downloadFolder = matlab.internal.examples.download金宝appSupportFile(“音频”，“L3DAS21_ov1.zip”）;dataFolder = tempdir;unzip(下载文件夹，数据文件夹)dataset = fullfile(数据文件夹，数据文件夹)“L3DAS21_ov1”）;

speedupExample =假；

adsTrain = audioDatastore(fullfile(dataset，“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(dataset，“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))“麦克风”) ylabel (“W”nexttile plot(micB(:，1))B“麦克风”nexttile plot(micA(:，2))“X”nexttile plot(micB(:，2)) nexttile plot(micA(:，3))“Y”nexttile plot(micB(:，3))“Z”nexttile plot(micB(:，4))

麦克风=1；频道=1；时间=10；Fs = 32e3;数据的已知采样率。s = [micA,micB];数据= s(1:轮(时长*fs)，通道+(麦克风-1)*4);声音(数据、fs)

keySet = [“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 = {1,2,3,4,5,6,7,8,9,10,11,12,13,14};参数个数。SoundClasses = containers.Map(keySet,valueSet);

[folder,fn] = fileparts(adsTrainA.Files);targetPath = fullfile(strrep(文件夹，filesep+“数据”filesep +“标签”)，“label_”+ strrep (fn,“_A”，＂＂) +. csv”）;ttdsTrain = tabularTextDatastore(targetPath);labelTable =预览(ttdsTrain)

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

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

dsTTrain = transform(ttdsTrain，@(x)extractSEDTargets(x,params));sedTTrain = readall(dsTTrain);[folder,fn] = fileparts(adsValidationA.Files);targetPath = fullfile(strrep(文件夹，filesep+“数据”filesep +“标签”)，“label_”+ strrep (fn,“_A”，＂＂) +. csv”）;ttdsValidation = tabularTextDatastore(targetPath);dsTValidation = transform(ttdsValidation，@(x)extractSEDTargets(x,params));sedTValidation = readall(dsTValidation);

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

dsTTrain = transform(ttdsTrain，@(x)extractDOATargets(x,params));doaTTrain = readall(dsTTrain);[folder,fn] = fileparts(adsValidationA.Files);targetPath = fullfile(strrep(文件夹，filesep+“数据”filesep +“标签”)，“label_”+ strrep (fn,“_A”，＂＂) +. csv”）;ttdsValidation = tabularTextDatastore(targetPath);dsTValidation = transform(ttdsValidation，@(x)extractDOATargets(x,params));doaTValidation = readall(dsTValidation);

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

stftfeat = extractSTFT({micA,micB}，params);[numfeatered, numframered, numchannelsed] = size(stftfeat)

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

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

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

trainoptionsed = struct(.．.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

trainseddmbq = 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([numfeatered, numframered,numchannelsSED]，Normalization= . seldnetCNNLayers = [imageInputLayer([numfeatered, numframered, numchannelsed]，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。美国有线电视新闻网;sedModel。RNN = netRNN;

迭代= 0;averageGrad = [];averageSqGrad = [];Epoch = 0;bestLoss = Inf;badEpochs = 0;learnRate = trainOptionsSED.InitialLearnRate;

pp = progressPlotterSELD();

rng (0)而epoch < trainOptionsSED。MaxEpochs && badEpochs < trainoptionsed .;ValidationPatience epoch = epoch + 1;打乱迷你批处理队列。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] = next(trainSEDmbq);使用dlfeval和modelLoss函数评估模型梯度和损失。[loss,grad,state] = dlfeval(@modelLoss,sedModel,X,T);损耗=损耗/尺寸(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, averageqgrad] = adamupdate(sedModel,grad,averageGrad，.．.迭代,averageSqGrad learnRate、trainOptionsSED.GradientDecayFactor trainOptionsSED.SquaredGradientDecayFactor);更新培训进度图。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。loss = 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)如果验证损失得到改善，则创建检查点。如果验证%损失没有改善，增加了坏epoch的数量。如果损失< bestLoss bestLoss =损失;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 . learnratdroppperiod)==0 learnRate = learnRate* trainoptionssed . learnratdroppfactor;结束结束

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

gccphatfeat = extractGCCPHAT({micA,micB}，params);[numfeaturesDOA,timestepsDOA,numchannelsDOA] = size(gccphatfeat)

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

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

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

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

trainOptionsDOA = trainoptionsed;

