语音数字识别与小波散射和深度学习

pathToRecordingsFolder = fullfile (tempdir,“free-spoken-digit-dataset”,“录音”);位置= pathToRecordingsFolder;

广告= audioDatastore(位置);

ads.Labels = helpergenLabels(广告);总结(ads.Labels)

LenSig = 0(元素个数(ads.Files), 1);nr = 1;而hasdata(广告)数字=阅读(广告);LenSig (nr) =元素个数(数字);nr = nr + 1;结束重置(广告)直方图(LenSig)网格在包含(的信号长度(样本))ylabel (“频率”)

科幻小说= waveletScattering (“SignalLength”,8192,“InvarianceScale”,0.22,…“SamplingFrequency”,8000,“OversamplingFactor”2);

rng默认的;广告= shuffle(广告);[adsTrain, adsTest] = splitEachLabel(广告,0.8);countEachLabel (adsTrain)

countEachLabel (adsTest)

Xtrain = [];scatds_Train =变换(adsTrain @ (x) helperReadSPData (x));而hasdata (scatds_Train) smat =阅读(scatds_Train);Xtrain =猫(2 Xtrain smat);结束

Xtest = [];scatds_Test =变换(adsTest @ (x) helperReadSPData (x));而hasdata (scatds_Test) smat =阅读(scatds_Test);Xtest =猫(2 Xtest smat);结束

应变= sf.featureMatrix (Xtrain);圣= sf.featureMatrix (Xtest);

TrainFeatures =应变(2:,:,);TrainFeatures =挤压(平均(TrainFeatures, 2))的;TestFeatures =圣(2:,:,);TestFeatures =挤压(平均(TestFeatures, 2))的;

模板= templateSVM (…“KernelFunction”,多项式的,…“PolynomialOrder”2,…“KernelScale”,“汽车”,…“BoxConstraint”,1…“标准化”,真正的);classificationSVM = fitcecoc (…TrainFeatures,…adsTrain.Labels,…“学习者”模板,…“编码”,“onevsone”,…“类名”分类({' 0 ';' 1 ';' 2 ';“3”;“4”;“5”;“6”;“7”;“8”;“9”}));

partitionedModel = crossval (classificationSVM,“KFold”5);[validationPredictions, validationScores] = kfoldPredict (partitionedModel);validationAccuracy = (1 - kfoldLoss (partitionedModel“LossFun”,“ClassifError”))* 100

predLabels =预测(classificationSVM TestFeatures);testAccuracy = (predLabels = = adsTest.Labels) /元素个数之和(predLabels) * 100

图(“单位”,“归一化”,“位置”(0.2 - 0.2 0.5 - 0.5));ccscat = confusionchart (adsTest.Labels predLabels);ccscat。Title =“小波散射分类”;ccscat。ColumnSummary =“column-normalized”;ccscat。RowSummary =“row-normalized”;

TrainFeatures =应变(2:,:,);TrainFeatures =挤压(num2cell (TrainFeatures [1 - 2]));TestFeatures =圣(2:,:,);TestFeatures =挤压(num2cell (TestFeatures [1 - 2]));

[inputSize ~] =大小(TrainFeatures {1});YTrain = adsTrain.Labels;numHiddenUnits = 512;numClasses =元素个数(独特(YTrain));层= […sequenceInputLayer inputSize lstmLayer (numHiddenUnits,“OutputMode”,“最后一次”)fullyConnectedLayer (numClasses) softmaxLayer classificationLayer];

maxEpochs = 300;miniBatchSize = 50;选择= trainingOptions (“亚当”,…“InitialLearnRate”,0.0001,…“MaxEpochs”maxEpochs,…“MiniBatchSize”miniBatchSize,…“SequenceLength”,“最短”,…“洗牌”,“every-epoch”,…“详细”假的,…“阴谋”,“训练进步”);

网= trainNetwork (TrainFeatures、YTrain层,选择);

predLabels =分类(净,TestFeatures);testAccuracy = (predLabels = = adsTest.Labels) /元素个数之和(predLabels) * 100

YTrain = adsTrain.Labels;欧美= adsTest.Labels;如果~ (”结果/”,“dir”mkdir)结果结束

optVars = [optimizableVariable (“InitialLearnRate”(1 e-5, 1 e 1),“转换”,“日志”)optimizableVariable (“NumHiddenUnits”(1000),“类型”,“整数”));

