负载(“network_Audio_SequentialFeatureSelection.mat”，“bestNet”，“安全的”，“标准化者”）;

fs = afe.SampleRate;deviceReader = audioDeviceReader (“SampleRate”fs,“SamplesPerFrame”, 256);audioBuffer = dsp.AsyncBuffer (fs * 3);steBuffer = dsp.AsyncBuffer (1000);predictionBuffer = dsp.AsyncBuffer (5);

无花果=图;streamAxes =次要情节(1,1);streamPlot =情节(0 (fs, 1));ylabel (“振幅”)包含(“时间(s)”)标题(“音频流”) streamAxes。XTick = (0, fs);streamAxes。XTickLabel = [0, 1];streamAxes。YLim = [1];analyzedAxes =次要情节(3、1、2);analyzedPlot =情节(0 (fs / 2,1));标题(分析了部分的) ylabel (“振幅”)包含(“时间(s)”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”)包含(“类”）

steThreshold = 0.015;idxVec = 1: fs;抽搐而toc < 20%从你的设备中读取音频帧。audioIn = deviceReader ();将音频写入缓冲区。写(audioBuffer audioIn);%当200毫秒的数据未使用时，继续此循环。而audioBuffer。NumUnreadSamples > 0.2 * fs%从音频缓冲区读取1秒。这一秒的800毫秒%是重新读取旧数据，200ms是新数据。audioToAnalyze =阅读(audioBuffer, fs, 0.8 * fs);%更新图表以绘制当前的音频数据。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 > steThreshold%使用detectspeech函数来确定一个区域是否有语音%。idx = detectSpeech (audioToAnalyze, fs);%如果存在语音区域，执行以下操作。如果~ isempty (idx)除语音外，信号的所有部分都调零%区域，并修剪到0.5秒。audioToAnalyze = HelperTrimOrPad (audioToAnalyze (idx (1,1): idx(1、2),fs / 2);使音频正常化。audioToAnalyze = audioToAnalyze / max (abs (audioToAnalyze));%更新分析的分段图analyzedPlot。YData = audioToAnalyze;提取特征和转置他们，使时间%在列。特点=(提取(afe audioToAnalyze) ';使特性正常化。feature = (features - normalizer . mean) ./ normalizer . 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;%更新概率和获胜的情节%的标签。probabilityPlot。YData =聚合氯化铝;写(predictionBuffer,聚合氯化铝);如果predictionBuffer。NumUnreadSamples = = predictionBuffer。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，则假定无语音%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(“正在下载免费语音数字数据集……\n”解压缩(url, datasetFolder)结束

广告= audioDatastore (datasetFolder,“IncludeSubfolders”,真正的);[~, adsInfo] =阅读(广告);fs = adsInfo.SampleRate;

(~,文件名)= cellfun (@ (x) fileparts (x)的广告。文件,“UniformOutput”、假);ads.Labels =分类(string (cellfun (@ (x) x(1),文件名)));

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

赢得=汉明(圆(0.03 * fs),“周期”）;overlapLength =圆(0.02 * fs);afe = audioFeatureExtractor (．..“窗口”,赢了,．..“OverlapLength”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,．..“spectralKurtosis”,真的,．..“spectralRolloffPoint”,真的,．..“spectralSkewness”,真的,．..“spectralSlope”,真的,．..“spectralSpread”,真的,．..．..“节”假的,．..“harmonicRatio”、假);

numUnits =One hundred.；层= [．..sequenceInputLayer (1) bilstmLayer (numUnits“OutputMode”，“最后一次”) full connectedlayer (numel(categories(adsTrain.Labels))) softmaxLayer classificationLayer];选择= trainingOptions (“亚当”，．..“LearnRateSchedule”，“分段”，．..“洗牌”，“every-epoch”，．..“详细”假的,．..“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”)

