downloadFolder = matlab.internal.examples.downloadSupportFile (“音频”,“L3DAS21_ov1.zip”);dataFolder = tempdir;unzip(downloadFolder,dataFolder) dataset = fullfile(dataFolder，“L3DAS21_ov1”);

speedupExample =假;

adsTrain = audioDatastore (fullfile(数据集,“训练”,“数据”));adsTrainA =子集(adsTrain cellfun (@ (c) endsWith (c,“A.wav”), adsTrain.Files));adsTrainB =子集(adsTrain cellfun (@ (c) endsWith (c,“B.wav”), adsTrain.Files));adsValidation = audioDatastore (fullfile(数据集,“确认”,“数据”));adsValidationA =子集(adsValidation cellfun (@ (c) endsWith (c,“A.wav”), adsValidation.Files));adsValidationB =子集(adsValidation cellfun (@ (c) endsWith (c,“B.wav”), adsValidation.Files));

如果speedupExample adsTrainA =子集(adsTrainA,1:2);adsTrainB =子集(adsTrainB 1:2);结束

云母=预览(adsTrainA);micB =预览(adsTrainB);tiledlayout (4,2 TileSpacing =“紧”) nexttile plot(micA(:，1)) title(“麦克风”) ylabel (“W”) nexttile plot(micB(:，1)) title(B“麦克风”nexttile plot(micA(:，2)) ylabel(“X”) nexttile plot(micB(:，2)) nexttile plot(micA(:，3)) ylabel(“Y”) nexttile plot(micB(:，3)) nexttile plot(micB(:，4)) ylabel(“Z”) nexttile情节(micB (:,)

麦克风=1;频道=1;时间=10;fs = 32 e3;%数据的已知采样率。s =(云母、micB);Data = s(1:round(duration*fs)，channel +(麦克风-1)*4);声音(数据、fs)

键盘= [“Chink_and_clink”,“Computer_keyboard”,“Cupboard_open_or_close”,“Drawer_open_or_close”,…“Female_speech_and_woman_speaking”,“Finger_snapping”,“Keys_jangling”,“敲门”,“笑”,…“Male_speech_and_man_speaking”,“打印机”,“剪刀”,“电话”,“写”];valueSet ={1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12日,13日,14日};参数个数。SoundClasses = containers.Map(键盘、valueSet);

(文件夹,fn) = fileparts (adsTrainA.Files);定位路径= fullfile (strrep(文件夹,filesep +“数据”filesep +“标签”),“label_”+ strrep (fn,“_A”,”“) +. csv”);ttdsTrain = tabularTextDatastore(定位路径);labelTable =预览(ttdsTrain)

params.Targets.TotalDuration = 60;params.Targets.NumFrames = 600;

SEDTargets = extractSEDTargets (labelTable params);[numframes, numclasses] =大小(SEDTargets {1})

dsTTrain =变换(ttdsTrain @ (x) extractSEDTargets (x, params));sedTTrain = readall (dsTTrain);(文件夹,fn) = fileparts (adsValidationA.Files);定位路径= fullfile (strrep(文件夹,filesep +“数据”filesep +“标签”),“label_”+ strrep (fn,“_A”,”“) +. csv”);ttdsValidation = tabularTextDatastore(定位路径);dsTValidation =变换(ttdsValidation @ (x) extractSEDTargets (x, params));sedTValidation = readall (dsTValidation);

params.DOA.ScaleFactor = 2;DOATargets = extractDOATargets (labelTable params);[numframes, numaxis] =大小(DOATargets {1})

dsTTrain =变换(ttdsTrain @ (x) extractDOATargets (x, params));doaTTrain = readall (dsTTrain);(文件夹,fn) = fileparts (adsValidationA.Files);定位路径= fullfile (strrep(文件夹,filesep +“数据”filesep +“标签”),“label_”+ strrep (fn,“_A”,”“) +. csv”);ttdsValidation = tabularTextDatastore(定位路径);dsTValidation =变换(ttdsValidation @ (x) extractDOATargets (x, params));doaTValidation = readall (dsTValidation);

params.SED.SampleRate = 32 e3;params.SED.HopLength = 400;params.SED.Window =汉明(512“周期”);

stftFeats = extractSTFT({云母,micB}, params);[numfeaturesSED, numframesSED numchannelsSED] =大小(stftFeats)

