Update app.py

app.py

@@ -8,7 +8,7 @@ import io
 import os
 import soundfile as sf
 from nltk.tokenize import sent_tokenize
-from pydub import AudioSegment
+from pydub import AudioSegment, silence
 import gradio as gr
 
 from chatterbox.src.chatterbox.tts import ChatterboxTTS
@@ -37,7 +37,7 @@ def get_model():
 get_model()
 
 # ===============================
-# SEED (Using their exact robust seeder)
+# SEED
 # ===============================
 def set_seed(seed):
     torch.manual_seed(seed)
@@ -64,7 +64,7 @@ def generate_tts(
     temperature=0.8,
     seed=0,
     cfg_weight=0.5,
-    vad_trim=True # The secret parameter from their code
+    vad_trim=True # We will process this manually now
 ):
 
     model = get_model()
@@ -73,25 +73,37 @@ def generate_tts(
         set_seed(int(seed))
 
     # --------------------------------
-    # HF Official Kwargs setup
+    # HF Official Kwargs (Removed vad_trim to fix the crash)
     # --------------------------------
     generate_kwargs = {
         "exaggeration": exaggeration,
         "temperature": temperature,
         "cfg_weight": cfg_weight,
-        "vad_trim": vad_trim,
     }
 
     temp_prompt = None
     if ref_audio:
         try:
-            # Chatterbox is highly sensitive. Standardize to WAV just in case.
             audio = AudioSegment.from_file(ref_audio)
+            
+            # ---------------------------------------------------------
+            # MANUAL REF VAD TRIMMING (Replaces the missing HF feature)
+            # This strips room tone/breaths from the reference upload
+            # so the model doesn't hallucinate.
+            # ---------------------------------------------------------
+            if vad_trim:
+                print("✂️ Sanitizing reference audio...")
+                non_silent_ranges = silence.detect_nonsilent(audio, min_silence_len=100, silence_thresh=-45)
+                if non_silent_ranges:
+                    start_trim = non_silent_ranges[0][0]
+                    end_trim = non_silent_ranges[-1][1]
+                    audio = audio[start_trim:end_trim]
+
             temp_prompt = "voice_prompt.wav"
             audio.export(temp_prompt, format="wav")
             generate_kwargs["audio_prompt_path"] = temp_prompt
-        except:
-            print("⚠️ Reference audio failed — using default voice.")
+        except Exception as e:
+            print(f"⚠️ Reference audio failed: {e} — using default voice.")
 
     # --------------------------------
     # Sentence chunking
@@ -123,10 +135,10 @@ def generate_tts(
     for i, chunk in enumerate(chunks):
         print(f"➡️ Generating chunk [{i+1}/{len(chunks)}]: {chunk[:50]}...")
 
-        # 1. Generate Raw Audio using HF exact params
+        # 1. Generate Raw Audio 
         wav = model.generate(chunk, **generate_kwargs)
         
-        # 2. Convert tensor to numpy exactly like HF
+        # 2. Convert tensor to numpy
         wav_np = wav.squeeze(0).cpu().numpy()
 
         # 3. Write to buffer
@@ -136,7 +148,13 @@ def generate_tts(
 
         segment = AudioSegment.from_wav(buffer)
 
-        # 4. Smooth stitch (No harsh volume choppers needed anymore)
+        # Optional Output Sanitizer (Prevents trailing hisses on chunks)
+        if vad_trim:
+             out_silent = silence.detect_nonsilent(segment, min_silence_len=100, silence_thresh=-45)
+             if out_silent:
+                 segment = segment[:out_silent[-1][1] + 50] # Leave 50ms padding so we don't eat consonants
+
+        # 4. Smooth stitch 
         if len(final_audio) > 0:
             final_audio = final_audio.append(segment, crossfade=CROSSFADE_MS)
         else:
@@ -160,8 +178,8 @@ def generate_tts(
 # GRADIO UI
 # ===============================
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown("## 🎙️ Chatterbox TTS (Unlimited Length + HF Official Engine)")
-    gr.Markdown("Uses the official HuggingFace backend parameters (`vad_trim`) to eliminate artifacts, wrapped in an unlimited sentence chunker.")
+    gr.Markdown("## 🎙️ Chatterbox TTS (Unlimited Length + Manual VAD)")
+    gr.Markdown("Uses the official HuggingFace default settings and custom Reference VAD trimming to eliminate artifacts, wrapped in an unlimited sentence chunker.")
 
     with gr.Row():
         with gr.Column():
@@ -178,7 +196,6 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
             )
             
             with gr.Accordion("⚙️ Engine Settings (Synced to HF Defaults)", open=True):
-                # Sliders updated to match HF Official defaults
                 exaggeration = gr.Slider(0.25, 2.0, value=0.5, step=0.05, label="Exaggeration (Neutral = 0.5)")
                 cfg = gr.Slider(0.2, 1.0, value=0.5, step=0.05, label="CFG / Pace Weight")
                 temperature = gr.Slider(0.05, 5.0, value=0.8, step=0.05, label="Temperature")