_chunk_size}, auto_apollo_overlap: {corrected_auto_overlap}")
 
         # Setup temporary directories
         auto_ensemble_temp = os.path.join(BASE_DIR, "auto_ensemble_temp")
@@ -627,7 +859,6 @@ def auto_ensemble_process(
 
         for i, model in enumerate(selected_models):
             clean_model_name = clean_model(model)
-            print(f"Processing model {i+1}/{total_models}: Original={model}, Cleaned={clean_model_name}")
             model_output_dir = os.path.join(auto_ensemble_temp, clean_model_name)
             os.makedirs(model_output_dir, exist_ok=True)
 
@@ -639,7 +870,12 @@ def auto_ensemble_process(
             )
 
             model_type, config_path, start_check_point = get_model_config(clean_model_name, auto_chunk_size, auto_overlap)
-            print(f"Model configuration: model_type={model_type}, config_path={config_path}, start_check_point={start_check_point}")
+
+            # Read the model's native chunk_size from its YAML after download
+            native_chunk = get_model_chunk_size(clean_model_name)
+            effective_chunk_size = native_chunk if native_chunk else auto_chunk_size
+            if native_chunk:
+                print(f"Using model's native chunk_size from YAML: {native_chunk} (UI value was: {auto_chunk_size})")
 
             cmd = [
                 "python", INFERENCE_PATH,
@@ -648,7 +884,7 @@ def auto_ensemble_process(
                 "--start_check_point", start_check_point,
                 "--input_folder", INPUT_DIR,
                 "--store_dir", model_output_dir,
-                "--chunk_size", str(auto_chunk_size),
+                "--chunk_size", str(effective_chunk_size),
                 "--overlap", str(auto_overlap),
                 "--export_format", f"{export_format.split()[0].lower()} FLOAT"
             ]
@@ -668,20 +904,83 @@ def auto_ensemble_process(
             )
 
             stderr_output = ""
+            last_yield_percent = -1
+            downloading_file = None
+            
             for line in process.stdout:
-                print(line.strip())
-                if "Progress:" in line:
+                line_stripped = line.strip()
+                
+                # Check for download progress [SESA_DOWNLOAD]
+                if line_stripped.startswith("[SESA_DOWNLOAD]"):
+                    try:
+                        dl_info = line_stripped.replace("[SESA_DOWNLOAD]", "")
+                        if dl_info.startswith("START:"):
+                            downloading_file = dl_info.replace("START:", "")
+                            yield None, i18n("downloading_model_file").format(downloading_file), update_progress_html(
+                                i18n("downloading_model_file").format(downloading_file),
+                                i * model_progress_per_step,
+                                download_info={"filename": downloading_file, "percent": 0}
+                            )
+                        elif dl_info.startswith("END:"):
+                            downloading_file = None
+                        elif ":" in dl_info:
+                            parts = dl_info.rsplit(":", 1)
+                            if len(parts) == 2:
+                                filename, percent_str = parts
+                                download_percent = int(percent_str)
+                                yield None, i18n("downloading_file_progress").format(filename, download_percent), update_progress_html(
+                                    i18n("downloading_model_file").format(filename),
+                                    i * model_progress_per_step,
+                                    download_info={"filename": filename, "percent": download_percent}
+                                )
+                    except (ValueError, TypeError):
+                        pass
+                # Check for unique progress prefix [SESA_PROGRESS]
+                elif line_stripped.startswith("[SESA_PROGRESS]"):
                     try:
-                        percentage = float(re.search(r"Progress: (\d+\.\d+)%", line).group(1))
+                        # Extract percentage from [SESA_PROGRESS]XX format
+                        percentage_str = line_stripped.replace("[SESA_PROGRESS]", "").strip()
+                        percentage = float(percentage_str) if percentage_str else 0
+                        percentage = min(max(percentage, 0), 100)  # Clamp to 0-100
+                        
                         model_percentage = (percentage / 100) * model_progress_per_step
                         current_progress = (i * model_progress_per_step) + model_percentage
                         current_progress = clamp_percentage(current_progress)
-                        yield None, i18n("loading_model").format(i+1, total_models, clean_model_name), update_progress_html(
-                            i18n("loading_model_progress").format(i+1, total_models, clean_model_name, current_progress),
-                            current_progress
-                        )
-                    except (AttributeError, ValueError) as e:
-                        print(f"Progress parsing error: {e}")
+                        
+                        # Yield on every percent change for smooth updates
+                        if int(percentage) != last_yield_percent:
+                            last_yield_percent = int(percentage)
+                            yield None, i18n("loading_model_progress_label").format(i+1, total_models, clean_model_name, int(percentage)), update_progress_html(
+                                f"Model {i+1}/{total_models}: {clean_model_name} - {int(percentage)}%",
+                                current_progress
+                            )
+                    except (ValueError, TypeError):
+                        # Silently ignore parsing errors for progress lines
+                        pass
+                # Also support legacy "Progress: XX%" format for backwards compatibility
+                elif line_stripped.startswith("Progress:"):
+                    try:
+                        match = re.search(r"Progress:\s*(\d+(?:\.\d+)?)%?", line_stripped)
+                        if match:
+                            percentage = float(match.group(1))
+                            percentage = min(max(percentage, 0), 100)
+                            
+                            model_percentage = (percentage / 100) * model_progress_per_step
+                            current_progress = (i * model_progress_per_step) + model_percentage
+                            current_progress = clamp_percentage(current_progress)
+                            
+                            if int(percentage) != last_yield_percent:
+                                last_yield_percent = int(percentage)
+                                yield None, i18n("loading_model_progress_label").format(i+1, total_models, clean_model_name, int(percentage)), update_progress_html(
+                                    f"Model {i+1}/{total_models}: {clean_model_name} - {int(percentage)}%",
+                                    current_progress
+                                )
+                    except (ValueError, TypeError):
+                        pass
+                else:
+                    # Print non-progress lines
+                    if line_stripped:
+                        print(line_stripped)
 
             for line in process.stderr:
                 stderr_output += line
@@ -868,6 +1167,8 @@ def auto_ensemble_process(
                 print(f"File not found in Google Drive, copying from local path: {output_path}")
                 shutil.copy(output_path, drive_output_path)
                 print(f"Copied to Google Drive: {drive_output_path}")
+        else:
+            drive_output_path = output_path
 
         yield output_path, i18n("success_output_created") + f" Saved to {drive_output_path if IS_COLAB else output_path}", update_progress_html(
             i18n("ensemble_completed"), 100

