dataFolder = tempdir;dataset = fullfile(数据文件夹，“Emo-DB”）;如果~datasetExists(dataset) url =“http://emodb.bilderbar.info/download/download.zip”；disp ("下载Emo-DB (40.5 MB)…"解压缩(url,数据集)结束

ads = audioDatastore(fullfile(dataset，“wav”));

fileppaths = ads.Files;emotionCodes = cellfun(@(x)x(end-5)， fileppaths,UniformOutput=false);emotions = replace(emotionCodes，[“W”，“L”，“E”，“一个”，“F”，“T”，“N”),.．.［“愤怒”，“无聊”，“厌恶”，“焦虑和恐惧”，“幸福”，“悲伤”，“中性”]);speakerCodes = cellfun(@(x)x(end-10:end-9)， fileppaths,UniformOutput=false);labelTable = cell2table([speakerCodes,emotions]，VariableNames=[“议长”，“情感”]);labelTable。情感= categorical(labelTable.Emotion); labelTable.Speaker = categorical(labelTable.Speaker); summary(labelTable)

ads.Labels =标签表;

downloadFolder = matlab.internal.examples.download金宝appSupportFile(“音频”，“SpeechEmotionRecognition.zip”）;dataFolder = tempdir;unzip(下载文件夹，数据文件夹)netFolder = fullfile(数据文件夹，数据文件夹)“SpeechEmotionRecognition”）;负载(fullfile (netFolder“network_Audio_SER.mat”));

fs = afe.SampleRate;

演讲者=分类(“03”)；情感=分类(“厌恶”)；ads子集=子集(ads,ads. labels。演讲者==speaker & ads.Labels.Emotion==emotion); audio = read(adsSubset); sound(audio,fs)

Features = (extract(afe,audio))';featuresNormalized = (features - normalizers.Mean)./normalizers.标准差;numOverlap =10；featureSequences = HelperFeatureVector2Sequence(featuresNormalized,20,numOverlap);

YPred = double(predict(net,featureSequences));平均=“模式”；开关平均情况下“的意思是”probs = mean(YPred,1);情况下“中值”probs = median(YPred,1);情况下“模式”probs = mode(YPred,1);结束派(probs. /笔(聚合氯化铝),字符串(net.Layers(结束). class))

numAugmentations =50；augmenter = audioDataAugmenter(NumAugmentations= numaugments，.．.TimeStretchProbability = 0,.．.VolumeControlProbability = 0,.．..．.PitchShiftProbability = 0.5,.．..．.TimeShiftProbability = 1,.．.TimeShiftRange = [-0.3, 0.3],.．..．.AddNoiseProbability = 1,.．.SNRRange = [-20, 40]);

currentDir = pwd;writeDirectory = fullfile(currentDir，“augmentedData”）;mkdir (writeDirectory)

N = nummel (ads.Files)* numaugations;reset(ads) numPartitions = 18;抽搐parforadsPart =分区(ads,numPartitions,ii);而hasdata(adsPart) [x,adsInfo] = read(adsPart);数据=增强(增强器，x,fs);[~，fn] = fileparts(adsInfo.FileName);为i = 1:size(data,1) augmentedAudio = data. audio {i};augmentedAudio = augmentedAudio/max(abs(augmentedAudio)，[]，“所有”）;augNum = num2str(i);如果数字(augNum)==1' 0 'augNum);其他的iString = augNum;结束audiowrite (fullfile (writeDirectory sprintf (“% s_aug % s.wav”、fn iString))、augmentedAudio fs);结束结束结束disp (“扩展完成”+ round(toc/60,2) +“分钟”。)

adsAug = audioDatastore(writeDirectory);adsAug。标签= repelem(ads.Labels,augmenter.NumAugmentations,1);

赢=汉明(轮(0.03*fs)，“周期”）;overlapLength = 0;afe = audioFeatureExtractor(.．.窗口=赢,.．.OverlapLength = OverlapLength,.．.SampleRate = fs,.．..．.gtcc = true,.．.gtccDelta = true,.．.mfccDelta = true,.．..．.SpectralDescriptorInput =“melSpectrum”，.．.spectralCrest = true);

adsTrain = adsAug;

tallTrain = tall(adsTrain);

featuresTallTrain = cellfun(@(x)extract(afe,x)，tallTrain,UniformOutput=false);featuresTallTrain = cellfun(@(x)x'，featuresTallTrain,UniformOutput=false);featuresTrain = gather(featurerestalltrain);

allFeatures = cat(2,featuresTrain{:});M = mean(allFeatures,2，“omitnan”）;S = std(allFeatures,0,2，“omitnan”）;featuresTrain = cellfun(@(x)(x- m)./S,featuresTrain,UniformOutput=false);