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]，归一化=“没有”、名称=“输入”）;seldnetCNNLayers(5) = maxPooling2dLayer([3,2]，Stride=[3,2]，Padding=“相同”、名称=“maxpool1”）;netCNN = dlnetwork(layerGraph(seldnetCNNLayers));seldnetGRULayers(11) = fullyConnectedLayer(3,Name=“fc4”）;seldnetGRULayers(12) = tanhLayer(Name=“输出”）;netRNN = dlnetwork(layerGraph(seldnetGRULayers));

doaModel。美国有线电视新闻网;doaModel。RNN = netRNN;

迭代= 0;averageGrad = [];averageSqGrad = [];Epoch = 0;bestLoss = Inf;badEpochs = 0;learnRate = trainOptionsDOA.InitialLearnRate;

pp = progressPlotterSELD();

rng (0)而epoch < trainOptionsDOA。MaxEpochs && badEpochs < trainOptionsDOA。ValidationPatience epoch = epoch + 1;打乱迷你批处理队列。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] = next(trainDOAmbq);使用dlfeval和modelLoss函数评估模型梯度和损失。[loss,grad,state] = dlfeval(@modelLoss,doaModel,X,T);损耗=损耗/尺寸(T,2);%更新状态。domodel . cnn . state = state.CNN;dommodell . rnn . state = state.RNN;使用Adam优化器更新网络参数。。[doaModel,averageGrad,averageSqGrad] = adamupdate(doaModel,grad,averageGrad，.．.迭代,averageSqGrad learnRate、trainOptionsDOA.GradientDecayFactor trainOptionsDOA.SquaredGradientDecayFactor);更新培训进度图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loss = 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)如果验证损失得到改善，则创建检查点。如果验证%损失没有改善，增加了坏epoch的数量。如果损失< bestLoss bestLoss =损失;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 . learnratdroppperiod)==0 learnRate = learnRate* trainoptionsdoa . learnratdroppfactor;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;

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

[results,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)提取到达方向(DOA)目标% T = extractDOATargets(fileName,params)解析CSV文件返回一个矩阵t。目标矩阵是一个n × 3的矩阵%矩阵，其中N对应帧数，3对应%表示三维空间中位置的3个轴。预分配目标矩阵。一个全为0的坐标系对应着没有声音%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，使其介于-1和1之间(tanh的边界)激活层)。将目标包装在单元格数组中以方便批处理%处理。T = {T/params.DOA.ScaleFactor};结束

函数T = extractSEDTargets(csvFile,params)提取声音事件检测(SED)目标% T = extractSEDTargets(fileName,params)解析CSV文件返回SED目标t的矩阵。目标矩阵是n × k的%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。一个全为0的坐标系对应着没有声音%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，填入目标矩阵的相应列%，带1，表示声音类出现在该帧中。为ii = 1:size(startendFrame,1) classID = params.SoundClasses(csvFile.Class{ii});T(startendFrame(ii,1):startendFrame(ii,2)-1,classID) = 1;结束将目标包装在单元格数组中，以便于批处理。T = {T};结束

函数X = extractSTFT(s,params)提取居中STFT的对数大小% X = extractSTFT({s1,s2}，params)连接s1和s2，然后%提取单边对数量级STFT。信号是预先填充的% STFT，使第一个窗口以第一个样本为中心。的%输出被修剪以去除第一个(DC)系数和最后一个%频谱。输入参数定义STFT。沿着第二个(信道)维度连接信号。音频= cat(2,s{:});提取居中的STFT。N = number (params.SED.Window);overlapLength = N - params.SED.HopLength;S = centeredSTFT(音频，params.SED.Window,overlapLength,N);从所有频谱中修剪第一个系数，并修剪最后一个频谱。S = S(2:end,1:end-1，:);%转换为对数大小。使用偏移量来防止log为零。mag = log(abs(S) + eps);%铸造输出单精度。X = single(mag);结束

