url ="http://emodb.bilderbar.info/download/download.zip"；downloadFolder=tempdir；datasetFolder=fullfile（downloadFolder，“Emo-DB”）;如果~存在(datasetFolder“dir”) disp ('正在下载Emo DB（40.5 MB）…')解压（url，数据集文件夹）结束

ads=音频数据存储（完整文件（数据集文件夹，“wav”））;

filepath=ads.Files；emotioncode=cellfun（@（x）x（end-5），filepath，“UniformOutput”、假);情感=取代(emotionCodes, {' W ',“L”,“E”,“一个”,“F”,“不”,“不”}，．..{“愤怒”,“无聊”,“厌恶”,焦虑和恐惧的,“幸福”,“悲伤”,“中性”})；speakerCodes=cellfun（@（x）x（end-10:end-9），文件路径，“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;

url ='http://ssd.mathworks.com/万博1manbetxsupportfiles/audio/SpeechEmotionRecognition.zip'；downloadNetFolder = tempdir;netFolder = fullfile (downloadNetFolder,“言语情感识别”）;如果~存在(netFolder“dir”) disp ('下载预训练网络(1文件- 1.5 MB)…'解压缩(url, downloadNetFolder)结束负载(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。发言者==speaker & ads.Labels.Emotion == emotion); audio = read(adsSubset); sound(audio,fs)

特点=(提取(afe、音频))';feature归一化= (features - normalizer . mean)./normalizer . standard deviation;numOverlap =10; featureSequences=HelperFeatureVector2Sequence（featuresNormalized，20，numOverlap）；

YPred =双(预测(网络,featureSequences));平均=“模式”；开关平均的情况下“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);结束饼图（probs./sum（probs），字符串（net.Layers（end.Classes））

numAugmentations =50; 增强器=音频数据增强器(“NumAugmentations”numAugmentations,．..“时间树概率”0,．..“VolumeControlProbability”0,．..．..“PitchShiftProbability”, 0.5,．..．..“TimeShiftProbability”,1，．..“TimeShiftRange”,[-0.3,0.3],．..．..“AddNoiseProbability”,1，．..“SNRRange”, -20年,40);

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

N =元素个数(ads.Files) * numAugmentations;myWaitBar = HelperPoolWaitbar (N,“正在扩充数据集…”); 重置（ads）数值=18；抽搐帕弗ii = 1:numPartitions adsPart = partition(ads,numPartitions,ii);而hasdata(adsPart) [x,adsInfo] = read(adsPart);data =增加(增压器,x, fs);[~, fn] = fileparts (adsInfo.FileName);对于i = 1:size(data,1) augmentedAudio = data. audio {i};augmentedAudio = augmentedAudio / max (abs (augmentedAudio), [],“所有”）;augNum = num2str(我);如果numel(augNum)==1'0'augNum);其他的iString = augNum;结束audiowrite（完整文件）（writeDirectory，sprintf(“% s_aug % s.wav”、fn iString))、augmentedAudio fs);增量(myWaitBar)结束结束结束

删除(myWaitBar)流('扩增完成(%0.2f分钟)'，toc/60）

adsagug=音频数据存储（writeDirectory）；adsagug.Labels=repelem（ads.Labels，augmenter.numumations，1）；

赢得=汉明(圆(0.03 * fs),“定期”); 重叠长度=0；afe=音频特征提取程序(．..“窗口”,赢了,．..“OverlapLength”，重叠长度，．..“SampleRate”fs,．..．..“gtcc”,真的,．..“gtccDelta”,真的,．..“mfccDelta”,真的,．..．..“SpectralDescriptorInput”,“梅尔斯波谱”,．..“spectralCrest”,真正的);

adsTrain = adsAug;

tallTrain =高(adsTrain);

featuresTallTrain=cellfun（@（x）提取物（afe，x），tallTrain，“统一输出”、假);featuresTallTrain = cellfun (@ (x) x ', featuresTallTrain,“统一输出”、假);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”、假);

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（大小（sequencesTrain{1}，1））DropOutputLayer（dropoutProb1）BilstLayer（numUnits，“OutputMode”,“最后一次”) dropoutLayer(dropoutProb2) fulllyconnectedlayer (numel(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”,．..“冗长”假的,．..“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);

对于ii = 1:numel(labeltrue) foldAcc = mean(labeltrue {ii}==labelsPred{ii})*100;流('折叠%1.0f，精度=%0.1f\n'，ii，foldac）；结束

labelsTrueMat =猫(1,labelsTrue {:});labelsPredMat =猫(1,labelsPred {:});figure cm = confusionchart(labelsTrueMat,labelsPredMat);valAccuracy =意味着(labelsTrueMat = = labelsPredMat) * 100;厘米。标题= sprintf (' 10倍交叉验证的混淆矩阵\nAverage Accuracy = %0.1f', valAccuracy);sortClasses(厘米、类别(emptyEmotions))厘米。ColumnSummary =“列规格化”；厘米。RowSummary =“row-normalized”；

函数[序列,sequencePerFile] = HelperFeatureVector2Sequence(特性、featureVectorsPerSequence featureVectorOverlap)MathWorks, Inc.版权所有如果featureVectorsPerSequence <= featureVectorOverlap error('重叠特征向量的数量必须小于每个序列的特征向量数量。')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);对于ii = 1:numel(features) sequencePerFile{ii} = floor(((size(features{ii}，2) - featureVectorsPerSequence)/hopLength) + 1;idx2 = 1;对于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]=HelperTrain和ValidateNet工作（varargn）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 (:) = [];%在每个折叠处打圈。trueLabelsCrossFold = {};predictedLabelsCrossFold = {};对于我= 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%以便时间沿着行兼容% sequenceInputLayer。tallTrain = cellfun (@ x (x) / max (abs (x)、[]“所有”)，塔勒列，“统一输出”、假);tallFeaturesTrain = cellfun (@ (x)提取(萃取器,x), tallTrain,“统一输出”、假);tallFeaturesTrain = cellfun (@ (x) x ', tallFeaturesTrain,“统一输出”、假);% #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“统一输出”、假);tallFeaturesValidation = cellfun (@ (x)提取(萃取器,x), tallValidation,“统一输出”、假);tallFeaturesValidation = cellfun (@ (x) x ', tallFeaturesValidation,“统一输出”、假);% #好< NASGU >[~, featuresValidation] = evalc (“收集(tallFeaturesValidation)”）;%使用evalc抑制命令行输出。% 3。使用训练集确定平均值和标准每个特征的偏差%。规范培训和验证%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”、假);对于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”、假);对于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。复制列车和验证集的标签，以便%它们与序列是一一对应的。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））DropOutputLayer（dropoutProb1）BilstLayer（numUnits，“OutputMode”,“最后一次”) dropoutLayer(dropoutProb2) fulllyconnectedlayer (numel(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”,．..“冗长”、假);% 8. 培训网络。网= trainNetwork (sequencesTrain、labelsTrain层,选择);% 9。评估网络。调用分类以得到预测的标签%为每个序列。获取每个预测标签的模式%序列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;对于ii=1:numel（trueLabels）predictedLabels（ii，：）=模式（predictedLabelsPerSequence（idx1:idx1+sequencePerFileValidation{ii}-1，：），1）；idx1=idx1+sequencePerFileValidation{ii}；结束trueLabelsCrossFold{i}=trueLabels；% #好< AGROW >predictedLabelsCrossFold{我}= predictedLabels;% #好< AGROW >结束结束