app.py

import argparse
import logging
import os
import re
import gradio.processing_utils as gr_pu
import gradio as gr
import librosa
import numpy as np
import soundfile
from scipy.io import wavfile
import tempfile
import edge_tts
import utils

from inference.infer_tool import Svc

logging.getLogger('numba').setLevel(logging.WARNING)
logging.getLogger('markdown_it').setLevel(logging.WARNING)
logging.getLogger('urllib3').setLevel(logging.WARNING)
logging.getLogger('matplotlib').setLevel(logging.WARNING)

sampling_rate = 44100

tts_voice = {
    "中文男": "zh-CN-YunxiNeural",
    "中文女": "zh-CN-XiaoyiNeural",
    "英文男": "en-US-EricNeural",
    "英文女": "en-US-AnaNeural"
}

hubert_dict = {
    "vec768l12": utils.get_speech_encoder("vec768l12", device="cpu"),
    "vec256l9": utils.get_speech_encoder("vec256l9", device="cpu")
}


def create_fn(model, spk):
    def svc_fn(input_audio, vc_transform, auto_f0, f0p):
        if input_audio is None:
            return 0, None
        sr, audio = input_audio
        audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
        audio = librosa.resample(audio, orig_sr=sr, target_sr=sampling_rate)
        if len(audio.shape) > 1:
            audio = librosa.to_mono(audio.transpose(1, 0))
        temp_path = "temp.wav"
        soundfile.write(temp_path, audio, sampling_rate, format="wav")

        model.hubert_model = hubert_dict[model.speech_encoder]
        out_audio = model.slice_inference(raw_audio_path=temp_path,
                                          spk=spk,
                                          slice_db=-40,
                                          cluster_infer_ratio=0,
                                          noice_scale=0.4,
                                          clip_seconds=10,
                                          tran=vc_transform,
                                          f0_predictor=f0p,
                                          auto_predict_f0=auto_f0)
        model.clear_empty()
        os.remove(temp_path)
        return sampling_rate, out_audio

    async def tts_fn(input_text, gender, tts_rate, vc_transform, auto_f0, f0p):
        if input_text == '':
            return 0, None
        input_text = re.sub(r"[\n\,\(\) ]", "", input_text)
        voice = tts_voice[gender]
        ratestr = "+{:.0%}".format(tts_rate) if tts_rate >= 0 else "{:.0%}".format(tts_rate)
        communicate = edge_tts.Communicate(text=input_text, voice=voice, rate=ratestr)
        with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
            temp_path = tmp_file.name
        await communicate.save(temp_path)

        audio, sr = librosa.load(temp_path)
        audio = librosa.resample(audio, orig_sr=sr, target_sr=sampling_rate)
        os.remove(temp_path)
        temp_path = "temp.wav"
        wavfile.write(temp_path, sampling_rate, (audio * np.iinfo(np.int16).max).astype(np.int16))
        sr, audio = gr_pu.audio_from_file(temp_path)
        input_audio = (sampling_rate, audio)
        return svc_fn(input_audio, vc_transform, auto_f0, f0p)

    return svc_fn, tts_fn


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--device', type=str, default='cpu')
    parser.add_argument('--api', action="store_true", default=False)
    parser.add_argument("--share", action="store_true", default=False, help="share gradio app")
    args = parser.parse_args()
    models = []
    for f in os.listdir("models"):
        name = f
        model = Svc(fr"models/{f}/{f}.pth", f"models/{f}/config_{f}.json", device=args.device)
        cover = f"models/{f}/cover.png" if os.path.exists(f"models/{f}/cover.png") else f"models/{f}/cover.jpg"
        models.append((name, cover, create_fn(model, name)))
    with gr.Blocks() as app:
        gr.Markdown(
            """
            # <center> 圣安地列斯角色语音生成
            ## <center> 模型作者：B站[Cyber蝈蝈总](https://space.bilibili.com/37706580)
            #### <center> 传送门[GTAVC](https://huggingface.shushu.icu/spaces/GroveStreet/GTAVC_SOVITS)；[GTAV](https://huggingface.shushu.icu/spaces/GroveStreet/GTAV_SOVITS)
            <center> 使用此资源创作的作品请标出处，CJ有两个模型，carl1更清晰，carl2音域广
            """
        )
        with gr.Tabs():
            for (name, cover, (svc_fn, tts_fn)) in models:
                with gr.TabItem(name):
                    with gr.Row():
                        with gr.Column():
                            with gr.Row():
                                vc_transform = gr.Number(label="音高调整 (正负半音，12为1个八度)", value=0)
                                f0_predictor = gr.Radio(label="f0预测器 (推荐rmvpe)",
                                                        choices=['crepe', 'harvest', 'rmvpe'], value='rmvpe')
                            auto_f0 = gr.Checkbox(label="自动音高预测 (文本转语音或讲话可选,会导致唱歌跑调)",
                                                  value=False)
                            with gr.Tabs():
                                with gr.TabItem('语音转语音'):
                                    svc_input = gr.Audio(
                                        label="上传干声 (已支持无限长音频，处理时间约为原音频时间的5倍)")
                                    svc_submit = gr.Button("生成", variant="primary")

                                with gr.TabItem('文本转语音'):
                                    tts_input = gr.Textbox(label='说话内容', value='',
                                                           placeholder='已支持无限长内容，处理时间约为说完原内容时间的5倍')
                                    with gr.Row():
                                        gender = gr.Radio(label='说话人性别 (男音调低，女音调高)', value='中文男',
                                                          choices=['中文男', '中文女', '英文男', '英文女'])
                                        tts_rate = gr.Number(label='语速 (正负, 单位百分比)', value=0)
                                    tts_submit = gr.Button("生成", variant="primary")

                        with gr.Column():
                            gr.Image(cover, width=400, height=400)
                            vc_output = gr.Audio(label="输出音频")
                    svc_submit.click(svc_fn, [svc_input, vc_transform, auto_f0, f0_predictor], vc_output)
                    tts_submit.click(tts_fn, [tts_input, gender, tts_rate, vc_transform, auto_f0, f0_predictor],
                                     vc_output)
        app.queue(api_open=args.api).launch(share=args.share)