频道=7;figure imagesc(stftfeat (:，:，channel)) colorbar xlabel(“帧”) ylabel (“频率(本)”)设置(gca YDir =“正常”)

= ~ isempty(版本(pFlag“平行”) & & ~ speedupExample;trainDS =结合(adsTrainA adsTrainB);trainDS_T =变换(trainDS @ (x) {extractSTFT (x, params)}, IncludeInfo = false);XTrain = readall (pFlag trainDS_T UseParallel =);valDS =结合(adsValidationA adsValidationB);valDS_T =变换(valDS @ (x) {extractSTFT (x, params)}, IncludeInfo = false);XValidation = readall (pFlag valDS_T UseParallel =);

trainSedDS =结合(arrayDatastore (XTrain OutputType =“相同”), arrayDatastore (sedTTrain OutputType =“相同”));valSedDS =结合(arrayDatastore (XValidation OutputType =“相同”), arrayDatastore (sedTValidation OutputType =“相同”));

trainOptionsSED =结构(…MaxEpochs = 300,…MiniBatchSize = 4,…InitialLearnRate e-5 = 1,…GradientDecayFactor = 0.01,…SquaredGradientDecayFactor = 0.0,…ValidationPatience = 25,…LearnRateDropPeriod = 100,…LearnRateDropFactor = 1);如果speedupExample trainOptionsSED。MaxEpochs = 1;结束

trainSEDmbq = minibatchqueue (trainSedDS,…MiniBatchSize = trainOptionsSED。MiniBatchSize,…OutputAsDlarray = [1],…MiniBatchFormat = [“SSCB”,“TCB”),…OutputEnvironment = [“汽车”,“汽车”]);validationSEDmbq = minibatchqueue (valSedDS,…MiniBatchSize = trainOptionsSED。MiniBatchSize,…OutputAsDlarray = [1],…MiniBatchFormat = [“SSCB”,“TCB”),…OutputEnvironment = [“汽车”,“汽车”]);

seldnetCNNLayers = [imageInputLayer([numfeatuered, numframered,numchannelsSED]，Normalization= .“没有”、名称=“输入”) convolution2dLayer([3, 3], 64填充=“相同”、名称=“conv1”) batchNormalizationLayer (Name =“batchnorm1”) reluLayer (Name =“relu1”) maxPooling2dLayer([8 2],跨步=[8 2],填充=“相同”、名称=“maxpool1”) convolution2dLayer([3, 3], 128填充=“相同”、名称=“conv2”) batchNormalizationLayer (Name =“batchnorm2”) reluLayer (Name =“relu2”) maxPooling2dLayer([8 2],跨步=[8 2],填充=“相同”、名称=“maxpool2”) convolution2dLayer([3, 3], 256填充=“相同”、名称=“conv3”) batchNormalizationLayer (Name =“batchnorm3”) reluLayer (Name =“relu3”) maxPooling2dLayer(2, 2,步=(2,2)填充=“相同”、名称=“maxpool3”) convolution2dLayer([3, 3], 512填充=“相同”、名称=“conv4”) batchNormalizationLayer (Name =“batchnorm4”) reluLayer (Name =“relu4”) maxPooling2dLayer([1],跨步=[1],填充=“相同”、名称=“maxpool4”) reshapeLayer (“重塑”));netCNN = dlnetwork (layerGraph (seldnetCNNLayers));

seldnetGRULayers = [sequenceInputLayer(1024,Name= .“sequenceInputLayer”) bigruLayer(1024、256、Name =“gru1”) bigruLayer(512256年,Name =“gru2”) bigruLayer(512256年,Name =“gru3”) fullyConnectedLayer(1024年,Name =“fc1”) reluLayer (Name =“relu1”) fullyConnectedLayer(1024年,Name =“取得”) reluLayer (Name =“relu2”) fullyConnectedLayer(1024年,Name =“一个fc3”文件) reluLayer (Name =“relu3”) fullyConnectedLayer (params.SoundClasses.Count Name =“fc4”) sigmoidLayer (Name =“输出”));netRNN = dlnetwork (layerGraph (seldnetGRULayers));

sedModel。CNN = netCNN;sedModel。RNN = netRNN;

