加载（“network_Audio_SequentialFeatureSelection.mat”，“bestNet”，'afe'，“标准化者”）;

FS = AFE.SAMPLEDE;devicereader = audiodevicereader（“SampleRate”fs,“SamplesPerFrame”，256）;audiobuffer = dsp.asyncuffer（FS * 3）;stebuffer = dsp.asyncubuffer（1000）;predicationbuffer = dsp.asyncuffer（5）;

无花果=图;streamAxes =次要情节(1,1);streamPlot =情节(0 (fs, 1));ylabel (“振幅”）Xlabel（'时间'')标题('音频流'）Streamaxes.xtick = [0，FS];Streamaxes.xticklabel = [0,1];stribexes.ylim = [-1,1];分析=子图（3,1,2）;分析ZEDPLOT = PLOT（零（FS / 2,1））;标题（分析了部分的）ylabel（“振幅”）Xlabel（'时间''甘氨胆酸)组(,'xticklabel'[]) analyzedAxes。XTick = [0, fs / 2];analyzedAxes。XTickLabel = [0, 0.5];analyzedAxes。YLim = [1];probabilityAxes =情节(3、1,3);probabilityPlot =酒吧(0:9,0.1 * (10));轴([1,10,0,1])ylabel (“概率”）Xlabel（“类”）

steThreshold = 0.015;idxVec = 1: fs;抽搐尽管toc < 20%从设备中读取音频帧。audioIn = deviceReader ();将音频写入缓冲区。写(audioBuffer audioIn);%当有200毫秒的数据未使用时，继续此循环。尽管audioBuffer。NumUnreadSamples > 0.2 * fs%从音频缓冲区读取1秒。1秒，800毫秒%表示重新读取旧数据，200ms表示新数据。audioToAnalyze =阅读(audioBuffer, fs, 0.8 * fs);%更新图形以绘制当前音频数据。streamPlot。YData = audioToAnalyze;ste =意味着(abs (audioToAnalyze));写(steBuffer, ste);如果steBuffer。NumUnreadSamples > 5 abc = sort(peek(steBuffer));steThreshold = abc(圆(0.4 *元素个数(abc)));结束如果ste> tethreshold.%使用detectspeech功能来确定一个语音区域%。Idx =检测echech（录音萼，FS）;%如果存在语音区域，执行以下操作。如果~ isempty (idx)％零解出信号的所有部分除了演讲％区域，并修剪0.5秒。audioToAnalyze = HelperTrimOrPad (audioToAnalyze (idx (1,1): idx(1、2),fs / 2);%将音频正常化。audioToAnalyze = audioToAnalyze / max (abs (audioToAnalyze));％更新分析的段绘图analyzedPlot。YData = audioToAnalyze;%提取特征并转置，使时间为%在列。特点=(提取(afe audioToAnalyze) ';%将特征规格化。features = (features - normalize . mean) ./ normalize . standarddeviation;%调用分类，以确定概率和%获得标签。特性(isnan(特性))= 0;(标签,聚合氯化铝)= (bestNet、特性)进行分类;%用概率和获胜值更新图％ 标签。probaultsplot.ydata = probs;写（预测缓冲，probs）;如果predictionBuffer。NumUnreadSamples = = predictionBuffer。容量lastTen = peek(predictionBuffer);[~,决定]= max(平均(lastTen。*损害(大小(lastTen, 1)), 1));probabilityAxes.Title.String = num2str (decision-1);结束结束其他的%如果信号能量低于阈值，假设没有语音%检测。probabilityAxes.Title.String =＇＇;probabilityPlot。YData = 0.1 * 1 (1);analyzedPlot。YData = 0 (f / 2、1);重置(predictionBuffer)结束drawnowlimitrate结束结束

url =“https://zenodo.org/record/1342401/files/Jakobovski/free-spoken-digit-dataset-v1.0.8.zip”;downloadFolder = tempdir;datasetFolder = fullfile (downloadFolder,'fsdd'）;如果〜存在（DataSetFolder，“dir”)流(“下载免费语音数字数据集…\n”解压缩(url, datasetFolder)结束

广告= audioDatastore (datasetFolder,'insertumbfolders',真正的);[~, adsInfo] =阅读(广告);fs = adsInfo.SampleRate;

(~,文件名)= cellfun (@ (x) fileparts (x)的广告。文件,'统一输出'、假);ads.Labels =分类(string (cellfun (@ (x) x(1),文件名)));