pytorch_backend.py

@@ -0,0 +1,621 @@
+# coding: utf-8
+__author__ = 'PyTorch Backend Implementation'
+
+import os
+import pickle
+import numpy as np
+import torch
+import torch.nn as nn
+from typing import Dict, Tuple, Optional, Any
+import warnings
+import hashlib
+import time
+
+# Suppress channels_last warnings for 3D audio tensors
+warnings.filterwarnings("ignore", message=".*channels_last.*")
+warnings.filterwarnings("ignore", message=".*rank 3.*")
+
+
+class PyTorchBackend:
+    """
+    ULTRA-OPTIMIZED PyTorch backend for model inference.
+    Provides various optimization techniques for maximum speed.
+    """
+    
+    def __init__(self, device='cuda:0', optimize_mode='channels_last'):
+        """
+        Initialize ULTRA-OPTIMIZED PyTorch backend.
+        
+        Parameters:
+        ----------
+        device : str
+            Device to use for inference (cuda:0, cpu, mps, etc.)
+        optimize_mode : str
+            Optimization mode: 'channels_last' (recommended), 'compile', 'jit', or 'default'
+        """
+        self.device = device
+        self.optimize_mode = optimize_mode
+        self.model = None
+        self.compiled_model = None
+        
+        # Check device availability
+        if device.startswith('cuda') and not torch.cuda.is_available():
+            warnings.warn("CUDA not available, falling back to CPU")
+            self.device = 'cpu'
+        elif device == 'mps' and not torch.backends.mps.is_available():
+            warnings.warn("MPS not available, falling back to CPU")
+            self.device = 'cpu'
+        
+        # Apply ultra optimization settings
+        self._apply_ultra_optimizations()
+    
+    def _apply_ultra_optimizations(self):
+        """Apply ultra-speed optimizations globally."""
+        if self.device.startswith('cuda'):
+            # Enable all CUDA optimizations
+            torch.backends.cudnn.benchmark = True
+            torch.backends.cuda.matmul.allow_tf32 = True
+            torch.backends.cudnn.allow_tf32 = True
+            
+            # Set optimal CUDA settings
+            torch.backends.cudnn.deterministic = False
+            torch.backends.cudnn.enabled = True
+            
+            # Enable cuBLAS optimizations
+            os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'
+        
+        # Optimize CPU inference
+        if self.device == 'cpu':
+            import multiprocessing
+            num_threads = multiprocessing.cpu_count()
+            torch.set_num_threads(num_threads)
+            torch.set_num_interop_threads(num_threads)
+            print(f"CPU threads set to {num_threads}")
+    
+    def optimize_model(
+        self,
+        model: nn.Module,
+        example_input: Optional[torch.Tensor] = None,
+        use_amp: bool = True,
+        use_channels_last: bool = True
+    ) -> nn.Module:
+        """
+        Optimize PyTorch model for inference.
+        
+        Parameters:
+        ----------
+        model : nn.Module
+            PyTorch model to optimize
+        example_input : Optional[torch.Tensor]
+            Example input for optimization (required for some modes)
+        use_amp : bool
+            Use automatic mixed precision (AMP)
+        use_channels_last : bool
+            Use channels-last memory format
+            
+        Returns:
+        -------
+        nn.Module
+            Optimized model
+        """
+        print(f"Optimizing model with mode: {self.optimize_mode}")
+        
+        self.model = model.eval().to(self.device)
+        self.use_amp = use_amp
+        
+        # Disable gradients for all parameters (inference only)
+        for param in self.model.parameters():
+            param.requires_grad = False
+        
+        # Apply memory format optimization (default: channels_last for CUDA)
+        # Note: Audio models use 3D tensors, so channels_last is applied only where beneficial
+        if use_channels_last and self.device.startswith('cuda'):
+            print("  Using channels-last optimization")
+            # Only apply to model if it has 4D conv layers, otherwise skip silently
+            try:
+                with warnings.catch_warnings():
+                    warnings.simplefilter("ignore")
+                    self.model = self.model.to(memory_format=torch.channels_last)
+            except Exception:
+                pass  # Silently skip for models that don't support channels_last
+        
+        # Set model to inference mode
+        torch.set_grad_enabled(False)
+        
+        # Apply optimization based on mode
+        if self.optimize_mode == 'compile':
+            self.compiled_model = self._compile_model(self.model)
+        elif self.optimize_mode == 'jit':
+            if example_input is None:
+                raise ValueError("example_input required for JIT optimization")
+            self.compiled_model = self._jit_trace_model(self.model, example_input)
+        elif self.optimize_mode == 'channels_last':
+            self.compiled_model = self.model
+        else:
+            print("  Using default optimization")
+            self.compiled_model = self.model
+        
+        # Apply fusion optimizations if possible
+        try:
+            if hasattr(torch.nn.utils, 'fusion'):
+                self.compiled_model = torch.nn.utils.fusion.fuse_conv_bn_eval(self.compiled_model)
+                print("  Conv-BN fusion applied")
+        except:
+            pass
+        
+        print("Optimization complete")
+        return self.compiled_model
+    
+    def _compile_model(self, model: nn.Module) -> nn.Module:
+        """
+        Compile model using torch.compile (PyTorch 2.0+) with ULTRA optimization.
+        
+        Parameters:
+        ----------
+        model : nn.Module
+            Model to compile
+            
+        Returns:
+        -------
+        nn.Module
+            Compiled model
+        """
+        try:
+            if hasattr(torch, 'compile'):
+                print("  Compiling model with torch.compile")
+                # Try max-autotune for best performance
+                try:
+                    compiled = torch.compile(model, mode='max-autotune', fullgraph=True)
+                    print("  Using max-autotune mode")
+                    return compiled
+                except:
+                    # Fallback to reduce-overhead
+                    compiled = torch.compile(model, mode='reduce-overhead')
+                    print("  Using reduce-overhead mode")
+                    return compiled
+            else:
+                print("  torch.compile not available (requires PyTorch 2.0+)")
+                return model
+        except Exception as e:
+            print(f"  Compilation failed: {e}")
+            return model
+    
+    def _jit_trace_model(self, model: nn.Module, example_input: torch.Tensor) -> nn.Module:
+        """
+        Trace model using TorchScript JIT.
+        
+        Parameters:
+        ----------
+        model : nn.Module
+            Model to trace
+        example_input : torch.Tensor
+            Example input for tracing
+            
+        Returns:
+        -------
+        nn.Module
+            Traced model
+        """
+        try:
+            print("  → Tracing model with TorchScript JIT")
+            with torch.no_grad():
+                traced = torch.jit.trace(model, example_input)
+            traced = torch.jit.optimize_for_inference(traced)
+            return traced
+        except Exception as e:
+            print(f"  JIT tracing failed: {e}")
+            return model
+    
+    def save_optimized_model(self, save_path: str):
+        """
+        Save optimized model to file.
+        
+        Parameters:
+        ----------
+        save_path : str
+            Path to save the model
+        """
+        if self.compiled_model is None:
+            raise RuntimeError("No model has been optimized yet")
+        
+        try:
+            # Save based on optimization mode
+            if self.optimize_mode == 'jit':
+                torch.jit.save(self.compiled_model, save_path)
+            else:
+                torch.save(self.compiled_model.state_dict(), save_path)
+            print(f"✓ Model saved to: {save_path}")
+        except Exception as e:
+            print(f"✗ Failed to save model: {e}")
+    
+    def load_optimized_model(self, load_path: str, model_template: nn.Module) -> nn.Module:
+        """
+        Load optimized model from file.
+        
+        Parameters:
+        ----------
+        load_path : str
+            Path to the saved model
+        model_template : nn.Module
+            Model template for loading state dict
+            
+        Returns:
+        -------
+        nn.Module
+            Loaded model
+        """
+        try:
+            if self.optimize_mode == 'jit':
+                self.compiled_model = torch.jit.load(load_path, map_location=self.device)
+            else:
+                model_template.load_state_dict(torch.load(load_path, map_location=self.device, weights_only=False))
+                self.compiled_model = model_template.eval()
+            
+            print(f"✓ Model loaded from: {load_path}")
+            return self.compiled_model
+        except (pickle.UnpicklingError, RuntimeError, EOFError) as e:
+            error_details = f"""
+CHECKPOINT FILE CORRUPTED
+
+Error: {str(e)}
+
+The checkpoint file appears to be corrupted or was not downloaded correctly.
+File: {load_path}
+
+Common causes:
+  - File is an HTML page (wrong download URL, e.g., HuggingFace /blob/ instead of /resolve/)
+  - Incomplete or interrupted download
+  - Network issues during download
+  - File system corruption
+
+Solution:
+  1. Delete the corrupted checkpoint file:
+     {load_path}
+  2. Re-run the application - it will automatically re-download the model
+  3. If the problem persists, check that your model URL uses /resolve/ not /blob/
+     Example: https://huggingface.co/user/repo/resolve/main/model.ckpt
+"""
+            print(error_details)
+            raise
+        except Exception as e:
+            print(f"✗ Failed to load model: {e}")
+            raise
+    
+    def __call__(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Run inference with optimized model.
+        
+        Parameters:
+        ----------
+        x : torch.Tensor