迭代= 0;averageGrad = [];averageSqGrad = [];时代= 0;bestLoss =正;badEpochs = 0;learnRate = trainOptionsSED.InitialLearnRate;

页= progressPlotterSELD ();

rng (0)而时代< trainOptionsSED。MaxEpochs && badEpochs < trainOptionsSED. MaxEpochs && badEpochs < trainOptionsSED.ValidationPatience epoch = epoch + 1;% Shuffle迷你批处理队列。洗牌(trainSEDmbq)而hasdata (trainSEDmbq)%更新迭代计数器。迭代=迭代+ 1;%读取小批数据。[X, T] =下一个(trainSEDmbq);%评估模型的梯度和损失使用dlfeval和modelLoss函数。(损失、毕业生、国家)= dlfeval (@modelLoss sedModel X T);损失=损失/大小(T, 2);%更新状态。sedModel.CNN.State = state.CNN;sedModel.RNN.State = state.RNN;%使用Adam优化器更新网络参数。[sedModel, averageGrad averageSqGrad] = adamupdate (averageGrad sedModel,研究生,…迭代,averageSqGrad learnRate、trainOptionsSED.GradientDecayFactor trainOptionsSED.SquaredGradientDecayFactor);更新训练进度图。。updateTrainingProgress (pp、时代=时代,LearnRate = LearnRate迭代=迭代,损失=损失);结束%在每个纪元之后执行验证。损失= predictBatch (sedModel validationSEDmbq);%用验证结果更新训练进度图。updateValidation (pp、损失=损失、迭代=迭代)%如果验证损失改善，创建检查点。如果验证%的损失没有改善，增加了糟糕时期的数量。如果loss < bestLoss bestLoss = loss;badEpochs = 0;文件名=“SED-BestModel”;保存(文件名,“sedModel”);其他的badEpochs = badEpochs + 1;结束%更新学习率如果rem(epoch,trainOptionsSED.LearnRateDropPeriod)==0 learnRate = learnRate*trainOptionsSED.LearnRateDropFactor;结束结束

params.DOA.SampleRate = 32 e3;params.DOA.Window =损害(1024);params.DOA.NumBands = 96;params.DOA.HopLength = 400;

gccPhatFeats = extractGCCPHAT({云母,micB}, params);[numfeaturesDOA, timestepsDOA numchannelsDOA] =大小(gccPhatFeats)

channelpair =1;figure imagesc(gccphatfeat (:，:，channelpair)) colorbar xlabel(“帧”) ylabel (“乐队”)设置(gca YDir =“正常”)

= ~ isempty(版本(pFlag“平行”) & & ~ speedupExample;trainDS =结合(adsTrainA adsTrainB);trainDS_T =变换(trainDS @ (x) {extractGCCPHAT (x, params)}, IncludeInfo = false);XTrain = readall (pFlag trainDS_T UseParallel =);

valDS =结合(adsValidationA adsValidationB);valDS_T =变换(valDS @ (x) {extractGCCPHAT (x, params)}, IncludeInfo = false);XValidation = readall (pFlag valDS_T UseParallel =);

trainDOA =结合(arrayDatastore (XTrain OutputType =“相同”), arrayDatastore (doaTTrain OutputType =“相同”));validationDOA =结合(arrayDatastore (XValidation OutputType =“相同”), arrayDatastore (doaTValidation OutputType =“相同”));

trainOptionsDOA = trainOptionsSED;

