app.py

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

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


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)
        if len(audio.shape) > 1:
            audio = librosa.to_mono(audio.transpose(1, 0))
        temp_path = "temp.wav"
        soundfile.write(temp_path, audio, sr, format="wav")
        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=20,
                                          tran=vc_transform,
                                          f0_predictor=f0p,
                                          auto_predict_f0=auto_f0)
        os.remove(temp_path)
        return sr, out_audio

    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 = "zh-CN-XiaoyiNeural" if gender == '女' else "zh-CN-YunxiNeural"
        ratestr = "+{:.0%}".format(tts_rate) if tts_rate >= 0 else "{:.0%}".format(tts_rate)
        temp_path = "temp.wav"
        p = subprocess.Popen("edge-tts " +
                             " --text " + input_text +
                             " --write-media " + temp_path +
                             " --voice " + voice +
                             " --rate=" + ratestr, shell=True,
                             stdout=subprocess.PIPE,
                             stdin=subprocess.PIPE)
        p.wait()

        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 = (sr, 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 None
        models.append((name, cover, create_fn(model, name)))
    with gr.Blocks() as app:
        gr.Markdown(
            "# <center> GTASA人物语音生成\n"
            "## <center> 模型作者：B站Cyber蝈蝈总\n"
            "<center> 使用此处资源创作的作品，请显著标明出处（B站Cyber蝈蝈总）\n"
        )
        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 (对电音有影响)",
                                                        choices=['crepe', 'harvest', 'dio', 'pm'], value='crepe')
                            auto_f0 = gr.Checkbox(label="自动音高预测 (文本转语音或正常说话可选，唱歌会导致跑调)",
                                                  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'):
                                    gr.Markdown("<center>调用了外部服务，有可能超时报错，可以多试几次")
                                    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='男',
                                                          choices=['男', '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.Markdown(
                                '<div align="center">'
                                f'<img style="width:auto;height:400px;" src="file/{cover}">' if cover else ""
                                                                                                           '</div>'
                            )
                            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(concurrency_count=1, api_open=args.api).launch(share=args.share)