+            Input tensor
+            
+        Returns:
+        -------
+        torch.Tensor
+            Model output
+        """
+        if self.compiled_model is None:
+            raise RuntimeError("No model has been optimized yet")
+        
+        # Apply memory format if needed (only for 4D tensors - images)
+        # Audio models typically use 3D tensors, so we silently skip channels_last for them
+        if self.optimize_mode == 'channels_last' and x.dim() == 4:
+            x = x.to(memory_format=torch.channels_last)
+        
+        # Run inference with AMP if enabled
+        try:
+            if self.use_amp and self.device.startswith('cuda'):
+                with torch.cuda.amp.autocast():
+                    with torch.no_grad():
+                        return self.compiled_model(x)
+            else:
+                with torch.no_grad():
+                    return self.compiled_model(x)
+        except Exception as e:
+            # Fallback to non-compiled model if torch.compile fails at runtime
+            # This can happen with rotary embeddings that mutate class variables
+            if self.optimize_mode == 'compile' and self.model is not None:
+                print(f"  ⚠️ torch.compile runtime error: {type(e).__name__}")
+                print(f"  🔄 Falling back to non-compiled model...")
+                self.compiled_model = self.model
+                self.optimize_mode = 'fallback'
+                # Retry with non-compiled model
+                if self.use_amp and self.device.startswith('cuda'):
+                    with torch.cuda.amp.autocast():
+                        with torch.no_grad():
+                            return self.compiled_model(x)
+                else:
+                    with torch.no_grad():
+                        return self.compiled_model(x)
+            else:
+                raise
+
+
+class PyTorchOptimizer:
+    """
+    Helper class for various PyTorch optimization techniques.
+    """
+    
+    @staticmethod
+    def enable_cudnn_benchmark():
+        """Enable cuDNN benchmark mode."""
+        if torch.cuda.is_available():
+            torch.backends.cudnn.benchmark = True
+            torch.backends.cudnn.deterministic = False
+            print("cuDNN benchmark enabled")
+    
+    @staticmethod
+    def enable_cudnn_deterministic():
+        """Enable cuDNN deterministic mode for reproducible results."""
+        if torch.cuda.is_available():
+            torch.backends.cudnn.deterministic = True
+            torch.backends.cudnn.benchmark = False
+            print("✓ cuDNN deterministic mode enabled")
+    
+    @staticmethod
+    def enable_tf32():
+        """Enable TF32 for Ampere GPUs (RTX 30xx+)."""
+        if torch.cuda.is_available():
+            torch.backends.cuda.matmul.allow_tf32 = True
+            torch.backends.cudnn.allow_tf32 = True
+            # Also enable for float32 matmul precision
+            torch.set_float32_matmul_precision('high')  # or 'highest' for max speed
+            print("TF32 enabled")
+    
+    @staticmethod
+    def set_num_threads(num_threads: int):
+        """Set number of threads for CPU inference."""
+        torch.set_num_threads(num_threads)
+        print(f"✓ Number of threads set to: {num_threads}")
+    
+    @staticmethod
+    def optimize_for_inference(model: nn.Module) -> nn.Module:
+        """
+        Apply ULTRA optimization for inference.
+        
+        Parameters:
+        ----------
+        model : nn.Module
+            Model to optimize
+            
+        Returns:
+        -------
+        nn.Module
+            ULTRA-optimized model
+        """
+        model.eval()
+        torch.set_grad_enabled(False)
+        
+        # Disable gradient computation for all parameters
+        for param in model.parameters():
+            param.requires_grad = False
+        
+        # Fuse operations if possible
+        try:
+            # Try to fuse batch norm
+            model = torch.quantization.fuse_modules(model, inplace=True)
+            print("Batch norm fused")
+        except:
+            pass
+        
+        try:
+            # Try to fuse conv-bn if available
+            if hasattr(torch.nn.utils, 'fusion'):
+                model = torch.nn.utils.fusion.fuse_conv_bn_eval(model)
+                print("Conv-BN fused")
+        except:
+            pass
+        
+        return model
+
+
+def benchmark_pytorch_optimizations(
+    model: nn.Module,
+    input_shape: Tuple[int, ...],
+    device: str = 'cuda:0',
+    num_iterations: int = 100,
+    warmup_iterations: int = 10
+) -> Dict[str, float]:
+    """
+    Benchmark different PyTorch optimization techniques.
+    
+    Parameters:
+    ----------
+    model : nn.Module
+        Model to benchmark
+    input_shape : Tuple[int, ...]
+        Input tensor shape
+    device : str
+        Device to use
+    num_iterations : int
+        Number of benchmark iterations
+    warmup_iterations : int
+        Number of warmup iterations
+        
+    Returns:
+    -------
+    Dict[str, float]
+        Benchmark results with average inference times
+    """
+    results = {}
+    dummy_input = torch.randn(*input_shape).to(device)
+    
+    optimization_modes = ['default', 'compile', 'channels_last']
+    
+    for mode in optimization_modes:
+        print(f"\n{'='*60}")
+        print(f"Benchmarking: {mode}")
+        print('='*60)
+        
+        try:
+            backend = PyTorchBackend(device=device, optimize_mode=mode)
+            
+            # Optimize model
+            if mode == 'compile':
+                optimized_model = backend.optimize_model(model, use_amp=True)
+            else:
+                optimized_model = backend.optimize_model(
+                    model, 
+                    example_input=dummy_input,
+                    use_amp=True,
+                    use_channels_last=(mode == 'channels_last')
+                )
+            
+            # Warmup
+            for _ in range(warmup_iterations):
+                _ = backend(dummy_input)
+            
+            # Benchmark
+            if device.startswith('cuda'):
+                torch.cuda.synchronize()
+            
+            start = time.time()
+            for _ in range(num_iterations):
+                _ = backend(dummy_input)
+            
+            if device.startswith('cuda'):
+                torch.cuda.synchronize()
+            
+            elapsed = (time.time() - start) / num_iterations
+            results[mode] = elapsed * 1000  # Convert to ms
+            
+            print(f"  Average time: {results[mode]:.2f} ms")
+            
+        except Exception as e:
+            print(f"  Failed: {e}")
+            results[mode] = None
+    
+    return results
+
+
+def create_inference_session(
+    model: nn.Module,
+    device: str = 'cuda:0',
+    optimize_mode: str = 'default',
+    enable_amp: bool = True,
+    enable_tf32: bool = True,
+    enable_cudnn_benchmark: bool = True
+) -> PyTorchBackend:
+    """
+    Create an optimized inference session.
+    
+    Parameters:
+    ----------
+    model : nn.Module
+        Model to use for inference
+    device : str
+        Device to use
+    optimize_mode : str
+        Optimization mode
+    enable_amp : bool
+        Enable automatic mixed precision
+    enable_tf32 : bool
+        Enable TF32 (for Ampere GPUs)
+    enable_cudnn_benchmark : bool
+        Enable cuDNN benchmark
+        
+    Returns:
+    -------
+    PyTorchBackend
+        Configured inference session
+    """
+    # Apply global optimizations
+    optimizer = PyTorchOptimizer()
+    
+    if enable_cudnn_benchmark:
+        optimizer.enable_cudnn_benchmark()
+    
+    if enable_tf32 and device.startswith('cuda'):
+        optimizer.enable_tf32()
+    
+    # Create backend
+    backend = PyTorchBackend(device=device, optimize_mode=optimize_mode)
+    backend.optimize_model(model, use_amp=enable_amp)
+    
+    return backend
+
+
+def convert_model_to_onnx(
+    model: nn.Module,
+    input_shape: Tuple[int, ...],
+    output_path: str,
+    opset_version: int = 14
+):
+    """
+    Convert PyTorch model to ONNX format.
+    
+    Parameters:
+    ----------
+    model : nn.Module
+        Model to convert
+    input_shape : Tuple[int, ...]
+        Input tensor shape
+    output_path : str
+        Path to save ONNX model
+    opset_version : int
+        ONNX opset version
+    """
+    try:
+        import onnx
+        
+        model.eval()
+        dummy_input = torch.randn(*input_shape)
+        
+        print(f"Converting model to ONNX (opset {opset_version})...")
+        torch.onnx.export(
+            model,
+            dummy_input,
+            output_path,
+            export_params=True,
+            opset_version=opset_version,
+            do_constant_folding=True,
+            input_names=['input'],
+            output_names=['output'],
+            dynamic_axes={
+                'input': {0: 'batch_size'},
+                'output': {0: 'batch_size'}
+            }
+        )
+        
+        # Verify ONNX model
+        onnx_model = onnx.load(output_path)
+        onnx.checker.check_model(onnx_model)
+        
+        print(f"✓ ONNX model saved to: {output_path}")
+        
+    except ImportError:
+        print("✗ ONNX not available. Install with: pip install onnx")
+    except Exception as e:
+        print(f"✗ ONNX conversion failed: {e}")
+
+
+def get_model_info(model: nn.Module) -> Dict[str, Any]:
+    """
+    Get information about a PyTorch model.
+    
+    Parameters:
+    ----------
+    model : nn.Module
+        Model to analyze
+        
+    Returns:
+    -------
+    Dict[str, Any]
+        Model information
+    """
+    total_params = sum(p.numel() for p in model.parameters())
+    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
+    
+    # Estimate model size
+    param_size = sum(p.nelement() * p.element_size() for p in model.parameters())
+    buffer_size = sum(b.nelement() * b.element_size() for b in model.buffers())
+    size_mb = (param_size + buffer_size) / (1024 ** 2)
+    
+    return {
+        'total_parameters': total_params,
+        'trainable_parameters': trainable_params,
+        'model_size_mb': size_mb,
+        'device': next(model.parameters()).device,
+        'dtype': next(model.parameters()).dtype
+    }

