Update app.py

app.py

@@ -4,6 +4,7 @@ os.environ["CUDA_VISIBLE_DEVICES"] = ""
 
 import re
 import gc
+import pickle
 import inspect
 import tempfile
 import traceback
@@ -29,34 +30,67 @@ def _torch_load_cpu(*args, **kwargs):
 torch.load = _torch_load_cpu
 
 # =====================================================================
-# CONFIG & PATH MANAGEMENT (DIOPTIMALKAN AGAR SUARA MULUS)
+# CONFIG & PATH MANAGEMENT (SWEET SPOT PARAMETERS)
 # =====================================================================
 MODEL_REPO = "grandhigh/Chatterbox-TTS-Indonesian"
 CHECKPOINT_FILENAME = "t3_cfg.safetensors"
 DEVICE = "cpu"
 
 MAX_TOTAL_CHARS = int(os.getenv("MAX_TOTAL_CHARS", "2400"))
-# DIUBAH: Ditingkatkan ke 450 agar model membaca kalimat utuh (intonasi jauh lebih natural)
-MAX_CHARS_PER_CHUNK = int(os.getenv("MAX_CHARS_PER_CHUNK", "450"))
-MAX_CHUNKS = int(os.getenv("MAX_CHUNKS", "8"))
-# DIUBAH: Set ke 0.0 agar tidak ada jeda kosong robotik antar potongan file audio
+# Set ke 280 agar CPU ringan melakukan sampling awal dan tidak stuck di 0%
+MAX_CHARS_PER_CHUNK = int(os.getenv("MAX_CHARS_PER_CHUNK", "280"))
+MAX_CHUNKS = int(os.getenv("MAX_CHUNKS", "10"))
 PAUSE_SECONDS = float(os.getenv("PAUSE_SECONDS", "0.0"))
 DOWNLOAD_TIMEOUT = int(os.getenv("DOWNLOAD_TIMEOUT", "90"))
 
-# Jalur untuk suara default bawaan sistem langsung di root (/)
 ROOT_DIR = Path(__file__).parent
 DEFAULT_SPEAKER_PATH = ROOT_DIR / "default_speaker.wav"
 
-# Global Cache untuk mengunci embedding karakteristik suara ke dalam RAM
-CACHED_EMBEDDINGS = {}
-
-from chatterbox.tts import ChatterboxTTS
-from huggingface_hub import hf_hub_download
-from safetensors.torch import load_file
+# FILE BOBOT PERMANEN: Lokasi penyimpanan biner karakteristik suara kustom kamu
+EMBEDDING_CACHE_PATH = ROOT_DIR / "speaker_embedding.pth"
 
+# Global RAM Cache untuk menyimpan koordinat suara siap pakai
+CACHED_EMBEDDINGS = {}
 _model = None
 _model_lock = Lock()
 
+
+def inisialisasi_bobot_default_permanen(model_instance):
+    """
+    Fungsi untuk melatih/mengekstrak karakteristik file default_speaker.wav
+    HANYA SEKALI saja saat startup. Hasilnya disimpan permanen ke disk (.pth).
+    """
+    global CACHED_EMBEDDINGS
+    try:
+        # Jika file biner .pth sudah ada, langsung muat ke memori RAM secara instan
+        if EMBEDDING_CACHE_PATH.exists():
+            print("[INIT] Menemukan file bobot permanen speaker_embedding.pth. Memuat ke RAM...")
+            with open(EMBEDDING_CACHE_PATH, "rb") as f:
+                CACHED_EMBEDDINGS[str(DEFAULT_SPEAKER_PATH)] = pickle.load(f)
+            print("[INIT] Bobot kustom bawaan siap digunakan instan!")
+            return
+
+        # Jika belum ada file .pth tapi ada file .wav asli, lakukan ekstraksi awal
+        if DEFAULT_SPEAKER_PATH.exists():
+            print("[INIT] Membuat file bobot baru. Mengekstrak default_speaker.wav...")
+            if hasattr(model_instance, "extract_conditioning"):
+                embedding = model_instance.extract_conditioning(str(DEFAULT_SPEAKER_PATH))
+            elif hasattr(model_instance, "get_speaker_embedding"):
+                embedding = model_instance.get_speaker_embedding(str(DEFAULT_SPEAKER_PATH))
+            else:
+                embedding = str(DEFAULT_SPEAKER_PATH)
+
+            # Simpan secara fisik ke disk tingkat root agar permanen
+            if not isinstance(embedding, str):
+                with open(EMBEDDING_CACHE_PATH, "wb") as f:
+                    pickle.dump(embedding, f)
+                print("[INIT] File speaker_embedding.pth berhasil disimpan secara permanen!")
+            
+            CACHED_EMBEDDINGS[str(DEFAULT_SPEAKER_PATH)] = embedding
+    except Exception as e:
+        print(f"[WARN] Gagal inisialisasi pembentukan bobot permanen awal: {e}")
+
+
 def get_model():
     global _model
     if _model is None:
@@ -77,6 +111,9 @@ def get_model():
 
