多声源分离（三个声源以上）

代码已经调通，链接在下文，可以讲解哟，有需要，可以私信我，真实有效！！！，

1、数据集制作

采集三种不同声音，将其保存为A.wav,B.wav,C.wav;使用Matlab代码将其混合成ABC.wav，以此类推做成自己所需要的数据集。

2、模型设计

网络模型结构还是基于全时域卷积神经网络，将其改成三声源分离。

3、模型训练

描述	值
训练轮数	50
每个批量大小	4
分割块长度	4
初始学习率	0.0001
优化器	adam
激活函数	sigmoid

4、模型推理

import os
import torch
import sys
sys.path.append('./options')
from AudioReader import AudioReader, write_wav, read_wav
import argparse
from torch.nn.parallel import data_parallel
from Conv_TasNet import ConvTasNet
from utils import get_logger
# from option import parse
from option import parse
import tqdm
import time


class Separation():
    def __init__(self, mix_path, yaml_path, model, gpuid):
        super(Separation, self).__init__()
        self.mix = read_wav(mix_path)
        opt = parse(yaml_path, is_tain=False)
        net = ConvTasNet(**opt['net_conf'])
        dicts = torch.load(model, map_location='cpu')
        net.load_state_dict(dicts["model_state_dict"])
        self.logger = get_logger(__name__)
        self.logger.info('Load checkpoint from {}, epoch {: d}'.format(model, dicts["epoch"]))
       # self.net=net.cuda()
        self.net=net.cpu()
       # self.device=torch.device('cuda:{}'.format(
        #    gpuid[0]) if len(gpuid) > 0 else 'cpu')
        self.device = torch.device("cpu")
        self.gpuid=tuple(gpuid)

    def inference(self, file_path):
        with torch.no_grad():
            egs=self.mix.to(self.device)
            norm = torch.norm(egs,float('inf'))
            if len(self.gpuid) != 0:
                ests=self.net(egs)
                spks=[torch.squeeze(s.detach().cpu()) for s in ests]
            else:
                ests=self.net(egs)
                spks=[torch.squeeze(s.detach()) for s in ests]
            index=0
            for s in spks:
                s = s[:egs.shape[0]]
                #norm
                s = s*norm/torch.max(torch.abs(s))
                index += 1
                os.makedirs(file_path+'/spk'+str(index), exist_ok=True)
               # filename=file_path+'/spk'+str(index)+'/'+key
                filename = file_path + '/spk' + str(index)+'.wav'
                write_wav(filename, s, 44100)
        self.logger.info("Compute over {:d} utterances".format(len(self.mix)))


def main():
    parser=argparse.ArgumentParser()
    parser.add_argument(
        '-mix_scp', type=str, default='mix.wav', help='Path to mix scp file.')
    parser.add_argument(
        '-yaml', type=str, default='./options/train/train.yml', help='Path to yaml file.')
    parser.add_argument(
        '-model', type=str, default='./Conv-TasNet-with-skip/best.pt', help="Path to model file.")
    parser.add_argument(
        '-gpuid', type=str, default='0', help='Enter GPU id number')
    parser.add_argument(
        '-save_path', type=str, default='./non-pit-2', help='save result path')
    args=parser.parse_args()
    gpuid=[int(i) for i in args.gpuid.split(',')]
    start = time.time()
    separation=Separation(args.mix_scp, args.yaml, args.model, gpuid)
    separation.inference(args.save_path)
    end = time.time()
    st = str(end-start)
    print("分离所需时间"+st+"s")

if __name__ == "__main__":
    main()
    #python Separation_wav.py -mix_scp ./ceshi/0.wav

代码链接：https://download.csdn.net/download/m0_60103585/89241414