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

Win =汉明（圆形（0.03 * fs），“周期”）;overlapLength =圆(0.02 * fs);afe = audioFeatureExtractor (......“窗口”， 赢，......'overlaplencth'overlapLength,......“SampleRate”，fs，............“linearSpectrum”假的,......“melSpectrum”假的,......“barkSpectrum”假的,......'erbspectrum'假的,............“mfcc”,真的,......“mfccDelta”,真的,......“mfccDeltaDelta”,真的,......“gtcc”,真的,......“gtccDelta”,真的,......“gtccDeltaDelta”,真的,............'spectralcentroid',真的,......“spectralCrest”,真的,......“spectralDecrease”,真的,......“spectralEntropy”,真的,......“spectralFlatness”,真的,......“spectralFlux”,真的,......'spectarkurtosis',真的,......'spectralrolloffpoint'，真的，......'spectrancewswness',真的,......“spectralSlope”,真的,......“spectralSpread”,真的,............“节”假的,......'halmonicratio'、假);

numUnits =100.;层= [......sequenceInputLayer (1) bilstmLayer (numUnits“outputmode”，“最后一次”) full connectedlayer (numel(categories(adsTrain.Labels))) softmaxLayer classiationlayer];选择= trainingOptions (“亚当”，......“Learnrateschedule”，“分段”，......“洗牌”，“every-epoch”，......“详细”，错误的，......“MaxEpochs”, 20);

方向=“前进”;(日志,bestFeatures bestNet标准化者]= HelperSFS (adsTrain、adsValidation afe,层,选择,方向);

日志

bestFeatures

安全设置(afe bestFeatures)

%保存(“network_Audio_SequentialFeatureSelection.mat”、“bestNet”,“安全的”,“标准化者”)

功能[Truelabels，PredightLabels，Net，Indantmizers] = HelpertrainAndValidateNetwork（adstrain，Adsvalidation，AFE，图层，选项）培训并验证一个网络。％%的输入:% adsTrain -指向训练集的audioDatastore对象% adsValidation - audioDatastore对象，指向验证集％afe  -  audiofeatureextractor对象。% layers - LSTM或BiLSTM网络的层％选项 -  TrainingOptions对象％%输出:% trueLabels -验证集的真标签% predictedLabels -验证集的预测标签％网络培训网络%归一化器-被测特性的归一化因子版权所有2019 The MathWorks, Inc.。%将数据转换为高数组。tallTrain =高(adsTrain);tallValidation =高(adsValidation);%从训练集中提取特征。重新定位功能，以便% time是沿着行，以兼容sequenceInputLayer。FS = AFE.SAMPLEDE;Talltrain = Cellfun（@（x）Helpertrimorpad（x，fs / 2），塔特拉特，“UniformOutput”、假);tallTrain = cellfun (@ x (x) / max (abs (x)、[]“所有”)、tallTrain“UniformOutput”、假);tallFeaturesTrain = cellfun (@ (x)提取(afe x), tallTrain,“UniformOutput”、假);TallFeaturestain = Cellfun（@（x）x'，tallfeaturestain，“UniformOutput”、假);％＃OK [〜，featureStrain] = evalc（“收集(tallFeaturesTrain)”）;％使用evalc抑制命令行输出。tallvalidation = Cellfun（@（x）Helpertrimorpad（x，fs / 2），tallvalidation，“UniformOutput”、假);tallValidation = cellfun (@ x (x) / max (abs (x)、[]“所有”），TallValidation，“UniformOutput”、假);tallFeaturesValidation = cellfun (@ (x)提取(afe x), tallValidation,“UniformOutput”、假);tallFeaturesValidation = cellfun (@ (x) x ', tallFeaturesValidation,“UniformOutput”、假);％＃OK [〜，特点过验证] = evalc（“收集(tallFeaturesValidation)”）;％使用evalc抑制命令行输出。%使用训练集来确定每一个的平均值和标准差%的特性。规范化培训和验证集。AllFeatures = Cat（2，FeatureStrain {：}）;m =均值（全部，2，'omitnan'）;S =性病(allFeatures 0 2,'omitnan'）;featuresTrain = cellfun (@ (x)(即x m)。/ S, featuresTrain,'统一输出'、假);为i = 1:numel(featuresTrain) idx = find(isnan(featuresTrain{ii}));如果~isempty(idx) featuresTrain{ii}(idx) = 0;结束结束特征idation = Cellfun（@（x）（x-m）./ s，companyvalidation，'统一输出'、假);为idx = find(isnan(featuresValidation{ii}));如果~isempty(idx) featresvalidation {ii}(idx) = 0;结束结束%复制列车和验证集的标签，以便它们在%与序列一一对应。labelsTrain = adsTrain.Labels;%更新输入层，显示被测试特性的数量。层(1)= sequenceInputLayer(大小(featuresTrain {1}, 1));%训练网络。网= trainNetwork (featuresTrain、labelsTrain层,选择);%评估网络。调用classification来获得每个标签的预测值％ 顺序。predictedLabels =分类(净,featuresValidation);trueLabels = adsValidation.Labels;%将规范化因子保存为结构体。标准化者。意味着= M;标准化者。StandardDeviation = S;结束