featureVectorsPerSequence = 20;featureVectorOverlap = 10;[sequencesTrain,sequencePerFileTrain] = HelperFeatureVector2Sequence(featuresTrain,featureVectorsPerSequence,featureVectorOverlap);

labelsTrain = repelem(adsTrain.Labels.Emotion，[sequencePerFileTrain{:}]);emptyEmotions = ads.Labels.Emotion;emptyEmotions(:) = [];

dropoutProb1 = 0.3;numUnits = 200;dropoutProb2 = 0.6;层= [.．.sequenceInputLayer(afe.FeatureVectorLength) dropoutLayer(dropoutProb1) bilstmLayer(numUnits,OutputMode=“最后一次”) dropoutLayer(dropoutProb2) fullyConnectedLayer(nummel (categories(emptyEmotions))) softmaxLayer classificationLayer];

miniBatchSize = 512;initialLearnRate = 0.005;learnratdropperiod = 2;maxEpochs = 3;选项= trainingOptions(“亚当”，.．.MiniBatchSize = MiniBatchSize,.．.InitialLearnRate = InitialLearnRate,.．.LearnRateDropPeriod = LearnRateDropPeriod,.．.LearnRateSchedule =“分段”，.．.MaxEpochs = MaxEpochs,.．.洗牌=“every-epoch”，.．.Verbose = false,.．.情节=“训练进步”）;

net = trainNetwork(sequencesTrain,labelsTrain,layers,options);

saveSERSystem =假；如果saveSERSystem标准化者。均值= M;标准化者。标准差= S;保存(“network_Audio_SER.mat”，“净”，“安全的”，“标准化者”)结束

speaker = ads.Labels.Speaker;numFolds = nummel(扬声器);总结(扬声器)

[labelsTrue,labelsPred] = HelperTrainAndValidateNetwork(ads,adsAug,afe);

为ii = 1:数字(labelsTrue) foldAcc = mean(labelsTrue{ii}==labelsPred{ii})*100;disp (“折”+ ii +，精度= "+圆(foldAcc, 2))结束

labelsTrueMat = cat(1,labelsTrue{:});labelsPredMat = cat(1,labelsPred{:});figure cm = confusichart (labelsTrueMat,labelsPredMat，.．.Title = (10倍交叉验证的混淆矩阵，“平均准确度=”圆的(意味着(labelsTrueMat = = labelsPredMat) * 100, 1)),.．.ColumnSummary =“column-normalized”RowSummary =“row-normalized”）;sortClasses(厘米、类别(emptyEmotions))

函数[sequences,sequencePerFile] = HelperFeatureVector2Sequence(features,featureVectorsPerSequence,featureVectorOverlap)版权所有2019 MathWorks, Inc.如果featureVectorsPerSequence <= featureVectorOverlap错误(“重叠特征向量的数量必须小于每个序列的特征向量的数量。”)结束如果~iscell(features) features = {features};结束hopLength = featureVectorsPerSequence - featureVectorOverlap;Idx1 = 1;序列= {};sequencePerFile = cell(数字(特征)，1);为ii = 1:numel(features) sequencePerFile{ii} = floor((size(features{ii}，2) - featureVectorsPerSequence)/hopLength) + 1;Idx2 = 1;为j = 1:sequencePerFile{ii} sequences{idx1,1} = features{ii}(:，idx2:idx2 + featureVectorsPerSequence - 1);% #好< AGROW >Idx1 = Idx1 + 1;idx2 = idx2 + hopLength;结束结束结束