requirements.txt

@@ -5,6 +5,7 @@ numpy
 pandas
 scipy
 torch
+torchaudio
 torchvision
 torchmetrics
 ml_collections
@@ -17,7 +18,7 @@ demucs
 audiomentations
 torch_audiomentations
 httpx
-gradio
+gradio>=4.0
 google-api-python-client
 pytube
 tqdm
@@ -37,4 +38,10 @@ pyloudnorm
 gdown
 validators
 requests
-spaces
+sympy
+typing_extensions
+packaging
+opencv-python-headless
+pypresence
+mpmath
+pillow

utils.py

@@ -13,13 +13,39 @@ from omegaconf import OmegaConf
 from tqdm.auto import tqdm
 from typing import Dict, List, Tuple, Any, Union
 import loralib as lora
-import gc  # For garbage collection
-import logging  # Hata takibi için
 
-# Log ayarları
-logging.basicConfig(level=logging.INFO, filename='utils.log', format='%(asctime)s - %(message)s')
+# Fix matplotlib backend for isolated Python environments
+try:
+    import matplotlib
+    matplotlib.use('Agg')  # Non-interactive backend
+    import matplotlib.pyplot as plt
+except ImportError:
+    plt = None
+
 
 def load_config(model_type: str, config_path: str) -> Union[ConfigDict, OmegaConf]:
+    """
+    Load the configuration from the specified path based on the model type.
+
+    Parameters:
+    ----------
+    model_type : str
+        The type of model to load (e.g., 'htdemucs', 'mdx23c', etc.).
+    config_path : str
+        The path to the YAML or OmegaConf configuration file.
+
+    Returns:
+    -------
+    config : Any
+        The loaded configuration, which can be in different formats (e.g., OmegaConf or ConfigDict).
+
+    Raises:
+    ------
+    FileNotFoundError:
+        If the configuration file at `config_path` is not found.
+    ValueError:
+        If there is an error loading the configuration file.
+    """
     try:
         with open(config_path, 'r') as f:
             if model_type == 'htdemucs':
@@ -32,6 +58,7 @@ def load_config(model_type: str, config_path: str) -> Union[ConfigDict, OmegaCon
     except Exception as e:
         raise ValueError(f"Error loading configuration: {e}")
 