trainDOAmbq = minibatchqueue (trainDOA,…MiniBatchSize = trainOptionsDOA。MiniBatchSize,…OutputAsDlarray = [1],…MiniBatchFormat = [“SSCB”,“TCB”),…OutputEnvironment = [“汽车”,“汽车”]);validationDOAmbq = minibatchqueue (validationDOA,…MiniBatchSize = trainOptionsDOA。MiniBatchSize,…OutputAsDlarray = [1],…MiniBatchFormat = [“SSCB”,“TCB”),…OutputEnvironment = [“汽车”,“汽车”]);

seldnetCNNLayers (1) = imageInputLayer ([numfeaturesDOA、timestepsDOA numchannelsDOA],标准化=“没有”、名称=“输入”);seldnetCNNLayers (5) = maxPooling2dLayer(3 2,步=(3 2),填充=“相同”、名称=“maxpool1”);netCNN = dlnetwork (layerGraph (seldnetCNNLayers));seldnetGRULayers (11) = fullyConnectedLayer (Name =“fc4”);seldnetGRULayers (12) = tanhLayer (Name =“输出”);netRNN = dlnetwork (layerGraph (seldnetGRULayers));

doaModel。CNN = netCNN;doaModel。RNN = netRNN;

迭代= 0;averageGrad = [];averageSqGrad = [];时代= 0;bestLoss =正;badEpochs = 0;learnRate = trainOptionsDOA.InitialLearnRate;

页= progressPlotterSELD ();

rng (0)而时代< trainOptionsDOA。MaxEpochs && badEpochs < trainOptionsDOA。ValidationPatience epoch = epoch + 1;% Shuffle迷你批处理队列。洗牌(trainDOAmbq)而hasdata (trainDOAmbq)%更新迭代计数器。迭代=迭代+ 1;%读取小批数据。[X, T] =下一个(trainDOAmbq);%评估模型的梯度和损失使用dlfeval和modelLoss函数。(损失、毕业生、国家)= dlfeval (@modelLoss doaModel X T);损失=损失/大小(T, 2);%更新状态。doaModel.CNN.State = state.CNN;doModel.RNN.State = state.RNN;%使用Adam优化器更新网络参数。[doaModel, averageGrad averageSqGrad] = adamupdate (averageGrad doaModel,研究生,…迭代,averageSqGrad learnRate、trainOptionsDOA.GradientDecayFactor trainOptionsDOA.SquaredGradientDecayFactor);%更新训练进度图updateTrainingProgress (pp、时代=时代,LearnRate = LearnRate迭代=迭代,损失=损失);结束%在每个epoch之后执行验证损失= predictBatch (doaModel validationDOAmbq);%用验证结果更新训练进度图。updateValidation (pp、损失=损失、迭代=迭代)%如果验证损失改善，创建检查点。如果验证%的损失没有改善，增加了糟糕时期的数量。如果loss < bestLoss bestLoss = loss;badEpochs = 0;文件名=“DOA-BestModel”;保存(文件名,“doaModel”);其他的badEpochs = badEpochs + 1;结束%更新学习率如果rem(epoch,trainOptionsDOA.LearnRateDropPeriod)==0 learnRate = learnRate* trainoptionsdoa . learnratdropfactor;结束结束

sedModel = importdata (“SED-BestModel.mat”);doaModel = importdata (“DOA-BestModel.mat”);

参数个数。SpatialThreshold = 0.2;参数个数。SilenceThreshold = 0.1;

结果= computeMetrics (sedModel doaModel、validationSEDmbq validationDOAmbq, params);结果

(结果,cm) = computeMetrics (sedModel、doaModel validationSEDmbq, validationDOAmbq, params, ApplySmoothing = true);结果

figure(Position=[100 100 800 800]);confusionchart(厘米,密匙(params.SoundClasses))

