add ref wav vad trimming option

app.py

@@ -49,13 +49,14 @@ def generate_tts_audio(
     exaggeration_input: float = 0.5,
     temperature_input: float = 0.8,
     seed_num_input: int = 0,
-    cfgw_input: float = 0.5
+    cfgw_input: float = 0.5,
+    vad_trim_input: bool = False,
 ) -> tuple[int, np.ndarray]:
     """
     Generate high-quality speech audio from text using ChatterboxTTS model with optional reference audio styling.
-    
-    This tool synthesizes natural-sounding speech from input text. When a reference audio file 
-    is provided, it captures the speaker's voice characteristics and speaking style. The generated audio 
+
+    This tool synthesizes natural-sounding speech from input text. When a reference audio file
+    is provided, it captures the speaker's voice characteristics and speaking style. The generated audio
     maintains the prosody, tone, and vocal qualities of the reference speaker, or uses default voice if no reference is provided.
 
     Args:
@@ -78,17 +79,18 @@ def generate_tts_audio(
         set_seed(int(seed_num_input))
 
     print(f"Generating audio for text: '{text_input[:50]}...'")
-    
+
     # Handle optional audio prompt
     generate_kwargs = {
         "exaggeration": exaggeration_input,
         "temperature": temperature_input,
         "cfg_weight": cfgw_input,
+        "vad_trim": vad_trim_input,
     }
-    
+
     if audio_prompt_path_input:
         generate_kwargs["audio_prompt_path"] = audio_prompt_path_input
-    
+
     wav = current_model.generate(
         text_input[:300],  # Truncate text to max chars
         **generate_kwargs
@@ -126,6 +128,7 @@ with gr.Blocks() as demo:
             with gr.Accordion("More options", open=False):
                 seed_num = gr.Number(value=0, label="Random seed (0 for random)")
                 temp = gr.Slider(0.05, 5, step=.05, label="Temperature", value=.8)
+                vad_trim = gr.Checkbox(label="Ref VAD trimming", value=False)
 
             run_btn = gr.Button("Generate", variant="primary")
 
@@ -141,8 +144,9 @@ with gr.Blocks() as demo:
             temp,
             seed_num,
             cfg_weight,
+            vad_trim,
         ],
         outputs=[audio_output],
     )
 
-demo.launch(mcp_server=True)
+demo.launch(mcp_server=True)

chatterbox/src/chatterbox/tts.py

@@ -183,10 +183,11 @@ class ChatterboxTTS:
         # Trim out silence
         return wav[dilated_vad]
 
-    def prepare_conditionals(self, wav_fpath, exaggeration=0.5):
+    def prepare_conditionals(self, wav_fpath, exaggeration=0.5, vad_trim=False):
         # Load reference wav at high SR and trim silence
         ref_wav, highres_sr = librosa.load(wav_fpath, sr=48_000)
-        ref_wav = self.trim_excess_silence(ref_wav, highres_sr)
+        if vad_trim:
+            ref_wav = self.trim_excess_silence(ref_wav, highres_sr)
 
         # Resample down
         s3gen_ref_wav = librosa.resample(ref_wav, orig_sr=highres_sr, target_sr=S3GEN_SR)
@@ -219,9 +220,10 @@ class ChatterboxTTS:
         exaggeration=0.5,
         cfg_weight=0.5,
         temperature=0.8,
+        vad_trim=False,
     ):
         if audio_prompt_path:
-            self.prepare_conditionals(audio_prompt_path, exaggeration=exaggeration)
+            self.prepare_conditionals(audio_prompt_path, exaggeration=exaggeration, vad_trim=vad_trim)
         else:
             assert self.conds is not None, "Please `prepare_conditionals` first or specify `audio_prompt_path`"