+
 def get_model_from_config(model_type: str, config_path: str) -> Tuple:
     """
     Load the model specified by the model type and configuration file.
@@ -101,97 +128,251 @@ def get_model_from_config(model_type: str, config_path: str) -> Tuple:
     elif model_type == 'experimental_mdx23c_stht':
         from models.mdx23c_tfc_tdf_v3_with_STHT import TFC_TDF_net
         model = TFC_TDF_net(config)
+    elif model_type == 'mel_band_roformer_experimental':
+        from models.bs_roformer.mel_band_roformer_experimental import MelBandRoformer
+        model = MelBandRoformer(**dict(config.model))
+    elif model_type == 'bs_roformer_experimental':
+        from models.bs_roformer.bs_roformer_experimental import BSRoformer
+        model = BSRoformer(**dict(config.model))
+    elif model_type == 'bs_roformer_custom':
+        from models.bs_roformer.bs_roformer_custom.bs_roformer import BSRoformer
+        model = BSRoformer(**dict(config.model))
+    elif model_type == 'scnet_tran':
+        from models.scnet.scnet_tran import SCNet_Tran
+        model = SCNet_Tran(**config.model)
+    elif model_type == 'scnet_masked':
+        from models.scnet.scnet_masked import SCNet
+        model = SCNet(**config.model)
+    elif model_type == 'conformer':
+        from models.conformer_model import ConformerMSS, NeuralModel
+        model = ConformerMSS(
+            core=NeuralModel(**config.model),
+            n_fft=config.stft.n_fft,
+            hop_length=config.stft.hop_length,
+            win_length=getattr(config.stft, 'win_length', config.stft.n_fft),
+            center=config.stft.center
+        )
+    elif model_type == 'mel_band_conformer':
+        from models.mel_band_conformer import MelBandConformer
+        model = MelBandConformer(**config.model)
     else:
         raise ValueError(f"Unknown model type: {model_type}")
 
     return model, config
 
+
 def read_audio_transposed(path: str, instr: str = None, skip_err: bool = False) -> Tuple[np.ndarray, int]:
+    """
+    Reads an audio file, ensuring mono audio is converted to two-dimensional format,
+    and transposes the data to have channels as the first dimension.
+    Parameters
+    ----------
+    path : str
+        Path to the audio file.
+    skip_err: bool
+        If true, not raise errors
+    instr:
+        name of instument
+    Returns
+    -------
+    Tuple[np.ndarray, int]
+        A tuple containing:
+        - Transposed audio data as a NumPy array with shape (channels, length).
+          For mono audio, the shape will be (1, length).
+        - Sampling rate (int), e.g., 44100.
+    """
+
     try:
         mix, sr = sf.read(path)
-        if len(mix.shape) == 1:  # Mono audio
-            mix = np.expand_dims(mix, axis=-1)
-        return mix.T, sr
     except Exception as e:
         if skip_err:
             print(f"No stem {instr}: skip!")
             return None, None
-        raise RuntimeError(f"Error reading the file at {path}: {e}")
+        else:
+            raise RuntimeError(f"Error reading the file at {path}: {e}")
+    else:
+        if len(mix.shape) == 1:  # For mono audio
+            mix = np.expand_dims(mix, axis=-1)
+        return mix.T, sr
+
+
+def normalize_audio(audio: np.ndarray) -> tuple[np.ndarray, Dict[str, float]]:
+    """
+    Normalize an audio signal by subtracting the mean and dividing by the standard deviation.
+
+    Parameters:
+    ----------
+    audio : np.ndarray
+        Input audio array with shape (channels, time) or (time,).
+
+    Returns:
+    -------
+    tuple[np.ndarray, dict[str, float]]
+        - Normalized audio array with the same shape as the input.
+        - Dictionary containing the mean and standard deviation of the original audio.
+    """
 
-def normalize_audio(audio: np.ndarray) -> Tuple[np.ndarray, Dict[str, float]]:
     mono = audio.mean(0)
     mean, std = mono.mean(), mono.std()
-    return (audio - mean) / (std + 1e-8), {"mean": mean, "std": std}
+    return (audio - mean) / std, {"mean": mean, "std": std}
+
 
 def denormalize_audio(audio: np.ndarray, norm_params: Dict[str, float]) -> np.ndarray:
+    """
+    Denormalize an audio signal by reversing the normalization process (multiplying by the standard deviation
+    and adding the mean).
+
+    Parameters:
+    ----------
+    audio : np.ndarray
+        Normalized audio array to be denormalized.
+    norm_params : dict[str, float]
+        Dictionary containing the 'mean' and 'std' values used for normalization.
+
+    Returns:
+    -------
+    np.ndarray
+        Denormalized audio array with the same shape as the input.
+    """
+
     return audio * norm_params["std"] + norm_params["mean"]
 
+
 def apply_tta(
-    config,
-    model: nn.Module,
-    mix: torch.Tensor,
-    waveforms_orig: Dict[str, torch.Tensor],
-    device: str,
-    model_type: str,
-    progress=None  # Gradio progress nesnesi
+        config,
+        model: torch.nn.Module,
+        mix: torch.Tensor,
+        waveforms_orig: Dict[str, torch.Tensor],
+        device: torch.device,
+        model_type: str
 ) -> Dict[str, torch.Tensor]:
-    track_proc_list = [mix[::-1].clone(), -mix.clone()]
-    total_steps = len(track_proc_list)
-    processed_steps = 0
+    """
+    Apply Test-Time Augmentation (TTA) for source separation.
+
+    This function processes the input mixture with test-time augmentations, including
+    channel inversion and polarity inversion, to enhance the separation results. The
+    results from all augmentations are averaged to produce the final output.
 
+    Parameters:
+    ----------
+    config : Any
+        Configuration object containing model and processing parameters.
+    model : torch.nn.Module
+        The trained model used for source separation.
+    mix : torch.Tensor
+        The mixed audio tensor with shape (channels, time).
+    waveforms_orig : Dict[str, torch.Tensor]
+        Dictionary of original separated waveforms (before TTA) for each instrument.
+    device : torch.device
+        Device (CPU or CUDA) on which the model will be executed.
+    model_type : str
+        Type of the model being used (e.g., "demucs", "custom_model").
+
+    Returns:
+    -------
+    Dict[str, torch.Tensor]
+        Updated dictionary of separated waveforms after applying TTA.
+    """
+    # Create augmentations: channel inversion and polarity inversion
+    track_proc_list = [mix[::-1].copy(), -1.0 * mix.copy()]
+
+    # Process each augmented mixture
     for i, augmented_mix in enumerate(track_proc_list):
-        # TTA adımı için ilerleme güncellemesi
-        processed_steps += 1
-        progress_value = round((processed_steps / total_steps) * 50)  # TTA için 0-50% aralığı
-        if progress is not None and callable(getattr(progress, '__call__', None)):
-            progress(progress_value / 100, desc=f"Applying TTA step {processed_steps}/{total_steps}")
-            update_progress_html(f"Applying TTA step {processed_steps}/{total_steps}", progress_value)
-
-        waveforms = demix(config, model, augmented_mix, device, model_type=model_type, pbar=False, progress=progress)
+        waveforms = demix(config, model, augmented_mix, device, model_type=model_type)
         for el in waveforms:
             if i == 0:
-                waveforms_orig[el] += waveforms[el][::-1].clone()
+                waveforms_orig[el] += waveforms[el][::-1].copy()
             else:
                 waveforms_orig[el] -= waveforms[el]
-        del waveforms, augmented_mix
-        gc.collect()
-        if device.startswith('cuda'):
-            torch.cuda.empty_cache()
 