函数X = extractGCCPHAT(s,params)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%为所有通道对提取GCC-PHAT。连接两个麦克风对应的信号。音频= cat(2,s{:});统计输入通道总数。nChan = size(音频，2);计算输出通道总数。numOutputChannels = nchoosek(nChan,2);预先分配一个NumFeatures-by-NumFrames-by-NumChannels特性(预测器)%的矩阵。numFrames = size(audio,1)/params.DOA.HopLength;X = 0 (params.DOA.NumBands,numFrames,numOutputChannels);% -----------------------------------计算每对通道的GCC-PHAT。为每个通道预计算STFT。N = number (params.DOA.Window);overlapLength = N - params.DOA.HopLength;micAB_stft = centeredSTFT(音频，params.DOA.Window,overlapLength,N);conmicab_stft = conj(micAB_stft(:，:，2:end));Idx = 1;为2 = 1: nChan - 1 R = micAB_stft (:,:, ii)。* conjmicAB_stft (:,:, ii:结束);R = exp(1i .*角度(R));R = padarray(R, N/2 - 1，“职位”）;gcc = fftshift(ifft(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% -----------------------------------%铸造输出单精度。X =单(X);结束

函数s = centeredSTFT(音频，win,overlapLength,fftLength)居中STFT% s = centeredSTFT(audioIn,win,overlapLength,fftLength)计算一个STFT%，第一个窗口以第一个示例为中心。两端是%填充反射音频信号。%输入信号的前面和后面。firstR = flip(音频(1:fftLength/2，:)，1);lastR = flip(audio(end - fftLength/2 + 1:end，:)，1);sig = cat(1,firstR,audio,lastR);%执行STFT。s = stft(sig,Window=win,OverlapLength= OverlapLength,FFTLength= FFTLength,FrequencyRange=“单向的”）;结束

函数fnum = time2frame(t,dur,numFrames)将时间戳转换为帧号% fnum = time2frame(t,dur,numFrames)映射在dur中存在的时间t，如果dur被划分为numFrames，则%转换为帧号。stp = dur/numFrames;Qt = round(t /stp).*stp;fnum = floor(qt*(numFrames - 1)/dur) + 1;结束

函数[loss,cnnState,rnnState,Y3] = forwardAll(model,X,T)FORWARDALL模型通过CNN和RNN网络的前向传递% [loss,cnnState,rnnState] = forwardAll(model,X,T)传递预测因子X，并返回网络的损耗和状态%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)也返回最终值Y模型的%预测。%通过CNN传递预测器。[Y1,cnnState] = forward(model.CNN,X);标记从CNN输出的尺寸供RNN使用。Y2 = darray (Y1，“TCUB”）;%通过RNN传递预测器。[Y3,rnnState] = forward(model.RNN,Y2);计算损失。损失= seldNetLoss(Y3,T);结束

函数[损失，Y3] = predictAll(模型，X,T)通过CNN和RNN网络进行模型预测% [loss,prediction] = predictAll(model,X,T)传递预测因子X%的模型，并返回损失和模型预测。这个模型是结构，包含CNN网络和RNN网络。%通过CNN传递预测器。Y1 = predict(model.CNN,X);标记从CNN输出的尺寸供RNN使用。Y2 = darray (Y1，“TCUB”）;%通过RNN传递预测器。Y3 = predict(model.RNN,Y2);计算损失。损失= seldNetLoss(Y3,T);结束

函数损失= predictBatch(模型，mbq)计算迷你批队列的损失% loss = predictBatch(model,mbq)返回由%通过模型传递迷你批处理队列的全部内容。重置迷你批处理队列并初始化计数器。复位(mbq)损耗= 0;N = 0;而hasdata(兆贝可)从迷你批处理队列中读取预测器和目标。[X,T] = next(mbq);通过模型传递迷你批并计算损失。lss = predictAll(model,X,T);lss = lss/size(T,2);%更新总损失。损失=损失+ lss;%数据点总数。N = N + 1;结束用累计总损耗除以小批数量。损耗=损耗/n;结束

函数[loss,gradients,state] = modelLoss(model,X,T)计算模型丢失、梯度和网络状态% [loss,gradients,state] = modelLoss(model,X,T)通过%预测因子X通过模型，并返回损失，梯度和%模型中网络的状态。模型是一个包含CNN网络和RNN网络。通过模型传递预测因子。[loss,cnnState,rnnState] = forwardAll(model,X,T);把可学习的东西分离出来。allGrad。CNN = model.CNN.Learnables;allGrad。RNN =模型。RNN. learnables;状态。CNN = cnnState;状态。RNN = rnnState;计算梯度。gradients = dlgradient(loss,allGrad);结束