函数[trueLabels predictedLabels,净,标准化者]= HelperTrainAndValidateNetwork (adsTrain、adsValidation afe,层,选项)%培训和验证网络。％%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网络的层数% options - trainingOptions对象％%输出:% true标签-验证集的真标签% predictedLabels -验证集的预测标签% net -训练的网络正规化因子-被测特性的正规化因子版权所有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沿着行与sequenceInputLayer兼容。fs = afe.SampleRate;tallTrain = cellfun (@ (x) HelperTrimOrPad (x, fs / 2) tallTrain,“UniformOutput”、假);tallTrain = cellfun (@ x (x) / max (abs (x)、[]“所有”)、tallTrain“UniformOutput”、假);tallFeaturesTrain = cellfun (@ (x)提取(afe x), tallTrain,“UniformOutput”、假);tallFeaturesTrain = cellfun (@ (x) x ', tallFeaturesTrain,“UniformOutput”、假);% #好< NASGU >[~, 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);% #好< NASGU >[~, featuresValidation] = evalc (“收集(tallFeaturesValidation)”）;%使用evalc抑制命令行输出。%使用训练集确定每个的平均值和标准偏差%的特性。将训练和验证集规范化。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);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”、假);为ii = 1:numel(featuresValidation) idx = find(isnan(featuresValidation{ii}));如果~isempty(idx) featuresValidation{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;%更新输入层的测试特性的数量。层(1)= sequenceInputLayer(大小(featuresTrain {1}, 1));%培训网络。网= trainNetwork (featuresTrain、labelsTrain层,选择);%评估网络。调用classify以获得每个的预测标签%的序列。predictedLabels =分类(净,featuresValidation);trueLabels = adsValidation.Labels;%将标准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

函数(日志,bestFeatures bestNet bestNormalizers] = HelperSFS (adsTrain、adsValidate afeThis,层,选择,方向)％%的输入:% adsTrain -指向训练集的audioDatastore对象% adsValidate -指向验证集的audioDatastore对象% afe - audioFeatureExtractor对象。设置所有功能测试为真% layers - LSTM或BiLSTM网络的层数% options - trainingOptions对象% direction - SFS方向，指定为“forward”或“backward”％%输出:%日志-包含测试特性配置和相应验证精度的表% bestFeatures -包含最佳特性配置的结构体% bestNet -训练有素的网络，具有最高的验证精度% 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:开启向后选择%或全部关闭向前选择。featureConfig = info (afe);为i = 1: N如果strcmpi(方向,“落后”) featureConfig.(featuresToTest{i}) = true;其他的featureConfig。(featuresToTest{我})= false;结束结束%初始化日志，以跟踪特征配置和准确性。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;而wrapperIdx < = N创建包含要测试的所有功能配置的单元格数组在当前循环中。featureConfigsToTest =细胞(元素个数(featuresToTest), 1);为2 = 1:元素个数(featuresToTest)如果strcmpi(方向,“落后”) featureConfig.(featuresToTest{ii}) = false;其他的featureConfig。(featuresToTest {2}) = true;结束featureConfigsToTest {2} = featureConfig;如果strcmpi(方向,“落后”) featureConfig.(featuresToTest{ii}) = true;其他的featureConfig。(featuresToTest {2}) = false;结束结束%循环每个特性集。为2 = 1:元素个数(featureConfigsToTest)确定当前要测试的特性配置。更新%该功能安全。currentConfig = featureConfigsToTest {2};currentConfig集(afe)%培训并获得当前k倍交叉验证精度%的功能配置。[trueLabels predictedLabels,净,标准化者]= HelperTrainAndValidateNetwork (adsTrain、adsValidate afe,层,选项);valAccuracy =意味着(trueLabels = = predictedLabels) * 100;如果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%以最佳精度确定和打印设置。如果精度%没有提高，结束跑。[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;%根据最近的胜出者更新待测试的功能。赢家=日志{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%对日志进行排序，使其更具可读性。:日志(1)= [];%删除第一行占位符。日志= sortrows(日志,{“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:num (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

函数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，前后修剪均匀。如果x被填充，它就是%在前面和后面用0填充。对于奇数长度的修剪或%填充，额外的样品从后面修剪或填充。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 frontTrim = floor((a-n)/2)+1;backTrim = a - n - frontTrim;y = x (frontTrim: end-backTrim);其他的y = x;结束结束