函数T = extractDOATargets (csvFile params)%EXTRACTDOATARGETS提取到达方向(DOA)目标% T = extractDOATargets(fileName,params)解析CSV文件% fileName，并返回一个矩阵t。目标矩阵是一个n × 3矩阵%矩阵，其中N对应帧数，3对应3个轴描述三维空间中的位置。。%预分配目标矩阵。全0的帧对应无声音%的来源。T = 0 (params.Targets.NumFrames 3);%将声源的时间戳量化为帧。startendTime = [csvFile.Start, csvFile.End];startendFrame = time2frame (startendTime params.Targets.TotalDuration params.Targets.NumFrames);%对于每个声源，填充目标矩阵声源位置为%适当的帧数。为ii = 1:size(startendFrame,1) idx = startendFrame(ii,1):startendFrame(ii,2)-1;T (idx:) = repmat ([csvFile.X (ii), csvFile.Y (ii), csvFile.Z (ii)],元素个数(idx), 1);结束%缩放目标，使其在-1和1之间(tanh的边界%活化层)。将目标包裹在cell数组中，方便批量%处理。T = {T / params.DOA.ScaleFactor};结束

函数T = extractSEDTargets (csvFile params)%EXTRACTSEDTARGETS提取声音事件检测(SED)目标% T = extractSEDTargets(fileName,params)解析CSV文件% fileName，并返回SED目标t的矩阵。目标矩阵是一个n × k矩阵%矩阵，其中N对应帧数，K对应%声音类的数量。%预分配目标矩阵。全0的帧对应无声音%的来源。T = 0 (params.Targets.NumFrames params.SoundClasses.Count);%将声源的时间戳量化为帧。startendTime = [csvFile.Start, csvFile.End];startendFrame = time2frame (startendTime params.Targets.TotalDuration params.Targets.NumFrames);%对于每个声源，填充目标矩阵的相应列%，带1，表示该帧中存在声音类。为ii = 1:size(startendFrame,1) classID = params.SoundClasses(csvFile.Class{ii});T (startendFrame (2, 1): startendFrame (2, 2) 1, classID) = 1;结束%将目标包装在单元格数组中，以方便批处理。T = {T};结束

函数X = extractSTFT(年代,params)%EXTRACTSTFT提取log-量级的居中STFT% X = extractSTFT({s1,s2}，params)连接s1和s2，然后%提取单边对数量级STFT。之前对信号进行填充%的STFT，使第一个窗口以第一个样本为中心。的% output被修剪，以去除第一个(DC)系数和最后一个系数%频谱。输入参数定义了STFT。沿着第二个(通道)维度连接信号。。音频=猫(2 s {:});%提取居中的STFT。N =元素个数(params.SED.Window);overlapLength = N - params.SED.HopLength;S = centeredSTFT(音频、params.SED.Window overlapLength N);%修剪第一个系数从所有频谱和修剪最后一个频谱。S = S(2:结束,1:end-1,:);%转换为对数量级。使用偏移量来防止log为零。mag = log(abs(S) + eps);%投射输出到单精度。X =单(mag);结束

函数X = extractGCCPHAT(年代,params)%EXTRACTGCCPHAT提取广义互相关相位变换(GCC-PHAT)特征% X = extractGCCPHAT({s1,s2}，params)连接s1和s2，然后%提取所有通道对的GCC-PHAT。%连接两个麦克风对应的信号。音频=猫(2 s {:});%统计输入通道的总数。nChan =大小(音频,2);%计算输出通道总数。numOutputChannels = nchoosek (nChan, 2);%预分配一个NumFeatures-by-NumFrames-by-NumChannels特性(预测器)%的矩阵。numFrames =大小(音频,1)/ params.DOA.HopLength;X = 0 (params.DOA.NumBands numFrames numOutputChannels);% -----------------------------------%计算每一对渠道的GCC-PHAT。%预计算每个通道的STFT。N =元素个数(params.DOA.Window);overlapLength = N - params.DOA.HopLength;micAB_stft = centeredSTFT(音频、params.DOA.Window overlapLength N);conjmicAB_stft =连词(micAB_stft(:,:, 2:结束));idx = 1;为2 = 1: nChan - 1 R = micAB_stft (:,:, ii)。* conjmicAB_stft (:,:, ii:结束);R = exp(1i .* angle(R));R = padarray(R, N/2 - 1，“职位”);gcc = fftshift(传输线(R, [], 1,“对称”), 1);X (:,:, idx: idx +大小(R, 3) 1) = gcc(地板(N / 2 + 1 - (params.DOA.NumBands-1) / 2):地板(N / 2 + 1 + (params.DOA.NumBands-1) / 2), 1: end-1,:);idx = idx + size(R,3);结束% -----------------------------------%投射输出到单精度。X =单(X);结束

