Willing to explore ChartterBox

.gitignore

@@ -1 +0,0 @@
-__pycache__

app.py

@@ -45,30 +45,25 @@ def set_seed(seed: int):
 @spaces.GPU
 def generate_tts_audio(
     text_input: str,
-    audio_prompt_path_input: str = None,
-    exaggeration_input: float = 0.5,
-    temperature_input: float = 0.8,
-    seed_num_input: int = 0,
-    cfgw_input: float = 0.5,
-    vad_trim_input: bool = False,
+    audio_prompt_path_input: str,
+    exaggeration_input: float,
+    temperature_input: float,
+    seed_num_input: int,
+    cfgw_input: float
 ) -> 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
-    maintains the prosody, tone, and vocal qualities of the reference speaker, or uses default voice if no reference is provided.
+    Generates TTS audio using the ChatterboxTTS model.
 
     Args:
-        text_input (str): The text to synthesize into speech (maximum 300 characters)
-        audio_prompt_path_input (str, optional): File path or URL to the reference audio file that defines the target voice style. Defaults to None.
-        exaggeration_input (float, optional): Controls speech expressiveness (0.25-2.0, neutral=0.5, extreme values may be unstable). Defaults to 0.5.
-        temperature_input (float, optional): Controls randomness in generation (0.05-5.0, higher=more varied). Defaults to 0.8.
-        seed_num_input (int, optional): Random seed for reproducible results (0 for random generation). Defaults to 0.
-        cfgw_input (float, optional): CFG/Pace weight controlling generation guidance (0.2-1.0). Defaults to 0.5.
+        text_input: The text to synthesize (max 300 characters).
+        audio_prompt_path_input: Path to the reference audio file.
+        exaggeration_input: Exaggeration parameter for the model.
+        temperature_input: Temperature parameter for the model.
+        seed_num_input: Random seed (0 for random).
+        cfgw_input: CFG/Pace weight.
 
     Returns:
-        tuple[int, np.ndarray]: A tuple containing the sample rate (int) and the generated audio waveform (numpy.ndarray)
+        A tuple containing the sample rate (int) and the audio waveform (numpy.ndarray).
     """
     current_model = get_or_load_model()
 
@@ -79,21 +74,12 @@ 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
+        audio_prompt_path=audio_prompt_path_input,
+        exaggeration=exaggeration_input,
+        temperature=temperature_input,
+        cfg_weight=cfgw_input,
     )
     print("Audio generation complete.")
     return (current_model.sr, wav.squeeze(0).numpy())
@@ -128,7 +114,6 @@ 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")
 
@@ -144,9 +129,8 @@ with gr.Blocks() as demo:
             temp,
             seed_num,
             cfg_weight,
-            vad_trim,
         ],
         outputs=[audio_output],
     )
 
-demo.launch(mcp_server=True)
+demo.launch()

chatterbox/src/chatterbox/tts.py

@@ -2,12 +2,10 @@ from dataclasses import dataclass
 from pathlib import Path
 
 import librosa
-import numpy as np
 import torch
 import perth
 import torch.nn.functional as F
 from huggingface_hub import hf_hub_download
-from silero_vad import load_silero_vad, get_speech_timestamps
 
 from .models.t3 import T3
 from .models.s3tokenizer import S3_SR, drop_invalid_tokens
@@ -123,7 +121,6 @@ class ChatterboxTTS:
         self.device = device
         self.conds = conds
         self.watermarker = perth.PerthImplicitWatermarker()
-        self.silero_vad = load_silero_vad()
 
     @classmethod
     def from_local(cls, ckpt_dir, device) -> 'ChatterboxTTS':
@@ -165,33 +162,11 @@ class ChatterboxTTS:
 
         return cls.from_local(Path(local_path).parent, device)
 
-    def trim_excess_silence(self, wav, sr):
-        "Trim excess silence from speech. Input must be a multiple of 16kHz."
-        assert sr % 16_000 == 0, "Silero requires an integer multiple of 16kHz"
+    def prepare_conditionals(self, wav_fpath, exaggeration=0.5):
+        ## Load reference wav
+        s3gen_ref_wav, _sr = librosa.load(wav_fpath, sr=S3GEN_SR)
 
-        # Get VAD as sample-level bool array
-        silero_regions = get_speech_timestamps(wav, self.silero_vad, sampling_rate=sr)
-        vad = np.zeros_like(wav)
-        for region in silero_regions:
-            vad[region["start"]:region["end"]] = 1
-
-        # Dilate VAD
-        max_silence_ms = 400
-        cfilter = np.ones(int(sr * max_silence_ms / (2 * 1000)))
-        dilated_vad = np.convolve(vad, cfilter, mode="same") > 0
-
-        # Trim out silence
-        return wav[dilated_vad]
-
-    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)
-        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)
-        ref_16k_wav = librosa.resample(ref_wav, orig_sr=highres_sr, target_sr=S3_SR)
+        ref_16k_wav = librosa.resample(s3gen_ref_wav, orig_sr=S3GEN_SR, target_sr=S3_SR)
 
         s3gen_ref_wav = s3gen_ref_wav[:self.DEC_COND_LEN]
         s3gen_ref_dict = self.s3gen.embed_ref(s3gen_ref_wav, S3GEN_SR, device=self.device)
@@ -220,10 +195,9 @@ 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, vad_trim=vad_trim)
+            self.prepare_conditionals(audio_prompt_path, exaggeration=exaggeration)
         else:
             assert self.conds is not None, "Please `prepare_conditionals` first or specify `audio_prompt_path`"