ObjFcn = helperBayesOptLSTM (TrainFeatures YTrain TestFeatures,欧美);optimizeCondition = false;如果optimizeCondition BayesObject = bayesopt (ObjFcn optVars,…“MaxObjectiveEvaluations”15岁的…“IsObjectiveDeterministic”假的,…“UseParallel”,真正的);其他的url =“http://ssd.mathworks.com/万博1manbetxsupportfiles/audio/SpokenDigitRecognition.zip”;downloadNetFolder = tempdir;netFolder = fullfile (downloadNetFolder,“SpokenDigitRecognition”);如果~存在(netFolder“dir”)disp (文件下载pretrained网络(1 - 12 MB)……”解压缩(url, downloadNetFolder)结束负载(fullfile (netFolder“0.02.mat”));结束

numHiddenUnits = 768;numClasses =元素个数(独特(YTrain));层= […sequenceInputLayer inputSize lstmLayer (numHiddenUnits,“OutputMode”,“最后一次”)fullyConnectedLayer (numClasses) softmaxLayer classificationLayer];maxEpochs = 300;miniBatchSize = 50;选择= trainingOptions (“亚当”,…“InitialLearnRate”2.198827960269379 e-04…“MaxEpochs”maxEpochs,…“MiniBatchSize”miniBatchSize,…“SequenceLength”,“最短”,…“洗牌”,“every-epoch”,…“详细”假的,…“阴谋”,“训练进步”);网= trainNetwork (TrainFeatures、YTrain层,选择);predLabels =分类(净,TestFeatures);testAccuracy = (predLabels = = adsTest.Labels) /元素个数之和(predLabels) * 100

segmentDuration = 8192 * (1/8000);frameDuration = 0.22;hopDuration = 0.01;numBands = 40;

重置(adsTrain);重置(adsTest);

epsil = 1 e-6;XTrain = helperspeechSpectrograms (adsTrain segmentDuration、frameDuration hopDuration, numBands);

XTrain = log10 (XTrain + epsil);XTest = helperspeechSpectrograms (adsTest segmentDuration、frameDuration hopDuration, numBands);

XTest = log10 (XTest + epsil);YTrain = adsTrain.Labels;欧美= adsTest.Labels;

深圳=大小(XTrain);specSize =深圳(1:2);图象尺寸= [specSize 1];numClasses =元素个数(类别(YTrain));dropoutProb = 0.2;numF = 12;层= [imageInputLayer(图象尺寸)convolution2dLayer (5 numF“填充”,“相同”)batchNormalizationLayer reluLayer maxPooling2dLayer (3“步”2,“填充”,“相同”)convolution2dLayer (3 2 * numF“填充”,“相同”)batchNormalizationLayer reluLayer maxPooling2dLayer (3“步”2,“填充”,“相同”)convolution2dLayer (3、4 * numF,“填充”,“相同”)batchNormalizationLayer reluLayer maxPooling2dLayer (3“步”2,“填充”,“相同”)convolution2dLayer (3、4 * numF,“填充”,“相同”)batchNormalizationLayer reluLayer convolution2dLayer (3、4 * numF,“填充”,“相同”)batchNormalizationLayer reluLayer maxPooling2dLayer (2) dropoutLayer (dropoutProb) fullyConnectedLayer (numClasses) softmaxLayer classificationLayer (“类”类别(YTrain));];

miniBatchSize = 50;选择= trainingOptions (“亚当”,…“InitialLearnRate”1的军医,…“MaxEpochs”30岁的…“MiniBatchSize”miniBatchSize,…“洗牌”,“every-epoch”,…“阴谋”,“训练进步”,…“详细”假的,…“ExecutionEnvironment”,“cpu”);

trainedNet = trainNetwork (XTrain、YTrain层,选择);

[Ypredicted,聚合氯化铝]=分类(trainedNet XTest,“ExecutionEnvironment”,“CPU”);cnnAccuracy = (Ypredicted = =次)/元素个数之和(欧美)* 100

图(“单位”,“归一化”,“位置”(0.2 - 0.2 0.5 - 0.5));ccDCNN = confusionchart(欧美,Ypredicted);ccDCNN。Title =“DCNN混乱图”;ccDCNN。ColumnSummary =“column-normalized”;ccDCNN。RowSummary =“row-normalized”;