+    # Average the results across augmentations
     for el in waveforms_orig:
-        waveforms_orig[el] /= (len(track_proc_list) + 1)
-
-    # TTA tamamlandı
-    if progress is not None and callable(getattr(progress, '__call__', None)):
-        progress(0.5, desc="TTA completed")
-        update_progress_html("TTA completed", 50)
+        waveforms_orig[el] /= len(track_proc_list) + 1
 
     return waveforms_orig
 
+
 def _getWindowingArray(window_size: int, fade_size: int) -> torch.Tensor:
+    """
+    Generate a windowing array with a linear fade-in at the beginning and a fade-out at the end.
+
+    This function creates a window of size `window_size` where the first `fade_size` elements
+    linearly increase from 0 to 1 (fade-in) and the last `fade_size` elements linearly decrease
+    from 1 to 0 (fade-out). The middle part of the window is filled with ones.
+
+    Parameters:
+    ----------
+    window_size : int
+        The total size of the window.
+    fade_size : int
+        The size of the fade-in and fade-out regions.
+
+    Returns:
+    -------
+    torch.Tensor
+        A tensor of shape (window_size,) containing the generated windowing array.
+
+    Example:
+    -------
+    If `window_size=10` and `fade_size=3`, the output will be:
+    tensor([0.0000, 0.5000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 1.0000, 0.5000, 0.0000])
+    """
+
     fadein = torch.linspace(0, 1, fade_size)
     fadeout = torch.linspace(1, 0, fade_size)
+
     window = torch.ones(window_size)
     window[-fade_size:] = fadeout
     window[:fade_size] = fadein
     return window
 
+
 def demix(
-    config: ConfigDict,
-    model: nn.Module,
-    mix: torch.Tensor,
-    device: str,
-    model_type: str,
-    pbar: bool = False,
-    progress=None  # Gradio progress nesnesi
-) -> Dict[str, np.ndarray]:
-    logging.info(f"Starting demix for model_type: {model_type}, chunk_size: {config.audio.chunk_size}")
-    
-    # CPU'da FP16 ile başla
-    mix = torch.tensor(mix, dtype=torch.float16, device='cpu')
-    mode = 'demucs' if model_type == 'htdemucs' else 'generic'
+        config: ConfigDict,
+        model: torch.nn.Module,
+        mix: torch.Tensor,
+        device: torch.device,
+        model_type: str,
+        pbar: bool = False
+) -> Tuple[List[Dict[str, np.ndarray]], np.ndarray]:
+    """
+    Unified function for audio source separation with support for multiple processing modes.
+
+    This function separates audio into its constituent sources using either a generic custom logic
+    or a Demucs-specific logic. It supports batch processing and overlapping window-based chunking
+    for efficient and artifact-free separation.
 
-    # İşlem parametreleri
+    Parameters:
+    ----------
+    config : ConfigDict
+        Configuration object containing audio and inference settings.
+    model : torch.nn.Module
+        The trained model used for audio source separation.
+    mix : torch.Tensor
+        Input audio tensor with shape (channels, time).
+    device : torch.device
+        The computation device (CPU or CUDA).
+    model_type : str, optional
+        Processing mode:
+            - "demucs" for logic specific to the Demucs model.
+        Default is "generic".
+    pbar : bool, optional
+        If True, displays a progress bar during chunk processing. Default is False.
+
+    Returns:
+    -------
+    Union[Dict[str, np.ndarray], np.ndarray]
+        - A dictionary mapping target instruments to separated audio sources if multiple instruments are present.
+        - A numpy array of the separated source if only one instrument is present.
+    """
+
+    mix = torch.tensor(mix, dtype=torch.float32)
+
+    if model_type == 'htdemucs':
+        mode = 'demucs'
+    else:
+        mode = 'generic'
+    # Define processing parameters based on the mode
     if mode == 'demucs':
         chunk_size = config.training.samplerate * config.training.segment
         num_instruments = len(config.training.instruments)
@@ -201,146 +382,272 @@ def demix(
         chunk_size = config.audio.chunk_size
         num_instruments = len(prefer_target_instrument(config))
         num_overlap = config.inference.num_overlap
+
         fade_size = chunk_size // 10
         step = chunk_size // num_overlap
         border = chunk_size - step
         length_init = mix.shape[-1]
-        windowing_array = _getWindowingArray(chunk_size, fade_size).to('cpu', dtype=torch.float16)
+        windowing_array = _getWindowingArray(chunk_size, fade_size)
+        # Add padding for generic mode to handle edge artifacts
         if length_init > 2 * border and border > 0:
             mix = nn.functional.pad(mix, (border, border), mode="reflect")
 
-    batch_size = getattr(config.inference, 'batch_size', 1)  # Düşük bellek için varsayılan 1
-
-    # Modeli cihaza taşı (ZeroGPU için cuda:0)
-    model = model.to(device)
-    model.eval()
+    batch_size = config.inference.batch_size
 
-    # Toplam chunk sayısını hesapla
-    total_chunks = (mix.shape[1] + step - 1) // step
-    processed_chunks = 0
+    use_amp = getattr(config.training, 'use_amp', True)
 
-    with torch.no_grad():  # Çıkarım için gradyan yok
-        with torch.cuda.amp.autocast(enabled=device.startswith('cuda'), dtype=torch.float16):
+    with torch.cuda.amp.autocast(enabled=use_amp):
+        with torch.inference_mode():
+            # Initialize result and counter tensors
             req_shape = (num_instruments,) + mix.shape
-            result = torch.zeros(req_shape, dtype=torch.float16, device='cpu')
-            counter = torch.zeros(req_shape, dtype=torch.float16, device='cpu')
+            result = torch.zeros(req_shape, dtype=torch.float32)
+            counter = torch.zeros(req_shape, dtype=torch.float32)
 
             i = 0
             batch_data = []
             batch_locations = []
-            start_time = time.time()
+            progress_bar = tqdm(
+                total=mix.shape[1], desc="Processing audio chunks", leave=False
+            ) if pbar else None
 
             while i < mix.shape[1]:
-                part = mix[:, i:i + chunk_size]
+                # Extract chunk and apply padding if necessary
+                part = mix[:, i:i + chunk_size].to(device)
                 chunk_len = part.shape[-1]
-                pad_mode = "reflect" if mode == "generic" and chunk_len > chunk_size // 2 else "constant"
+                if mode == "generic" and chunk_len > chunk_size // 2:
+                    pad_mode = "reflect"
+                else:
+                    pad_mode = "constant"
                 part = nn.functional.pad(part, (0, chunk_size - chunk_len), mode=pad_mode, value=0)
 
                 batch_data.append(part)
                 batch_locations.append((i, chunk_len))
                 i += step
 
