dataFolder = tempdir;数据集= fullfile (dataFolder,“Emo-DB”）;如果~ datasetExists(集)的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

广告= audioDatastore (fullfile(数据集,“wav”））;

filepaths = ads.Files;emotionCodes = cellfun (@ (x) x(录得5个),filepaths, UniformOutput = false);情感=取代(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) filepaths, UniformOutput = false);labelTable = cell2table (speakerCodes、情感,VariableNames = [“议长”，“情感”]);labelTable。情感= categorical(labelTable.Emotion); labelTable.Speaker = categorical(labelTable.Speaker); summary(labelTable)

ads.Labels = labelTable;

downloadFolder = matlab.internal.examples.downloadSupportFile (“音频”，“SpeechEmotionRecognition.zip”）;dataFolder = tempdir;unzip(downloadFolder,dataFolder) netFolder = fullfile(dataFolder，“SpeechEmotionRecognition”）;负载(fullfile (netFolder“network_Audio_SER.mat”））;

fs = afe.SampleRate;

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

特点=(提取(afe、音频))';featuresNormalized = (features - normalizers.Mean)./normalizers.StandardDeviation;numOverlap =10；featureSequences = HelperFeatureVector2Sequence (featuresNormalized 20 numOverlap);

YPred =双(预测(网络,featureSequences));平均=“模式”；开关平均情况下“的意思是”聚合氯化铝=意味着(YPred, 1);情况下“中值”聚合氯化铝=值(YPred, 1);情况下“模式”聚合氯化铝=模式(YPred, 1);结束派(probs. /笔(聚合氯化铝),字符串(net.Layers(结束). class))

numAugmentations =50；增量= audioDataAugmenter (NumAugmentations = NumAugmentations,.．.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 =元素个数(ads.Files) * numAugmentations;reset(ads) numPartitions = 18;抽搐parforadsPart = partition(ads,numPartitions,ii);而hasdata(adsPart) [x,adsInfo] = read(adsPart);data =增加(增压器,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(我);如果numel(augNum)==1 iString = [' 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 (“增强完成”+ (toc / 60, 2) +“分钟”。）

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

赢得=汉明(圆(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 =高(adsTrain);

featuresTallTrain = cellfun (@ (x)提取(afe x), tallTrain, UniformOutput = false);featuresTallTrain = cellfun (@ (x) x ', featuresTallTrain, UniformOutput = false);featuresTrain =收集(featuresTallTrain);

allFeatures =猫(2,featuresTrain {:});M =意味着(allFeatures 2“omitnan”）;S =性病(allFeatures 0 2,“omitnan”）;featuresTrain = cellfun (@ (x)(即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) fulllyconnectedlayer (numl (categories(emptyEmotions))) softmaxLayer classificationLayer];

miniBatchSize = 512;initialLearnRate = 0.005;learnRateDropPeriod = 2;maxEpochs = 3;选择= 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”）;

网= trainNetwork (sequencesTrain、labelsTrain层,选择);

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

演讲者= ads.Labels.Speaker;numFolds =元素个数(扬声器);总结(扬声器)

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

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

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

函数[序列,sequencePerFile] = HelperFeatureVector2Sequence(特性、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 =细胞(元素个数(特性),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[trueLabelsCrossFold, predictedLabelsCrossFold] = HelperTrainAndValidateNetwork(变长度输入宗量)版权所有2019 The MathWorks, Inc.如果Nargin == 3 ads = varargin{1};augads =变长度输入宗量{2};器=变长度输入宗量{3};elseifNargin == 2 ads = varargin{1};augads =变长度输入宗量{1};器=变长度输入宗量{2};结束演讲者=类别(ads.Labels.Speaker);numFolds =元素个数(扬声器);emptyEmotions = (ads.Labels.Emotion);emptyEmotions (:) = [];%对每个折叠进行循环。trueLabelsCrossFold = {};predictedLabelsCrossFold = {};为我= 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 ~ =议长(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。重新定位的功能这样时间是顺着行来兼容的。% sequenceInputLayer。tallTrain = cellfun (@ x (x) / max (abs (x)、[]“所有”)、tallTrain UniformOutput = false);tallFeaturesTrain = cellfun (@ (x)提取(萃取器,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)提取(萃取器,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。利用训练集确定均值和标准每个特征的%偏差。规范化培训和验证%设置。allFeatures =猫(2,featuresTrain {:});M =意味着(allFeatures 2“omitnan”）;S =性病(allFeatures 0 2,“omitnan”）;featuresTrain = cellfun (@ (x)(即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)(即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，使每个序列包含20个10个特征向量重叠的%特征向量。featureVectorsPerSequence = 20;featureVectorOverlap = 10;[sequencesTrain, sequencePerFileTrain] = HelperFeatureVector2Sequence (featuresTrain、featureVectorsPerSequence featureVectorOverlap);[sequencesValidation, sequencePerFileValidation] = HelperFeatureVector2Sequence (featuresValidation、featureVectorsPerSequence featureVectorOverlap);% 5。复制训练集和验证集的标签，以便它们与序列一一对应。。labelsTrain = [emptyEmotions; augadsTrain.Labels];labelsTrain = labelsTrain (:);labelsTrain = repelem (labelsTrain [sequencePerFileTrain {:}));% 6。定义BiLSTM网络。dropoutProb1 = 0.3;numUnits = 200;dropoutProb2 = 0.6;层= [.．.sequenceInputLayer(大小(sequencesTrain {1}, 1)) dropoutLayer dropoutProb1 bilstmLayer (numUnits OutputMode =“最后一次”) dropoutLayer(dropoutProb2) fulllyconnectedlayer (numl (categories(emptyEmotions))) softmaxLayer classificationLayer];% 7。定义培训选项。miniBatchSize = 512;initialLearnRate = 0.005;learnRateDropPeriod = 2;maxEpochs = 3;选择= 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。网= trainNetwork (sequencesTrain、labelsTrain层,选择);% 9。评估网络。调用classification来获得预测的标签%为每个序列。获取每个预测标签的模式%序列来获取每个文件的预测标签。predictedLabelsPerSequence =分类(净,sequencesValidation);trueLabels =分类(adsValidation.Labels);predictedLabels = trueLabels;idx1 = 1;为ii = 1:numel(trueLabels) predictedLabels(ii，:) = mode(predictedLabelsPerSequence(idx1:idx1 + sequenceperfilevalination {ii} - 1，:)，1);idx1 = idx1 + sequenceperfilevalalization {ii};结束trueLabelsCrossFold{我}= trueLabels;% #好< AGROW >predictedLabelsCrossFold{我}= predictedLabels;% #好< AGROW >结束结束