基于神经网络的自监督学习方法音频分离器（Matlab代码实现）

       目录

💥1 概述

📚2 运行结果

🎉3 参考文献

👨‍💻4 Matlab代码

---

💥1 概述

神经网络的输入是混合（男性+女性）音频的振幅谱。神经网络的输出目标是男性说话者理想的软掩模。损失函数是使输出和输入目标之间的均方误差最小化。在输出端，利用混合信号的输出幅度谱和相位将音频STFT转换回时域，并采用自监督学习方法。训练集是混合音频的前40秒中理想软掩模和二进制掩模的分离结果，网络输出是验证集为混合音频后的后20秒中男性说话者预测的理想软掩膜。

📚2 运行结果

主函数部分代码：

firstTrainingAudioFile   = "f.mp3";

secondTrainingAudioFile = "m.mp3";

C=1;%用于强化背景音乐

firstsongTrain   = C*audioread(firstTrainingAudioFile);

secondsongTrain = audioread(secondTrainingAudioFile);

L=500000;%采样值

firstsongTrain   = firstsongTrain(L:2*L);

secondsongTrain = secondsongTrain(L:2*L);%训练集

firstValidationAudioFile   ="f.mp3";

secondValidationAudioFile =  "m.mp3";

firstsongValidate   = C*audioread(firstValidationAudioFile);

secondsongValidate = audioread(secondValidationAudioFile);

 L1=1000000;

firstsongValidate   = firstsongValidate(3*L1:4*L1);

secondsongValidate = secondsongValidate(2.5*L1:3.5*L1);%验证级

% 将训练信号缩放到相同的功率。将验证信号缩放到相同的功率。

firstsongTrain  =firstsongTrain/norm(firstsongTrain);%训练集,

secondsongTrain = secondsongTrain/norm(secondsongTrain);

firstsongValidate  = firstsongValidate/norm(firstsongValidate);%验证集

secondsongValidate = secondsongValidate/norm(secondsongValidate);

mixTrain = firstsongTrain + secondsongTrain;

mixTrain = mixTrain / max(mixTrain);

mixValidate = firstsongValidate + secondsongValidate;

mixValidate = mixValidate / max(mixValidate);

WindowLength  = 128;

FFTLength     = 128;

OverlapLength = 128-1;

Fs            = 44000;

win           = hann(WindowLength,"periodic");

audiowrite('est_mix.wav',mixValidate,Fs);

P_mix0 = stft(mixTrain,'Window',win,'OverlapLength',OverlapLength,'FFTLength',FFTLength);

P_f    = abs(stft(firstsongTrain,'Window',win,'OverlapLength',OverlapLength,'FFTLength',FFTLength));

P_s    = abs(stft(secondsongTrain,'Window',win,'OverlapLength',OverlapLength,'FFTLength',FFTLength));

N      = 1 + FFTLength/2;

🎉3 参考文献

[1]鲁玉军,周世豪,胡小勇.基于BP神经网络和小波神经网络的太阳辐射强度预测[J].软件工程,2023,26(01):5-8+4.DOI:10.19644/j.cnki.issn2096-1472.2023.001.002.

👨‍💻4 Matlab代码