函数s = centeredSTFT(音频、赢得overlapLength fftLength)% CENTEREDSTFT集中STFT% s = centeredSTFT(audioIn,win,overlapLength,fftLength)计算一个STFT%，第一个窗口以第一个样本为中心。两端分别为%填充的反射音频信号。% Pad前面和后面的输入信号。firstR =翻转(音频(1:fftLength / 2,:), 1);lastR = flip(audio(end - fftLength/2 + 1:end，:)，1);sig =猫(1 firstR音频,昨天深);% STFT执行。s = stft(团体、窗口=赢,OverlapLength = OverlapLength FFTLength = FFTLength FrequencyRange =“单向的”);结束

函数fnum = time2frame (t,杜尔numFrames)%TIME2FRAME将时间戳转换为帧号% fnum = time2frame(t,dur,numFrames)映射存在于dur中的时间t，%到帧号，如果dur被划分为numFrames。stp =大调的/ numFrames;qt = (t / stp)。* stp;fnum = floor(qt*(numFrames - 1)/dur) + 1;结束

函数(损失,cnnState rnnState, Y3) = forwardAll(模型、X T)%FORWARDALL模型通过CNN和RNN网络的正向传递% [loss,cnnState,rnnState] = forwardAll(模型，X,T)传递预测器X%通过模型，并返回损失和网络的状态%的模型。该模型是一个包含CNN网络和RNN的结构%网络。%% [loss,cnnState,rnnState,Y] = forwardAll(model,X,T)也返回最终结果模型Y. %的预测。%通过CNN传递预测器。向前(cnnState日元)= (model.CNN X);%标记从CNN输出的尺寸，以供RNN消费。Y2 = dlarray(日元“TCUB”);%通过RNN传递预测器。向前(Y3, rnnState) = (model.RNN, Y2);%计算损失。损失= seldNetLoss (Y3、T);结束

函数(损失、Y3) = predictAll(模型、X T)%PREDICTALL通过CNN和RNN网络的模型预测%[损耗，预测]= predictAll(模型，X,T)将预测器X通过%的模型，并返回损失和模型预测。模型是一个%结构包含一个CNN网络和一个RNN网络。%通过CNN传递预测器。日元=预测(model.CNN X);%标记从CNN输出的尺寸，以供RNN消费。Y2 = dlarray(日元“TCUB”);%通过RNN传递预测器。Y3 =预测(model.RNN, Y2);%计算损失。损失= seldNetLoss (Y3、T);结束

函数损失= predictBatch(模型,兆贝可)%PREDICTBATCH计算小批处理队列的丢失% loss = predictBatch(model,mbq)返回由。计算的总损失%通过模型传递小批处理队列的全部内容。%重置小批队列并初始化计数器。复位(mbq)损耗= 0;n = 0;而hasdata(兆贝可)%从小批队列中读取预测器和目标。[X, T] =下一个(兆贝可);将小批量产品通过模型并计算损失。lss = predictAll(模型、X T);lss = lss /大小(T, 2);%更新总损失。损失=损失+ lss;%数据点总数。N = N + 1;结束%用累计的总损失除以小批量的数量。损失= / n损失;结束