                 _model = m
                 print("[INIT] Model ready.")
+                
+                # JALANKAN PROSES TRAINING/EKSTRAKSI SEKALI SAAT STARTUP
+                inisialisasi_bobot_default_permanen(_model)
     return _model
 
 
@@ -146,7 +183,7 @@ def _split_text_safely(text: str, max_chars: int = MAX_CHARS_PER_CHUNK):
     return chunks
 
 # =====================================================================
-# ENGINE UTAMA (SMOOTH STREAMING AUDIO CONCATENATION)
+# ENGINE UTAMA (FAST INFERENCE - INSTANT EMBEDDING LOADING)
 # =====================================================================
 def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(track_tqdm=False)):
     global CACHED_EMBEDDINGS
@@ -166,17 +203,17 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
         sr = getattr(model, "sr", 24000)
         torch.manual_seed(42)
 
+        # Caching dinamis sebagai pengaman cadangan untuk Opsi Custom Voice Baru
         if prompt_path not in CACHED_EMBEDDINGS:
-            progress(0.0, desc="Mengekstrak karakteristik gelombang audio ke RAM...")
-            
+            progress(0.0, desc="Mengekstrak karakteristik gelombang audio kustom baru ke RAM...")
             if hasattr(model, "extract_conditioning") or hasattr(model, "get_speaker_embedding"):
                 extract_fn = getattr(model, "extract_conditioning", getattr(model, "get_speaker_embedding", None))
                 CACHED_EMBEDDINGS[prompt_path] = extract_fn(prompt_path)
             else:
                 CACHED_EMBEDDINGS[prompt_path] = prompt_path
-                
-            print(f"[CACHE] Fitur karakteristik untuk {prompt_path} berhasil dikunci di RAM.")
+            print(f"[CACHE] Fitur karakteristik kustom untuk {prompt_path} berhasil dikunci di RAM.")
 
+        # LOAD INSTAN: Mengambil data biner kustom tanpa membedah file .wav lagi
         speaker_embedding = CACHED_EMBEDDINGS[prompt_path]
 
         wav_parts = []
@@ -212,7 +249,7 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
                 if "cfg_weight" in params:
                     kwargs["cfg_weight"] = 0.3
                 if "max_new_tokens" in params:
-                    kwargs["max_new_tokens"] = 512 # Dinaikkan seiring bertambahnya ukuran panjang karakter chunk
+                    kwargs["max_new_tokens"] = 350
 
                 try:
                     wav = model.generate(ch, **kwargs)
@@ -226,7 +263,7 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
                 if wav.dim() == 1:
                     wav = wav.unsqueeze(0)
 
-                # Masukkan potongan audio murni langsung tanpa diselipkan tensor kosong
+                # Gabungkan potongan audio secara mulus (Seamless Concatenation)
                 wav_parts.append(wav.detach().cpu().clone())
 
         progress(0.95, desc="Menyambungkan seluruh fragmentasi gelombang secara natural...")
@@ -250,12 +287,12 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
 # =====================================================================
 # INTERFACE DESIGN
 # =====================================================================
-with gr.Blocks(title="Chatterbox Seamless Engine") as demo:
-    gr.Markdown("## EduScanner AI Voice Backend - Seamless Continuous Synthesis")
-    gr.Markdown("Kode ini menghapus jeda mati buatan antar-chunk agar audio mengalir menyatu secara natural.")
+with gr.Blocks(title="Chatterbox Persistent Weight Engine") as demo:
+    gr.Markdown("## EduScanner AI Voice Backend - Pre-computed Weights Edition")
+    gr.Markdown("Sistem memuat bobot suara bawaan secara permanen untuk memotong durasi penundaan CPU.")
 
     text_in = gr.Textbox(label="Teks Rangkuman Materi Kuliah", lines=8, placeholder="Ketik teks di sini...")
-    wav_in = gr.Audio(label="Opsi Custom Voice (Kosongkan jika ingin pakai suara default Mythia Batford 1:15 di root)", type="filepath")
+    wav_in = gr.Audio(label="Opsi Custom Voice (Kosongkan jika ingin pakai bobot default Mythia Batford)", type="filepath")
     url_in = gr.Textbox(label="Opsi URL File Audio Custom", placeholder="https://domain-kamu.com/audio.wav")
     
     btn = gr.Button("Sintesis Audio Kloning Suara", variant="primary")
@@ -268,6 +305,10 @@ with gr.Blocks(title="Chatterbox Seamless Engine") as demo:
         api_name="clone_voice"
     )
 
+from chatterbox.tts import ChatterboxTTS
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file
+
 if __name__ == "__main__":
     port = int(os.getenv("PORT", "7860"))
     demo.queue(default_concurrency_limit=1)