+                # Process batch if it's full or the end is reached
                 if len(batch_data) >= batch_size or i >= mix.shape[1]:
-                    # Veriyi GPU'ya taşı
-                    arr = torch.stack(batch_data, dim=0).to(device, non_blocking=True)
-                    x = model(arr)  # Model çıkarımı GPU'da
-
-                    # Sonuçları hemen CPU'ya taşı
-                    x = x.cpu()
+                    arr = torch.stack(batch_data, dim=0)
+                    x = model(arr)
 
                     if mode == "generic":
-                        window = windowing_array.clone()
-                        if i - step == 0:
+                        window = windowing_array.clone() # using clone() fixes the clicks at chunk edges when using batch_size=1
+                        if i - step == 0:  # First audio chunk, no fadein
                             window[:fade_size] = 1
-                        elif i >= mix.shape[1]:
+                        elif i >= mix.shape[1]:  # Last audio chunk, no fadeout
                             window[-fade_size:] = 1
 
                     for j, (start, seg_len) in enumerate(batch_locations):
                         if mode == "generic":
-                            result[..., start:start + seg_len] += (x[j, ..., :seg_len] * window[..., :seg_len])
+                            result[..., start:start + seg_len] += x[j, ..., :seg_len].cpu() * window[..., :seg_len]
                             counter[..., start:start + seg_len] += window[..., :seg_len]
                         else:
-                            result[..., start:start + seg_len] += x[j, ..., :seg_len]
+                            result[..., start:start + seg_len] += x[j, ..., :seg_len].cpu()
                             counter[..., start:start + seg_len] += 1.0
 
-                    # İlerleme güncellemesi
-                    processed_chunks += len(batch_data)
-                    progress_value = min(round((processed_chunks / total_chunks) * 100), 100)  # %1 hassasiyet
-                    if progress is not None and callable(getattr(progress, '__call__', None)):
-                        progress(progress_value / 100, desc=f"Processing chunk {processed_chunks}/{total_chunks}")
-                        update_progress_html(f"Processing chunk {processed_chunks}/{total_chunks}", progress_value)
-
-                    del arr, x
                     batch_data.clear()
                     batch_locations.clear()
-                    gc.collect()
-                    if device.startswith('cuda'):
-                        torch.cuda.empty_cache()
-                        logging.info("Cleared CUDA cache")
 
-            elapsed_time = time.time() - start_time
-            logging.info(f"Demix completed in {elapsed_time:.2f} seconds")
+                if progress_bar:
+                    progress_bar.update(step)
+
+            if progress_bar:
+                progress_bar.close()
 
-            estimated_sources = result / (counter + 1e-8)
-            estimated_sources = estimated_sources.numpy().astype(np.float32)
+            # Compute final estimated sources
+            estimated_sources = result / counter
+            estimated_sources = estimated_sources.cpu().numpy()
             np.nan_to_num(estimated_sources, copy=False, nan=0.0)
 
-            if mode == "generic" and length_init > 2 * border and border > 0:
-                estimated_sources = estimated_sources[..., border:-border]
+            # Remove padding for generic mode
+            if mode == "generic":
+                if length_init > 2 * border and border > 0:
+                    estimated_sources = estimated_sources[..., border:-border]
+
+    # Return the result as a dictionary or a single array
+    if mode == "demucs":
+        instruments = config.training.instruments
+    else:
+        instruments = prefer_target_instrument(config)
 
-    instruments = config.training.instruments if mode == "demucs" else prefer_target_instrument(config)
     ret_data = {k: v for k, v in zip(instruments, estimated_sources)}
-    logging.info("Demix completed successfully")
 
-    # Son ilerleme güncellemesi
-    if progress is not None and callable(getattr(progress, '__call__', None)):
-        progress(1.0, desc="Demix completed")
-        update_progress_html("Demix completed", 100)
+    if mode == "demucs" and num_instruments <= 1:
+        return estimated_sources
+    else:
+        return ret_data
 
-    return ret_data
 
 def prefer_target_instrument(config: ConfigDict) -> List[str]:
-    return [config.training.target_instrument] if getattr(config.training, 'target_instrument', None) else config.training.instruments
+    """
+        Return the list of target instruments based on the configuration.
+        If a specific target instrument is specified in the configuration,
+        it returns a list with that instrument. Otherwise, it returns the list of instruments.
+
+        Parameters:
+        ----------
+        config : ConfigDict
+            Configuration object containing the list of instruments or the target instrument.
+
+        Returns:
+        -------
+        List[str]
+            A list of target instruments.
+        """
+    if getattr(config.training, 'target_instrument', None):
+        return [config.training.target_instrument]
+    else:
+        return config.training.instruments
+
+
+def load_not_compatible_weights(model: torch.nn.Module, weights: str, verbose: bool = False) -> None:
+    """
+    Load weights into a model, handling mismatched shapes and dimensions.
+
+    Args:
+        model: PyTorch model into which the weights will be loaded.
+        weights: Path to the weights file.
+        verbose: If True, prints detailed information about matching and mismatched layers.
+    """
 
-def load_not_compatible_weights(model: nn.Module, weights: str, verbose: bool = False) -> None:
     new_model = model.state_dict()
-    old_model = torch.load(weights, map_location='cpu')
+    old_model = torch.load(weights, weights_only=False)
     if 'state' in old_model:
+        # Fix for htdemucs weights loading
         old_model = old_model['state']
     if 'state_dict' in old_model:
+        # Fix for apollo weights loading
         old_model = old_model['state_dict']
+
     for el in new_model:
-        if el in old_model and new_model[el].shape == old_model[el].shape:
-            new_model[el] = old_model[el]
-    model.load_state_dict(new_model)
+        if el in old_model:
+            if verbose:
+                print(f'Match found for {el}!')
+            if new_model[el].shape == old_model[el].shape:
+                if verbose:
+                    print('Action: Just copy weights!')
+                new_model[el] = old_model[el]
+            else:
+                if len(new_model[el].shape) != len(old_model[el].shape):
+                    if verbose:
+                        print('Action: Different dimension! Too lazy to write the code... Skip it')
+                else:
+                    if verbose:
+                        print(f'Shape is different: {tuple(new_model[el].shape)} != {tuple(old_model[el].shape)}')
+                    ln = len(new_model[el].shape)
+                    max_shape = []
+                    slices_old = []
+                    slices_new = []
+                    for i in range(ln):
+                        max_shape.append(max(new_model[el].shape[i], old_model[el].shape[i]))
+                        slices_old.append(slice(0, old_model[el].shape[i]))
+                        slices_new.append(slice(0, new_model[el].shape[i]))
+                    # print(max_shape)
+                    # print(slices_old, slices_new)
+                    slices_old = tuple(slices_old)
+                    slices_new = tuple(slices_new)
+                    max_matrix = np.zeros(max_shape, dtype=np.float32)
+                    for i in range(ln):
+                        max_matrix[slices_old] = old_model[el].cpu().numpy()
+                    max_matrix = torch.from_numpy(max_matrix)
+                    new_model[el] = max_matrix[slices_new]
+        else:
+            if verbose:
+                print(f'Match not found for {el}!')
+    model.load_state_dict(
+        new_model
+    )
 