函数标签= helpergenLabels(广告)%这个函数只用于小波工具箱的例子。它可能是%在将来发布的版本中修改或删除。tmp =细胞(元素个数(ads.Files), 1);表达=“[0 - 9]+ _”;为nf = 1:元素个数(ads.Files) idx = regexp (ads.Files {nf},表达式);tmp {nf} = ads.Files {nf} (idx);结束标签=分类(tmp);结束

函数x = helperReadSPData (x)%这个函数只用于小波工具箱的例子。它可能改变或%在将来的版本中被删除。N =元素个数(x);如果N > 8192 x = x (1:8192);elseifN < 8192垫= 8192 - N;前置液=地板(垫/ 2);postpad =装天花板(垫/ 2);x =[0(前置液,1);x;0 (postpad 1)];结束x = x / max (abs (x));结束

函数x = helperBayesOptLSTM (X_train Y_train、X_val Y_val)%这个函数只用于%”语音数字识别与小波散射和深度学习”%的例子。它可能改变或在将来的版本中被删除。x = @valErrorFun;函数[valError,缺点,文件名]= valErrorFun (optVars)% % LSTM架构[inputSize ~] =大小(X_train {1});numClasses =元素个数(独特(Y_train));层= […sequenceInputLayer inputSize bilstmLayer (optVars.NumHiddenUnits,“OutputMode”,“最后一次”)%使用的隐藏层价值优化变量fullyConnectedLayer (numClasses) softmaxLayer classificationLayer];%的情节不显示在训练选择= trainingOptions (“亚当”,…“InitialLearnRate”optVars.InitialLearnRate,…%利用优化变量的初始学习速率值“MaxEpochs”,300,…“MiniBatchSize”30岁的…“SequenceLength”,“最短”,…“洗牌”,“永远”,…“详细”、假);% %训练网络网= trainNetwork (X_train、Y_train层,选择);% %训练精度X_val_P = net.classify (X_val);accuracy_training = (X_val_P = = Y_val)。/元素个数(Y_val);valError = 1 - accuracy_training;% %保存结果的网络和期权在垫文件在文件夹和错误值的结果文件名= fullfile (“结果”num2str (valError) +“.mat”);保存(文件名,“净”,“valError”,“选项”)缺点= [];结束%对内部函数结束%为外部函数结束

函数X = helperspeechSpectrograms(广告、segmentDuration frameDuration、hopDuration numBands)%这个函数只用于%”语音数字识别与小波散射和深度学习”%的例子。它可能改变或在将来的版本中被删除。%% helperspeechSpectrograms(广告、segmentDuration frameDuration、hopDuration numBands)%计算语音谱图数据存储文件的广告。% segmentDuration是演讲片段的总时间(以秒为单位),% frameDuration每个光谱图帧的持续时间,hopDuration每个光谱图帧之间%时移,numBands的数量%频段。disp (“语音谱图计算……”);numHops =装天花板((segmentDuration - frameDuration) / hopDuration);numFiles =长度(ads.Files);X = 0 ([numBands numHops 1, numFiles],“单一”);为i = 1: numFiles [x,信息]=阅读(广告);x = normalizeAndResize (x);fs = info.SampleRate;frameLength =圆(frameDuration * fs);hopLength =圆(hopDuration * fs);规范= melSpectrogram (x, fs,…“窗口”汉明(frameLength“周期”),…“OverlapLength”frameLength - hopLength…“FFTLength”,2048,…“NumBands”numBands,…“FrequencyRange”[4000]);%如果不如numHops宽光谱图,然后把谱图% X的中间。w =大小(规范,2);地板左= ((numHops-w) / 2) + 1;印第安纳州=左:左+ w1;X(:,印第安纳州,1,i) =规范;如果国防部(我500)= = 0 disp (“加工”+我+“文件”+ numFiles)结束结束disp (“…”);结束% - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -函数x = normalizeAndResize (x)%这个函数只用于%”语音数字识别与小波散射和深度学习”%的例子。它可能改变或在将来的版本中被删除。N =元素个数(x);如果N > 8192 x = x (1:8192);elseifN < 8192垫= 8192 - N;前置液=地板(垫/ 2);postpad =装天花板(垫/ 2);x =[0(前置液,1);x;0 (postpad 1)];结束x = x / max (abs (x));结束