功能[Logbook，Bestfeatures，BestNet，Bestnormalizers] = allowersfs（adstrain，Adsvalidate，Afethis，图层，选项，方向）％%的输入:% adsTrain -指向训练集的audioDatastore对象％AdsValidate  -  AudioDataStore对象指向验证集％afe  -  audiofeatureextractor对象。将所有功能设置为true% layers - LSTM或BiLSTM网络的层％选项 -  TrainingOptions对象％方向 -  SFS方向，指定为“转发”或“向后”％%输出:%日志-包含功能配置测试和相应的验证准确性的表％bestfeatures  - 结构可包含最佳功能配置% bestNet -具有最高验证精度的训练网络% bestnormalizer -最佳特征的特征归一化因子版权所有2019 The MathWorks, Inc.。afe = (afeThis)复印件;featuresToTest =字段名(信息(afe));N =元素个数(featuresToTest);bestValidationAccuracy = 0;%设置初始特性配置:全部打开，向后选择%或全部关闭，用于向前选择。featureConfig = info (afe);为我= 1：n如果strcmpi(方向,“落后”) featureConfig.(featuresToTest{i}) = true;其他的featureconfig。（featureestotest {i}）= false;结束结束%初始化日志以跟踪特性配置和准确性。日志=表(featureConfig 0'variablenames'，[“功能配置”，“准确性”]);%执行顺序特征评估。Wrapperidx = 1;bestaccuracy = 0;尽管wrapperIdx < = N%创建一个包含所有要测试的特性配置的单元格数组%在当前循环中。featureconfigstotest = cell（numel（featureasttist），1）;为2 = 1:元素个数(featuresToTest)如果strcmpi(方向,“落后”) featureConfig.(featuresToTest{ii}) = false;其他的featureconfig。（featureestotest {ii}）= true;结束featureconfigstotest {ii} = featureconfig;如果strcmpi(方向,“落后”) featureConfig.(featuresToTest{ii}) = true;其他的featureConfig。(featuresToTest {2}) = false;结束结束%循环遍历每个特性集。为2 = 1:元素个数(featureConfigsToTest)%确定要测试的当前特性配置。更新特性安全。currentConfig = featureConfigsToTest {2};currentConfig集(afe)％列车并获得当前的k折叠交叉验证精度%的功能配置。[Truelabels，PredightLabels，Net，Normanizers] = HelpertrainAndValidateNetwork（adstrain，Adsvalidate，AFE，图层，选项）;valaccuracy =平均值（truelabels == predightlabels）* 100;如果valAccuracy > bestValidationAccuracy bestValidationAccuracy = valAccuracy;bestNet =净;bestNormalizers =标准化者;结束%更新日志结果=表(currentConfig valAccuracy,'variablenames'，[“功能配置”，“准确性”]);日志=(日志;结果);% #好< AGROW >结束以最佳的精度确定和打印设置。如果精度％没有改善，结束运行。[a，b] = max（日志{:,'准确性'});如果a <= bestAccuracy wrapperIdx = inf;其他的wrapperIdx = wrapperIdx + 1;结束bestAccuracy =一个;%根据最近的获胜者更新特性到测试。获胜者= logbook {b，'功能配置'};fn =字段名(冠军);tf = structfun (@ (x) (x)赢家);如果strcmpi(方向,“落后”) featuresToRemove = fn(~tf);其他的featuresToRemove = fn (tf);结束为ii = 1:numel(featuresToRemove) loc = strcmp(featuresToTest,featuresToRemove{ii});featuresToTest (loc) = [];如果strcmpi(方向,“落后”) featureConfig.(featuresToRemove{ii}) = false;其他的featureConfig。(featuresToRemove {2}) = true;结束结束结束%对日志簿进行排序，使其更具可读性。:日志(1)= [];％删除占位符第一行。logbook = sortrows（logbook，{'准确性'}，{“下”});bestFeatures =日志{1,'功能配置'};m =日志{:,'功能配置'};fn =字段名(m);myString =字符串(元素个数(m), 1);为wrapperidx = 1：numel（m）tf = structfun（@（x）（x），logbook {wrapperidx，'功能配置'});myString (wrapperIdx) = strjoin (fn (tf),“，”）;结束日志=表(myString日志{:,'准确性'}，'variablenames'，[“特征”，“准确性”]);结束

功能y = helpertrimorpad（x，n）％y = Helpertrimorpad（x，n）修剪或填充输入x到n个样本。如果x是％修剪，它在正面和背面平均修剪。如果x被填充，则是%在前面和后面相等地用零填充。奇数长度的修剪或％填充，从后面修剪或填充额外的样品。版权所有2019 The MathWorks, Inc.。一个=大小(x, 1);如果a < n frontPad = floor((n-a)/2);backPad = n - a - frontPad;(y = 0 (frontPad 1); x; 0(挤压垫,1)];elseifa> n正面=地板（（a-n）/ 2）+1;回溯= a  -  n  -  fronttrim;y = x（Fronttrim：end-backtrim）;其他的y = x;结束结束