-def load_lora_weights(model: nn.Module, lora_path: str, device: str = 'cpu') -> None:
-    lora_state_dict = torch.load(lora_path, map_location=device)
+
+def load_lora_weights(model: torch.nn.Module, lora_path: str, device: str = 'cpu') -> None:
+    """
+    Load LoRA weights into a model.
+    This function updates the given model with LoRA-specific weights from the specified checkpoint file.
+    It does not require the checkpoint to match the model's full state dictionary, as only LoRA layers are updated.
+
+    Parameters:
+    ----------
+    model : Module
+        The PyTorch model into which the LoRA weights will be loaded.
+    lora_path : str
+        Path to the LoRA checkpoint file.
+    device : str, optional
+        The device to load the weights onto, by default 'cpu'. Common values are 'cpu' or 'cuda'.
+
+    Returns:
+    -------
+    None
+        The model is updated in place.
+    """
+    lora_state_dict = torch.load(lora_path, map_location=device, weights_only=False)
     model.load_state_dict(lora_state_dict, strict=False)
 
-def load_start_checkpoint(args: argparse.Namespace, model: nn.Module, type_='train') -> None:
+
+def load_start_checkpoint(args: argparse.Namespace, model: torch.nn.Module, type_='train') -> None:
+    """
+    Load the starting checkpoint for a model.
+
+    Args:
+        args: Parsed command-line arguments containing the checkpoint path.
+        model: PyTorch model to load the checkpoint into.
+        type_: how to load weights - for train we can load not fully compatible weights
+    """
+
     print(f'Start from checkpoint: {args.start_check_point}')
-    device = 'cpu'
-    state_dict = torch.load(args.start_check_point, map_location=device, weights_only=True)
-    if args.model_type in ['htdemucs', 'apollo'] and isinstance(state_dict, dict):
-        state_dict = state_dict.get('state', state_dict.get('state_dict', state_dict))
-    model.load_state_dict(state_dict)
-    if args.lora_checkpoint:
-        print(f"Loading LoRA weights from: {args.lora_checkpoint}")
-        load_lora_weights(model, args.lora_checkpoint, device)
+    if type_ in ['train']:
+        if 1:
+            load_not_compatible_weights(model, args.start_check_point, verbose=False)
+        else:
+            model.load_state_dict(torch.load(args.start_check_point, weights_only=False))
+    else:
+        device='cpu'
+        if args.model_type in ['htdemucs', 'apollo']:
+            state_dict = torch.load(args.start_check_point, map_location=device, weights_only=False)
+            # Fix for htdemucs pretrained models
+            if 'state' in state_dict:
+                state_dict = state_dict['state']
+            # Fix for apollo pretrained models
+            if 'state_dict' in state_dict:
+                state_dict = state_dict['state_dict']
+        else:
+            state_dict = torch.load(args.start_check_point, map_location=device, weights_only=True)
+        model.load_state_dict(state_dict)
+
 
 def bind_lora_to_model(config: Dict[str, Any], model: nn.Module) -> nn.Module:
+    """
+    Replaces specific layers in the model with LoRA-extended versions.
+
+    Parameters:
+    ----------
+    config : Dict[str, Any]
+        Configuration containing parameters for LoRA. It should include a 'lora' key with parameters for `MergedLinear`.
+    model : nn.Module
+        The original model in which the layers will be replaced.
+
+    Returns:
+    -------
+    nn.Module
+        The modified model with the replaced layers.
+    """
+
     if 'lora' not in config:
         raise ValueError("Configuration must contain the 'lora' key with parameters for LoRA.")
-    replaced_layers = 0
+
+    replaced_layers = 0  # Counter for replaced layers
+
     for name, module in model.named_modules():
         hierarchy = name.split('.')
         layer_name = hierarchy[-1]
+
+        # Check if this is the target layer to replace (and layer_name == 'to_qkv')
         if isinstance(module, nn.Linear):
             try:
+                # Get the parent module
                 parent_module = model
                 for submodule_name in hierarchy[:-1]:
                     parent_module = getattr(parent_module, submodule_name)
+
+                # Replace the module with LoRA-enabled layer
                 setattr(
                     parent_module,
                     layer_name,
@@ -351,23 +658,41 @@ def bind_lora_to_model(config: Dict[str, Any], model: nn.Module) -> nn.Module:
                         **config['lora']
                     )
                 )
-                replaced_layers += 1
+                replaced_layers += 1  # Increment the counter
+
             except Exception as e:
                 print(f"Error replacing layer {name}: {e}")
-    print(f"Number of layers replaced with LoRA: {replaced_layers}")
+
+    if replaced_layers == 0:
+        print("Warning: No layers were replaced. Check the model structure and configuration.")
+    else:
+        print(f"Number of layers replaced with LoRA: {replaced_layers}")
+
     return model
 
+
 def draw_spectrogram(waveform, sample_rate, length, output_file):
+    if plt is None:
+        print("Warning: matplotlib not available, skipping spectrogram")
+        return
+    
     import librosa.display
+
+    # Cut only required part of spectorgram
     x = waveform[:int(length * sample_rate), :]
-    X = librosa.stft(x.mean(axis=-1))
-    Xdb = librosa.amplitude_to_db(np.abs(X), ref=np.max)
+    X = librosa.stft(x.mean(axis=-1))  # perform short-term fourier transform on mono signal
+    Xdb = librosa.amplitude_to_db(np.abs(X), ref=np.max)  # convert an amplitude spectrogram to dB-scaled spectrogram.
     fig, ax = plt.subplots()
+    # plt.figure(figsize=(30, 10))  # initialize the fig size
     img = librosa.display.specshow(
-        Xdb, cmap='plasma', sr=sample_rate, x_axis='time', y_axis='linear', ax=ax
+        Xdb,
+        cmap='plasma',
+        sr=sample_rate,
+        x_axis='time',
+        y_axis='linear',
+        ax=ax
     )
     ax.set(title='File: ' + os.path.basename(output_file))
     fig.colorbar(img, ax=ax, format="%+2.f dB")
-    if output_file:
+    if output_file is not None:
         plt.savefig(output_file)
-        plt.close()