说话者识别使用自定义SincNet层更深的学习

dataFolder = tempdir;数据集= fullfile (dataFolder,“LibriSpeech”,“培训-清洗- 100”);如果~ datasetExists(集)文件名=“火车——清洁- 100. tar.gz”;url =“http://www.openSLR.org/resources/12/”+文件名;gunzip (url, dataFolder);unzippedFile = fullfile(数据集,文件名);解压(unzippedFile {1} (1: end-3),数据集);结束

广告= audioDatastore(数据集,IncludeSubfolders = true);

ads.Labels =分类(extractBetween (ads.Files fullfile(数据集,filesep) filesep));

speedupExample =假;如果speedupExample allSpeakers =独特(ads.Labels);subsetSpeakers = allSpeakers (1:50);广告=子集(广告,ismember (ads.Labels subsetSpeakers));ads.Labels = removecats (ads.Labels);结束广告= splitEachLabel(广告,0.1);

[adsTrain, adsTest] = splitEachLabel(广告,0.8);

[audioIn, dsInfo] =阅读(adsTrain);Fs = dsInfo.SampleRate;声音(audioIn Fs) t = (1 / Fs) *(0:长度(audioIn) 1);情节(t, audioIn)标题(“音频样本”)包含(“时间(s)”)ylabel (“振幅”网格)在

重置(adsTrain)

frameDuration = 200 e - 3;overlapDuration = 40 e - 3;frameLength =地板(Fs * frameDuration);overlapLength =圆(Fs * overlapDuration);

= ~ isempty(版本(pFlag“平行”));adsTrainTransform =变换(adsTrain @ (x) {preprocessAudioData (x, frameLength overlapLength Fs)});XTrain = readall (pFlag adsTrainTransform UseParallel =);

chunksPerFile = cellfun (@ (x)大小(x, 4), XTrain);TTrain = repelem (adsTrain.Labels chunksPerFile 1);

XTrain =猫(4,XTrain {:});

adsTestTransform =变换(adsTest @ (x) {preprocessAudioData (x, frameLength overlapLength Fs)});XTest = readall (adsTestTransform UseParallel = true);chunksPerFile = cellfun (@ (x)大小(x, 4), XTest);tt = repelem (adsTest.Labels chunksPerFile 1);XTest =猫(4 XTest {:});

numFilters = 80;filterLength = 251;numSpeakers =元素个数(独特(removecats (ads.Labels)));层= [imageInputLayer ([1 frameLength 1])%的第一个回旋的层numFilters filterLength convolution2dLayer ([1]) batchNormalizationLayer leakyReluLayer (0.2) maxPooling2dLayer (1 [3])%这一层是紧随其后的是2卷积层convolution2dLayer ([1 - 5], 60) batchNormalizationLayer leakyReluLayer (0.2) maxPooling2dLayer ([1 3]) convolution2dLayer ([1 - 5], 60) batchNormalizationLayer leakyReluLayer (0.2) maxPooling2dLayer (1 [3])%这是紧随其后的是3全层fullyConnectedLayer (256) batchNormalizationLayer leakyReluLayer (0.2) fullyConnectedLayer (256) batchNormalizationLayer leakyReluLayer (0.2) fullyConnectedLayer (256) batchNormalizationLayer leakyReluLayer (0.2) fullyConnectedLayer (numSpeakers) softmaxLayer classificationLayer];

analyzeNetwork(层)

numEpochs = 15;miniBatchSize = 128;validationFrequency =地板(元素个数(TTrain) / miniBatchSize);选择= trainingOptions (“亚当”,…洗牌=“every-epoch”,…MiniBatchSize = MiniBatchSize,…情节=“训练进步”,…Verbose = false, MaxEpochs = numEpochs,…ValidationData = {XTest,直言(tt)},…ValidationFrequency = ValidationFrequency);

(事先convNetInfo] = trainNetwork (XTrain、TTrain层,选择);

F =挤压(convNet.Layers (2, 1) .Weights);H = 0(大小(F));频率= 0(大小(F));为2 = 1:尺寸(F, 2) [h F] = freqz (F(:,(二),1251年,Fs);H(:,(二)= abs (H);频率(:,(二)= f;结束idx = linspace(1、大小(F, 2), 9);idx =圆(idx);图为jj = 1:9次要情节(3,3,jj)情节(频率(:,idx (jj)), H (:, idx (jj))) sgtitle (“学会了标准的CNN过滤器的频率响应”)包含(“频率(赫兹)”)结束

numFilters = 80;filterLength = 251;numChannels = 1;name =“constant_sinc”;

cSL = constantSincLayer (numChannels numFilters filterLength, Fs,名称)

层(2)= cSL;

[constSincNet, constSincInfo] = trainNetwork (XTrain、TTrain层,选择);

图9 n =;plotNFilters (constSincNet.Layers (2), n)

numFilters = 80;filterLength = 251;numChannels = 1;name =“sinc”;

sNL = sincNetLayer (numChannels numFilters filterLength, Fs,名称)

层(2)= sNL;

[sincNet, sincNetInfo] = trainNetwork (XTrain、TTrain层,选择);

图plotNFilters (sincNet.Layers (2), 9)

NetworkType = [“标准CNN”;“恒Sinc层”;“SincNet层”];精度= [convNetInfo.FinalValidationAccuracy; constSincInfo.FinalValidationAccuracy sincNetInfo.FinalValidationAccuracy];resultsSummary =表(NetworkType、准确性)

时代= linspace (0 numEpochs元素个数(sincNetInfo.ValidationAccuracy (~ isnan (sincNetInfo.ValidationAccuracy))));时代=[时代,numEpochs];sinc_valAcc = [sincNetInfo.ValidationAccuracy (~ isnan (sincNetInfo.ValidationAccuracy)),…sincNetInfo.FinalValidationAccuracy];const_sinc_valAcc = [constSincInfo.ValidationAccuracy (~ isnan (constSincInfo.ValidationAccuracy)),…constSincInfo.FinalValidationAccuracy];conv_valAcc = [convNetInfo.ValidationAccuracy (~ isnan (convNetInfo.ValidationAccuracy)),…convNetInfo.FinalValidationAccuracy];图绘制(sinc_valAcc时代,“- *”MarkerSize = 4)在情节(const_sinc_valAcc时代,“- *”MarkerSize = 4)情节(conv_valAcc时代,“- *”MarkerSize = 4) ylabel (“框架水准仪准确性(测试集)”)包含(“时代”)xlim ([0 numEpochs + 0.3])标题(“框架水准仪精度与时代”)传说(“sincNet”,“constantSincLayer”,“conv2dLayer”位置=“东南”网格)在

函数xp = preprocessAudioData (x, frameLength overlapLength Fs) speechIdx = detectSpeech (x, Fs);xp = 0 (1 frameLength 1,0);为2 = 1:尺寸(speechIdx, 1)%隔离段演讲audioChunk = x (speechIdx (2, 1): speechIdx (2, 2));%分成200 ms的块audioChunk =缓冲区(audioChunk frameLength overlapLength);audioChunk =重塑(audioChunk 1 frameLength 1大小(audioChunk, 2));%与现有的音频连接xp =猫(4 xp audioChunk);结束结束