sovits/sovits_inferencer.py

import os

import gradio as gr
import soundfile
import torch
import utils
import infer_tool
from sovits import ROOT_PATH

class SovitsInferencer:
    def __init__(self, hps_path, device="cpu"):
        print("init")
        self.device = torch.device(device)
        self.hps = utils.get_hparams_from_file(hps_path)
        self.model_path = self.get_latest_model_path()
        self.svc = infer_tool.Svc(self.model_path, hps_path, device=device)
    
    def get_latest_model_path(self):
        model_dir_path = os.path.join(ROOT_PATH, "models")
        return utils.latest_checkpoint_path(model_dir_path, "G_*.pth")
    
    def infer(self, audio_record, audio_upload, tran):
        if audio_upload is not None:
            audio_path = audio_upload
        elif audio_record is not None:
            audio_path = audio_record
        else:
            return "你需要上传wav文件或使用网页内置的录音！", None

        audio, sampling_rate = self.svc.format_wav(audio_path)
        duration = audio.shape[1] / sampling_rate
        if duration > 60:
            return "请上传小于60s的音频，需要转换长音频请使用colab", None

        o_audio, out_sr = self.svc.infer(0, tran, audio_path)
        out_path = f"./out_temp.wav"
        soundfile.write(out_path, o_audio.cpu(), self.svc.target_sample)
        mistake, var = self.svc.calc_error(audio_path, out_path, tran)
        return f"半音偏差：{mistake}\n半音方差：{var}", (self.hps.data.sampling_rate, o_audio.cpu().numpy())
    
    def render(self):
        gr.Markdown("""
        未完成，效果有待提升\n
        该模型适合**歌声**的声线转换，目前仅支持**45s以内**、**无伴奏**、**单声道**的**wav或mp3格式**文件
        """)
        record_input = gr.Audio(source="microphone", label="录制你的声音", type="filepath", elem_id="audio_inputs")
        upload_input = gr.Audio(source="upload", label="上传音频（长度小于45秒）", type="filepath",
                                elem_id="audio_inputs")
        # vc_speaker = gr.Number(label="Speaker", value=0)                        
        vc_transform = gr.Number(label="升降半音（整数，可以正负，半音数量，升高八度就是12）", value=0)
        vc_submit = gr.Button("转换", variant="primary")
        out_message = gr.Textbox(label="Output Message")
        out_audio = gr.Audio(label="Output Audio")
        # vc_submit.click(self.infer, [vc_speaker, record_input, upload_input, vc_transform], [out_message, out_audio])
        vc_submit.click(self.infer, [record_input, upload_input, vc_transform], [out_message, out_audio])