| |
| import os |
| import gradio as gr |
| import numpy as np |
| import torch |
| from huggingface_hub import snapshot_download |
| from qwen_tts import Qwen3TTSModel |
|
|
| MODEL_SIZES = ["0.6B", "1.7B"] |
| LANGUAGES = ["Auto", "Chinese", "English", "Japanese", "Korean", "French", "German", "Spanish", "Portuguese", "Russian", "Italian"] |
|
|
|
|
| def get_model_path(model_type: str, model_size: str) -> str: |
| return snapshot_download(f"Qwen/Qwen3-TTS-12Hz-{model_size}-{model_type}") |
|
|
|
|
| print("Loading Base 1.7B model...") |
| base_model_1_7b = Qwen3TTSModel.from_pretrained( |
| get_model_path("Base", "1.7B"), |
| device_map="cuda", |
| dtype=torch.float16, |
| ) |
| print("Model loaded successfully!") |
|
|
| BASE_MODELS = { |
| "1.7B": base_model_1_7b, |
| } |
|
|
|
|
| def _normalize_audio(wav, eps=1e-12, clip=True): |
| x = np.asarray(wav) |
| if np.issubdtype(x.dtype, np.integer): |
| info = np.iinfo(x.dtype) |
| if info.min < 0: |
| y = x.astype(np.float32) / max(abs(info.min), info.max) |
| else: |
| mid = (info.max + 1) / 2.0 |
| y = (x.astype(np.float32) - mid) / mid |
| elif np.issubdtype(x.dtype, np.floating): |
| y = x.astype(np.float32) |
| m = np.max(np.abs(y)) if y.size else 0.0 |
| if m > 1.0 + 1e-6: |
| y = y / (m + eps) |
| else: |
| raise TypeError(f"Unsupported dtype: {x.dtype}") |
| if clip: |
| y = np.clip(y, -1.0, 1.0) |
| if y.ndim > 1: |
| y = np.mean(y, axis=-1).astype(np.float32) |
| return y |
|
|
|
|
| def _audio_to_tuple(audio): |
| if audio is None: |
| return None |
| if isinstance(audio, tuple) and len(audio) == 2 and isinstance(audio[0], int): |
| sr, wav = audio |
| wav = _normalize_audio(wav) |
| return wav, int(sr) |
| if isinstance(audio, dict) and "sampling_rate" in audio and "data" in audio: |
| sr = int(audio["sampling_rate"]) |
| wav = _normalize_audio(audio["data"]) |
| return wav, sr |
| return None |
|
|
|
|
| def generate_voice_clone(ref_audio, ref_text, target_text, language, use_xvector_only, model_size, progress=gr.Progress(track_tqdm=True)): |
| if not target_text or not target_text.strip(): |
| return None, "Error: Target text is required." |
| audio_tuple = _audio_to_tuple(ref_audio) |
| if audio_tuple is None: |
| return None, "Error: Reference audio is required." |
| if not use_xvector_only and (not ref_text or not ref_text.strip()): |
| return None, "Error: Reference text is required when 'Use x-vector only' is not enabled." |
| try: |
| tts = BASE_MODELS.get(model_size, base_model_1_7b) |
| wavs, sr = tts.generate_voice_clone( |
| text=target_text.strip(), |
| language=language, |
| ref_audio=audio_tuple, |
| ref_text=ref_text.strip() if ref_text else None, |
| x_vector_only_mode=use_xvector_only, |
| max_new_tokens=2048, |
| ) |
| return (sr, wavs[0]), "Voice clone generation completed successfully!" |
| except Exception as e: |
| return None, f"Error: {type(e).__name__}: {e}" |
|
|
|
|
| def build_ui(): |
| with gr.Blocks(title="Worder TTS") as demo: |
| gr.Markdown("# Worder Voice Clone (Base 1.7B)") |
| with gr.Row(): |
| with gr.Column(): |
| clone_ref_audio = gr.Audio(label="Reference Audio", type="numpy") |
| clone_ref_text = gr.Textbox(label="Reference Text", lines=2) |
| clone_xvector = gr.Checkbox(label="Use x-vector only", value=False) |
| with gr.Column(): |
| clone_target_text = gr.Textbox(label="Target Text", lines=4) |
| with gr.Row(): |
| clone_language = gr.Dropdown(label="Language", choices=LANGUAGES, value="Auto") |
| clone_model_size = gr.Dropdown(label="Model Size", choices=["1.7B"], value="1.7B") |
| clone_btn = gr.Button("Generate", variant="primary") |
| clone_audio_out = gr.Audio(label="Generated Audio", type="numpy") |
| clone_status = gr.Textbox(label="Status", lines=2, interactive=False) |
| clone_btn.click( |
| generate_voice_clone, |
| inputs=[clone_ref_audio, clone_ref_text, clone_target_text, clone_language, clone_xvector, clone_model_size], |
| outputs=[clone_audio_out, clone_status], |
| ) |
| return demo |
|
|
|
|
| if __name__ == "__main__": |
| demo = build_ui() |
| demo.launch() |