函数(损失、渐变、状态)= modelLoss(模型、X T)%MODELLOSS计算模型损耗、梯度和网络状态%[损耗，梯度，状态]= modelLoss(模型，X,T)通过%预测X通过模型，并返回损失，梯度，和%模型中网络的状态。该模型是一个包含CNN网络和RNN网络。%通过模型传递预测器。(损失、cnnState rnnState] = forwardAll(模型、X T);%隔离可学习对象。allGrad。CNN = model.CNN.Learnables;allGrad。RNN = model.RNN.Learnables;状态。CNN = cnnState;状态。RNN = rnnState;%计算梯度。梯度= dlgradient(损失、allGrad);结束

函数损失= seldNetLoss (Y, T)%SELDNETLOSS计算DOA或SED模型的SELDnet损失函数% loss = seldNetLoss(Y,T)返回给定预测Y和的SELDnet损失%目标t损失函数取决于网络(DOA或SED)。的%网络由目标的维数推断。对于DOA网络，%损失函数为均方误差。对于SED网络，损失%函数是crossentropy。%确定目标是否对应于DOA网络或SED%网络。isDOAModel =大小(T, find (dim (T) = =“C”) = = 3;如果isDOAModel%计算MSE损失。doaLoss = mse (Y, T);doaLossFactor = 2 / (size(Y,1) * size(Y,3));loss = doaLoss * doaLossFactor;%与原始实现保持一致其他的%计算交叉熵损失。损失= crossentropy (Y, T, TargetCategories =“独立”NormalizationFactor =“所有元素”);结束loss = loss * size(T,2);结束

函数(r cm) = computeMetrics (sedModel、doaModel sedMBQ, doaMBQ, params, nvargs)%COMPUTEMETRICS计算性能指标% [r,cm] = computeMetrics(sedModel,doaModel,sedMBQ,doaMBW,params)返回根据SED和DOA计算的性能指标结构%验证小批队列，和混淆矩阵cm有效SED%的地区。参数sedModel doaModel sedMBQ doaMBQ params nvargs.;ApplySmoothing = false;结束%初始化计数器。TP = 0;FP = 0;FN = 0;= 0;ct = 0;呃= 0;sedYAll = [];sedTAll = [];%遍历所有数据。重置(sedMBQ)重置(doaMBQ)而hasdata (sedMBQ)%获取SED模型的预测器、目标和预测。[sedXb, sedTb] =下一个(sedMBQ);[~, sedYb] = predictAll (sedModel、sedXb sedTb);sedTb = extractdata(收集(sedTb));sedYb = extractdata(收集(sedYb));%获取DOA模型的预测器、目标和预测。[doaXb, doaTb] =下一个(doaMBQ);[~, doaYb] = predictAll (doaModel、doaXb doaTb);doaTb = extractdata(收集(doaTb));doaYb = extractdata(收集(doaYb));doaYb = doaYb * params.DOA.ScaleFactor;doaTb = doaTb * params.DOA.ScaleFactor;%对小批进行循环。为批= 1:尺寸(sedYb, 2)%隔离当前数据点的预测器和目标。sedY =挤压(sedYb(:,批:));sedT =挤压(sedTb(:,批:));doaY =挤压(doaYb(:,批:));衰老=挤压(doaTb(:,批:));%如果一个帧的SED预测都是用low做出的%置信度(低于阈值)，假设没有声音%的来源。isActive = ~(总和(双(sedY < params.SilenceThreshold), 1) = =大小(sedY,1));%将SED预测器和目标从单一热点向量转换为%标量。[~, sedY] = max (sedY [], 1);sedY = sedY。* isActive;[isActive, sedT] = max (sedT [], 1);sedT = sedT。* isActive;%平稳输出。如果nvargs。ApplySmoothing [doaY,sedY] = smoothOutputs(doaY,sedY,params);结束%执行位置敏感检测。[tp, fp, fn、e、c] = locationSensitiveDetection (sedY、sedT doaY,衰老,params);%累积性能指标。TP = TP + TP;FP = FP + FP;FN = FN + FN;Err = Err + e;Ct = Ct + c;sedYAll = [sedYAll sedY.*isActive];% #好< AGROW >sedTAll = [sedTAll sedT.*isActive];% #好< AGROW >结束It = It + 1;结束%计算性能指标。r.precision = TP/(TP + FP + eps);r.recall = TP / (TP + FN + eps);r.f1Score = 2 * (r.precision * r.recall) / (r。Precision + r.recall + eps);r.avgErr =呃/ ct;%计算混淆矩阵。confmat = confusionmat (sedTAll、单(sedYAll),订单= 0:14);厘米= confmat(2:结束,2:结束);%从混淆矩阵中去除沉默。结束

