Update app.py

app.py

@@ -29,19 +29,21 @@ def _torch_load_cpu(*args, **kwargs):
 torch.load = _torch_load_cpu
 
 # =====================================================================
-# CONFIG & PATH MANAGEMENT
+# CONFIG & PATH MANAGEMENT (DIOPTIMALKAN AGAR SUARA MULUS)
 # =====================================================================
 MODEL_REPO = "grandhigh/Chatterbox-TTS-Indonesian"
 CHECKPOINT_FILENAME = "t3_cfg.safetensors"
 DEVICE = "cpu"
 
 MAX_TOTAL_CHARS = int(os.getenv("MAX_TOTAL_CHARS", "2400"))
-MAX_CHARS_PER_CHUNK = int(os.getenv("MAX_CHARS_PER_CHUNK", "220"))
-MAX_CHUNKS = int(os.getenv("MAX_CHUNKS", "12"))
-PAUSE_SECONDS = float(os.getenv("PAUSE_SECONDS", "0.15"))
+# 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
+PAUSE_SECONDS = float(os.getenv("PAUSE_SECONDS", "0.0"))
 DOWNLOAD_TIMEOUT = int(os.getenv("DOWNLOAD_TIMEOUT", "90"))
 
-# Sesuai request: Suara default ditaruh langsung di root folder (/)
+# Jalur untuk suara default bawaan sistem langsung di root (/)
 ROOT_DIR = Path(__file__).parent
 DEFAULT_SPEAKER_PATH = ROOT_DIR / "default_speaker.wav"
 
@@ -96,7 +98,6 @@ def _resolve_audio_input(audio_file, audio_url: str):
         except Exception:
             pass
             
-    # Jika parameter audio kosong, otomatis gunakan file suara default langsung dari root (/)
     if DEFAULT_SPEAKER_PATH.exists():
         return str(DEFAULT_SPEAKER_PATH)
         
@@ -145,7 +146,7 @@ def _split_text_safely(text: str, max_chars: int = MAX_CHARS_PER_CHUNK):
     return chunks
 
 # =====================================================================
-# ENGINE UTAMA
+# ENGINE UTAMA (SMOOTH STREAMING AUDIO CONCATENATION)
 # =====================================================================
 def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(track_tqdm=False)):
     global CACHED_EMBEDDINGS
@@ -154,7 +155,6 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
         if not raw_text:
             raise gr.Error("Text prompt tidak boleh kosong.")
 
-        # Ambil jalur audio (Custom atau Default Root)
         prompt_path = _resolve_audio_input(audio_file, audio_url)
         if not prompt_path:
             raise gr.Error("Suara acuan default_speaker.wav tidak ditemukan di root server.")
@@ -166,9 +166,8 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
         sr = getattr(model, "sr", 24000)
         torch.manual_seed(42)
 
-        # Proses pembuatan cache embedding jika file audio terdeteksi baru
         if prompt_path not in CACHED_EMBEDDINGS:
-            progress(0.0, desc="Mengekstrak karakteristik gelombang audio ke RAM (Hanya sekali)...")
+            progress(0.0, desc="Mengekstrak karakteristik gelombang audio 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))
@@ -181,10 +180,8 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
         speaker_embedding = CACHED_EMBEDDINGS[prompt_path]
 
         wav_parts = []
-        pause = torch.zeros(1, int(sr * PAUSE_SECONDS))
         total = len(chunks)
         
-        # Bedah blueprint fungsi model asli untuk memvalidasi parameter masuk
         sig = inspect.signature(model.generate)
         params = sig.parameters
 
@@ -196,9 +193,7 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
                 if not re.search(r"[.!?…]$", ch):
                     ch += "."
 
-                # Saring parameter agar hanya memasukkan key yang dikenal oleh library model
                 kwargs = {}
-                
                 if "audio_prompt_path" in params:
                     kwargs["audio_prompt_path"] = speaker_embedding
                 elif "speaker_embedding" in params:
@@ -208,7 +203,6 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
                     if len(list_keys) > 1:
                         kwargs[list_keys[1]] = speaker_embedding
                 
-                # Parameter opsional tambahan (Hanya disuntikkan jika disupport oleh versi model)
                 if "temperature" in params:
                     kwargs["temperature"] = 0.05
                 if "top_p" in params:
@@ -218,9 +212,8 @@ 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"] = 260
+                    kwargs["max_new_tokens"] = 512 # Dinaikkan seiring bertambahnya ukuran panjang karakter chunk
 
-                # Eksekusi aman bebas dari bug TypeError
                 try:
                     wav = model.generate(ch, **kwargs)
                 except TypeError:
@@ -233,19 +226,15 @@ 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
                 wav_parts.append(wav.detach().cpu().clone())
-                wav_parts.append(pause)
-
-        if wav_parts:
-            wav_parts = wav_parts[:-1]
 
-        progress(0.95, desc="Menggabungkan kompilasi audio...")
+        progress(0.95, desc="Menyambungkan seluruh fragmentasi gelombang secara natural...")
         full_wav = torch.cat(wav_parts, dim=1)
 
         out_path = tempfile.NamedTemporaryFile(delete=False, suffix=".wav").name
         ta.save(out_path, full_wav, sr)
 
-        # Bersihkan memori RAM kontainer secara agresif
         del wav_parts
         del full_wav
         gc.collect()
@@ -261,9 +250,9 @@ def clone_voice(text: str, audio_file, audio_url: str, progress=gr.Progress(trac
 # =====================================================================
 # INTERFACE DESIGN
 # =====================================================================
-with gr.Blocks(title="Chatterbox Flat Root Safe Engine") as demo:
-    gr.Markdown("## EduScanner AI Voice Backend - Flat Root Configuration")
-    gr.Markdown("Sistem membaca `default_speaker.wav` langsung dari tingkat root (/) folder Space.")
+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.")
 
     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")