app.py

import io

import gradio as gr
import librosa
import numpy as np
import soundfile
from inference import slicer
from inference.infer_tool import Svc
import logging
from logmmse import logmmse
from typing import Tuple
import time

logging.getLogger('numba').setLevel(logging.WARNING)

model_sing = "logs/32k/sing1.pth"
model_talk = "logs/32k/talk1.pth"
config_name = "configs/config.json"

sid_map = {
    "唱歌特化": "sing",
    "杂谈特化": "talk"
}


class YukieGradio:
    def __init__(self):
        self.UI = gr.Blocks()
        with self.UI:
            with gr.Tabs():
                with gr.TabItem("Basic"):
                    gr.Markdown(value="""
                        # 前言
                        * 本demo基于[sovits 3.0 32khz版本](https://github.com/innnky/so-vits-svc)训练的
                        
                        * 由于雪宝唱歌和说话声线差异较大，一个模型包括说话和唱歌两部分声线会使得工作量较大（声纹识别等），故而选择将这两部分分开训练；目前这一版中，杂谈声线数据较差，仍需后续更新

                        # 声明
                        在使用此模型前请阅读[AI雪绘Yukie模型使用协议](https://huggingface.co/spaces/yukie/yukie-sovits3/edit/main/terms.md)
                        
                        # start！
                        上传一段**纯人声**干音（推荐60s以内），或者直接使用网站录音（二者只能选其一，优先使用上传音频）
                        
                        然后点击提交即可开始推理！

                        **请使用无bgm，无混响的人声来进行生成推理，否则效果可能会较差**
                        """)
                    self.sid = gr.Dropdown(label="音色", choices=[
                                           "唱歌特化", "杂谈特化"], value="唱歌特化", interactive=True)
                    self.dev = gr.Dropdown(label="设备（云端一般请勿切换，使用默认值即可）", choices=[
                                           "cuda", "cpu"], value="cpu", interactive=True)
                    self.inMic = gr.Microphone(label="录音")
                    self.inAudio = gr.Audio(label="上传音频")
                    self.needLogmmse = gr.Checkbox(label="是否使用自带降噪")
                    self.slice_db = gr.Slider(label="切片阈值(较嘈杂时-30，保留呼吸声时-50，一般默认-40)",
                                              maximum=32767, minimum=-32768, step=0.1, value=-40)
                    self.vcTransform = gr.Number(
                        label="升降调（整数，可以正负，半音数量，升高八度就是12）", value=0)
                    self.vcSubmit = gr.Button("转换", variant="primary")
                    self.outVcText = gr.Textbox(
                        label="音高平均偏差半音数量，体现转换音频的跑调情况（一般小于0.5）")
                    self.outAudio = gr.Audio(
                        source="upload", type="numpy", label="Output Audio")
                    self.f0_image = gr.Image(
                        label="f0曲线，蓝色为输入音高，橙色为合成音频的音高（代码有误差）")
                    gr.Markdown(value="""
                        ## 注意
                        如果要在本地使用该demo，请使用  `git lfs clone https://huggingface.co/spaces/yukie/yukie-sovits3`克隆该仓库([简单教程](https://huggingface.co/spaces/yukie/yukie-sovits3/edit/main/local.md))
                    """)
                    self.vcSubmit.click(infer, inputs=[self.inMic, self.inAudio, self.vcTransform, self.slice_db, self.needLogmmse, self.sid, self.dev], outputs=[
                        self.outVcText, self.outAudio, self.f0_image])


def infer(inMic, inAudio, transform, slice_db, lm, sid, dev):
    if inAudio != None:
        sampling_rate, inaudio = inAudio
    else:
        if inMic != None:
            sampling_rate, inaudio = inMic
        else:
            return "请上传一段音频后再次尝试", None

    print("start inference")
    start_time = time.time()
    # 预处理，重编码
    inaudio = (inaudio / np.iinfo(inaudio.dtype).max).astype(np.float32)
    if len(inaudio.shape) > 1:
        inaudio = librosa.to_mono(inaudio.transpose(1, 0))
    if sampling_rate != 32000:
        inaudio = librosa.resample(
            inaudio, orig_sr=sampling_rate, target_sr=32000)
    if lm:
        inaudio = logmmse(inaudio, 32000)

    ori_wav_path = "tmp_ori.wav"
    soundfile.write(ori_wav_path, inaudio, 32000, format="wav")
    chunks = slicer.cut(ori_wav_path, db_thresh=slice_db)
    audio_data, audio_sr = slicer.chunks2audio(ori_wav_path, chunks)

    audio = []
    sid = sid_map[sid]
    if sid == "sing":
        svc_model = Svc(model_sing, config_name, dev=dev)
    else:
        svc_model = Svc(model_talk, config_name, dev=dev)

    for (slice_tag, data) in audio_data:
        length = int(np.ceil(len(data) / audio_sr * svc_model.target_sample))
        raw_path = io.BytesIO()
        soundfile.write(raw_path, data, audio_sr, format="wav")
        raw_path.seek(0)
        if slice_tag:
            _audio = np.zeros(length)
        else:
            out_audio, out_str = svc_model.infer("yukie", transform, raw_path)
            _audio = out_audio.cpu().numpy()
        audio.extend(list(_audio))
    audio = (np.array(audio) * 32768.0).astype('int16')
    used_time = time.time() - start_time

    out_wav_path = "tmp.wav"
    soundfile.write(out_wav_path, audio, 32000, format="wav")

    mistake, var = svc_model.calc_error(ori_wav_path, out_wav_path, transform)
    out_picture = svc_model.f0_plt(ori_wav_path, out_wav_path, transform)
    out_str = ("Success! total use time:{}s\n半音偏差:{}\n半音方差:{}".format(
        used_time, mistake, var))

    return out_str, (32000, audio), gr.Image.update("temp.jpg")


if __name__ == "__main__":
    app = YukieGradio()
    app.UI.launch()