函数[TP, FP, FN、totErr ct) = locationSensitiveDetection (sedY、sedT doaY,衰老,params)%LOCATIONSENSITIVEDETECTION位置敏感检测% (TP、FP FN、totErr ct) =% locationSensitiveDetection(sedY,sedT,doaY,doaT,params)计算%真阳性，假阳性，假阴性，DOA总错误，和活动目标的%数量。每个指标的定义在%[4]。%计算距离。dist = vecnorm (doaY-doaT);%判断声音是否活跃，以供参考和预测。isReferenceActive = sedT ~ = 0;isPredictedActive = sedY ~ = 0;%计算引用活动段的总DOA误差。totErr = (dist。* isReferenceActive)之和;%统计活动目标的总数。ct =总和(isReferenceActive);%确定DOA是否在每帧阈值内。isDOAnear = dist < params.SpatialThreshold;%真阳性:TP = sum(isDOAnear & isReferenceActive & isPredictedActive & (sedT==sedY));%假阳性:FP1 = sum(~isReferenceActive & isPredictedActive);FP2 = sum(isReferenceActive & isPredictedActive & (sedT~=sedY | ~isDOAnear));Fp = fp1 + fp2;%假阴性:FN1 = sum(isReferenceActive & ~isPredictedActive);FN2 = sum(isReferenceActive & (sedT~=sedY | ~isDOAnear));Fn = fn1 + fn2;结束

函数[doaYSmooth, sedYSmooth] = smoothOutputs (doaY、sedY params)%SMOOTHOUTPUTS随时间的平滑DOA和SED预测% [doaYSmooth,sedYSmooth] = smoothOutputs(doaY,sedY,params)平滑DOA%和SED随时间的预测。%预分配平滑输出。doaYSmooth = doaY;sedYSmooth = sedY;%聚类DOA预测。集群= = clusterdata (doaY”标准“距离”,截止= params.SpatialThreshold);stt = 1;新奥集团= 1;而新奥集团< = params.Targets.NumFrames如果Clusters (stt) == Clusters (enn) enn = enn + 1;其他的doaYSmooth (: stt: enn-1) = smoothDOA (doaY (:, stt: enn-1));sedYSmooth (: stt: enn-1) =平滑(sedY (:, stt: enn-1));stt =新奥集团;结束结束doaYSmooth (: stt: enn-1) = smoothDOA (doaY (:, stt: enn-1));sedYSmooth (: stt: enn-1) =平滑(sedY (:, stt: enn-1));sedYSmooth =圆(movmedian (sedYSmooth 5));结束

函数平滑= smoothDOA(块)SMOOTHDOA平滑的DOA预测% smoothened = smoothDOA(chunk)通过替换%的值，每个轴的平均值在块。的意思是丢弃数据的上下四分之一后计算的%。%确定块的长度，然后索引切出中间%一半的数据。chlen =大小(块,2);圣= max(圆(chlen * 1/4), 1);en = max(圆(chlen * 3/4), 1);%对空间轴(列)进行排序。昏暗的=排序(块,2);%取内半部的平均值。平滑= repmat(意思是(暗(:,圣:en), 2), 1, chlen);结束

函数平滑=平滑(块)%SMOOTHSED SED平滑预测% smoothened = smoothSED(chunk)使用模式平滑SED预测。平滑= repmat(模式(块),1,大小(块,2));结束