app.py

import spaces
import sys
import torch

import gradio as gr
import opencc

# 添加第三方库路径
sys.path.append('third_party/Matcha-TTS')

from cosyvoice.cli.cosyvoice import CosyVoice2
from cosyvoice.utils.file_utils import load_wav

from huggingface_hub import hf_hub_download

# 繁简转换
converter = opencc.OpenCC('s2t.json')

# 预设参考音频选项
PRESET_AUDIO_OPTIONS = {
    "ZoengJyutGaai": "asset/张悦楷.wav",
    "Trump": "asset/特朗普.wav",
    "Taiyi Zhenren": "asset/太乙真人.wav",
    "Custom Upload": None
}

# 加载模型
cosyvoice_base = CosyVoice2(
    'ASLP-lab/Cosyvoice2-Yue',
    load_jit=False, load_trt=False, load_vllm=False, fp16=False
)
print('load model 1')
cosyvoice_zjg = CosyVoice2(
    'ASLP-lab/Cosyvoice2-Yue-ZoengJyutGaai',
    load_jit=False, load_trt=False, load_vllm=False, fp16=False
)
print('load model 2')
# cosyvoice_biaobei = CosyVoice2(
#     'pretrained_models/CosyVoice2-yue-biaobei',
#     load_jit=False, load_trt=False, load_vllm=False, fp16=False
# )

@spaces.GPU 
def tts_inference(model_choice, text, preset_audio_choice, custom_audio):
    # 选择模型和默认音频
    if model_choice == "CosyVoice2-Yue-ZoengJyutGaai":
        model = cosyvoice_zjg
        prompt_audio = "asset/sg_017_090.wav"
    elif model_choice == "CosyVoice2-精品女音":
        model = cosyvoice_base
        prompt_audio = "asset/F01_中立_20054.wav"
    elif model_choice == "CosyVoice2-Yue":
        model = cosyvoice_base
        if preset_audio_choice == "Custom Upload":
            if custom_audio is None:
                return None, "请上传参考音频或选择预设音频"
            prompt_audio = custom_audio
        else:
            prompt_audio = PRESET_AUDIO_OPTIONS[preset_audio_choice]
            if prompt_audio is None:
                return None, "请选择有效的参考音频"
    else:
        return None, "未知模型"
    
    model.model.cuda()

    # 繁简转换
    text = converter.convert(text)
    prompt_speech_16k = load_wav(prompt_audio, 16000)

    all_speech = []
    for _, j in enumerate(
        model.inference_instruct2(
            text, "用粤语说这句话", prompt_speech_16k, stream=False
        )
    ):
        all_speech.append(j['tts_speech'])

    concatenated_speech = torch.cat(all_speech, dim=1)
    audio_numpy = concatenated_speech.squeeze(0).cpu().numpy()
    sample_rate = model.sample_rate

    return (sample_rate, audio_numpy), f"successfully generated：{text}"


# ---- Gradio Interface ----
with gr.Blocks() as demo:
    gr.Markdown("# WenetSpeech-Yue TTS")
    # 添加模型说明文字
    gr.Markdown("""
    - Cosyvoice2-Yue: Finetuned CosyVoice2 with WenetSpeech-Yue.
    - Cosyvoice2-Yue-ZoengJyutGaai: Finetuned Cosyvoice2-Yue with The Zoeng Jyut Gaai Story-telling Speech Dataset (https://canclid.github.io/zoengjyutgaai/).
    """)
    
    with gr.Row():
        with gr.Column(scale=1):
            model_choice = gr.Dropdown(
                ["CosyVoice2-Yue", "CosyVoice2-Yue-ZoengJyutGaai"],
                label="Select model", value="CosyVoice2-Yue"
            )
            text_input = gr.Textbox(lines=2, label="Input text")
            preset_audio = gr.Dropdown(
                choices=list(PRESET_AUDIO_OPTIONS.keys()),
                value="Custom Upload",
                label="Select speaker prompt (CosyVoice2-Yue only)"
            )
            custom_audio = gr.Audio(sources=["upload"], type="filepath", label="Upload prompt audio(CosyVoice2-Yue only)")
            generate_btn = gr.Button("Generate")
        
        with gr.Column(scale=1):
            output_audio = gr.Audio(type="numpy", label="Generated audio")
            status_text = gr.Textbox(label="Status")
    
    generate_btn.click(
        fn=tts_inference,
        inputs=[model_choice, text_input, preset_audio, custom_audio],
        outputs=[output_audio, status_text]
    )

demo.launch()