函数[truelabelscrosfold, predictedlabelscrosfold] = HelperTrainAndValidateNetwork(varargin)The MathWorks, Inc.版权所有如果Nargin == 3 ads = varargin{1};Augads = varargin{2};Extractor = varargin{3};elseifNargin == 2 ads = varargin{1};Augads = varargin{1};Extractor = varargin{2};结束speaker = categories(ads.Labels.Speaker);numFolds = nummel(扬声器);emptyEmotions = (ads.Labels.Emotion);emptyEmotions(:) = [];%循环每一个折叠。truelabelcrossfold = {};predictedlabelcrosssfold = {};为i = 1:numFolds% 1。将音频数据存储分为训练集和验证集。将数据转换为高数组。idxTrain = augads.Labels.Speaker~=speaker(i);augadsTrain =子集(augads,idxTrain);augadsTrain。标签= augadsTrain.Labels.Emotion; tallTrain = tall(augadsTrain); idxValidation = ads.Labels.Speaker==speaker(i); adsValidation = subset(ads,idxValidation); adsValidation.Labels = adsValidation.Labels.Emotion; tallValidation = tall(adsValidation);% 2。从训练集中提取特征。重新定位功能%，使时间与沿线行相兼容% sequenceInputLayer。tallTrain = cellfun(@(x)x/max(abs(x)，[]，“所有”)、tallTrain UniformOutput = false);tallFeaturesTrain = cellfun(@(x)extract(extractor,x)，tallTrain,UniformOutput=false);tallFeaturesTrain = cellfun(@(x)x'，tallFeaturesTrain,UniformOutput=false);% #好< NASGU >[~，featuresTrain] = evalc(“收集(tallFeaturesTrain)”）;%使用evalc抑制命令行输出。tallValidation = cellfun(@(x)x/max(abs(x)，[]，“所有”)、tallValidation UniformOutput = false);tallFeaturesValidation = cellfun(@(x)extract(extractor,x)，tallValidation，“UniformOutput”、假);tallFeaturesValidation = cellfun(@(x)x'，tallFeaturesValidation,UniformOutput=false);% #好< NASGU >[~，featuresValidation] = evalc(“收集(tallFeaturesValidation)”）;%使用evalc抑制命令行输出。% 3。使用训练集来确定平均值和标准每个特征的%偏差。规范化培训和验证%设置。allFeatures = cat(2,featuresTrain{:});M = mean(allFeatures,2，“omitnan”）;S = std(allFeatures,0,2，“omitnan”）;featuresTrain = cellfun(@(x)(x- m)./S,featuresTrain,UniformOutput=false);为ii = 1:numel(featuresTrain) idx = find(isnan(featuresTrain{ii}));如果~isempty(idx) featuresTrain{ii}(idx) = 0;结束结束featuresValidation = cellfun(@(x)(x- m)./S,featuresValidation,UniformOutput=false);为ii = 1:numel(featuresValidation) idx = find(isnan(featuresValidation{ii}));如果~isempty(idx) featuresValidation{ii}(idx) = 0;结束结束% 4。缓冲序列，使每个序列由20个组成%特征向量与10个特征向量重叠。featureVectorsPerSequence = 20;featureVectorOverlap = 10;[sequencesTrain,sequencePerFileTrain] = HelperFeatureVector2Sequence(featuresTrain,featureVectorsPerSequence,featureVectorOverlap);[sequencesValidation,sequencePerFileValidation] = HelperFeatureVector2Sequence(featuresValidation,featureVectorsPerSequence,featureVectorOverlap);% 5。复制训练和验证集的标签，以便它们与序列一一对应。labelsTrain = [emptyEmotions;augadsTrain.Labels];labelsTrain = labelsTrain(:);labelsTrain = repelem(labelsTrain，[sequencePerFileTrain{:}]);% 6。定义一个BiLSTM网络。dropoutProb1 = 0.3;numUnits = 200;dropoutProb2 = 0.6;层= [.．.sequenceInputLayer(size(sequencesTrain{1}，1)) dropoutLayer(dropoutProb1) bilstmLayer(numUnits,OutputMode=“最后一次”) dropoutLayer(dropoutProb2) fullyConnectedLayer(nummel (categories(emptyEmotions))) softmaxLayer classificationLayer];% 7。定义培训选项。miniBatchSize = 512;initialLearnRate = 0.005;learnratdropperiod = 2;maxEpochs = 3;选项= trainingOptions(“亚当”，.．.MiniBatchSize = MiniBatchSize,.．.InitialLearnRate = InitialLearnRate,.．.LearnRateDropPeriod = LearnRateDropPeriod,.．.LearnRateSchedule =“分段”，.．.MaxEpochs = MaxEpochs,.．.洗牌=“every-epoch”，.．.Verbose = false);% 8。培训网络。net = trainNetwork(sequencesTrain,labelsTrain,layers,options);% 9。评估网络。调用classification来获得预测的标签%为每个序列。得到每个预测标签的模式%序列来获得每个文件的预测标签。predictedLabelsPerSequence =分类(net,sequencesValidation);trueLabels = categorical(adsValidation.Labels);predictedLabels = trueLabels;Idx1 = 1;为ii = 1:numel(trueLabels) predictedLabels(ii，:) = mode(predictedLabelsPerSequence(idx1:idx1 + sequencePerFileValidation{ii} - 1，:)，1);idx1 = idx1 + sequencePerFileValidation{ii};结束truelabelcrossfold {i} = trueLabels;% #好< AGROW >predictedlabelcrosssfold {i} = predictedLabels;% #好< AGROW >结束结束