函数= seldNetLoss(Y,T)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SELDnet损失损失函数依赖于网络(DOA或SED)。的%网络是由目标的维度推断出来的。对于DOA网络，损失函数为均方误差。对于SED网络，损失%函数是交叉的。确定目标是否对应于DOA网络或SED%网络。isDOAModel = size(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));损失= doaLoss * doaLossFactor;与原始实现保持一致其他的计算交叉熵损失。loss = crossentropy(Y,T,TargetCategories=“独立”NormalizationFactor =“所有元素”）;结束损耗=损耗*尺寸(T,2);结束

函数[r,cm] = computeMetrics(sedModel,doaModel,sedMBQ,doaMBQ,params,nvargs)计算性能指标% [r,cm] = computeMetrics(sedModel,doaModel,sedMBQ,doaMBW,params)返回通过SED和DOA计算的性能指标结构。%验证迷你批处理队列，和一个混淆矩阵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;它= 0;Ct = 0;Err = 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] = next(sedMBQ);[~，sedYb] = predictAll(sedModel,sedXb,sedTb);sedTb = extractdata(gather(sedTb));sedYb = extractdata(收集(sedYb));获取DOA模型的预测器、目标和预测。。[doaXb,doaTb] = next(doaMBQ);[~，doaYb] = predictAll(doaModel,doaXb,doaTb);doaTb = extractdata(gather(doaTb));doaYb = extractdata(gather(doaYb));doaYb = doaYb*params.DOA.ScaleFactor;doaTb = doaTb*params.DOA.ScaleFactor;遍历小批。为batch = 1:size(sedYb,2)隔离当前数据点的预测因子和目标。sedY = squeeze(sedYb(:，batch，:));sedT = squeeze(sedTb(:，batch，:));doaY = squeeze(doaYb(:，batch，:));doaT = squeeze(doaTb(:，batch，:));%如果一个帧的SED预测都是用low%置信度(低于阈值)，假设没有声音%源当前。isActive = ~(sum(double(sedY将SED预测器和目标从单一热向量转换为%标量。[~，sedY] = max(sedY，[]，1);sedY = sedY.*isActive;[isActive,sedT] = max(sedT，[]，1);sedT = sedT.*isActive;平滑输出。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,doaT,params);积累绩效指标。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。精度+ r.recall + eps);r.avgErr = err/ct;计算混淆矩阵。confmat = confusionmat(sedTAll,single(sedYAll)，Order=0:14);Cm = confmat(2:end,2:end);从混乱矩阵中移除沉默。结束

函数[TP,FP,FN,totErr,ct] = locationSensitiveDetection(sedY,sedT,doaY,doaT,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总误差，和活动目标的百分比。中提供了每个度量的定义%[4]。计算距离。dist = vecnorm(doaY-doaT);确定听起来是否活跃，以供参考和预测。。isReferenceActive = sedT~=0;isPredictedActive = sedY~=0;计算引用活动部分的总DOA误差。totErr = sum(dist.*isReferenceActive);%活动目标总数。ct = sum(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)平滑DOA和SED随时间的预测% [doaYSmooth,sedYSmooth] = smoothOutputs(doaY,sedY,params)平滑DOA%和SED预测随时间的变化。预分配平滑输出。doaYSmooth = doaY;sedYSmooth = sedY;聚类DOA预测。clusters = clusterdata(doaY'，Criterion=“距离”,截止= params.SpatialThreshold);STT = 1;Enn = 1;而enn <= params.Targets.NumFrames如果集群(stt) ==集群(enn) enn = enn + 1;其他的doaYSmooth(:，stt:enn-1) = smoothDOA(doaY(:，stt:enn-1));sedYSmooth(:，stt:enn-1) = smoothSED(sedY(:，stt:enn-1));STT = enn;结束结束doaYSmooth(:，stt:enn-1) = smoothDOA(doaY(:，stt:enn-1));sedYSmooth(:，stt:enn-1) = smoothSED(sedY(:，stt:enn-1));sedYSmooth = round(movmedian(sedYSmooth,5));结束

函数平滑=平滑doa(块)平滑DOA平滑DOA预测平滑DOA(chunk)平滑DOA预测每个轴的%值与该轴在数据块中的平均值。均值是丢弃数据的上下四分之一后计算的%。确定块的长度，然后将中间的索引剪掉%一半的数据。Chlen = size(chunk,2);St = max(round(chlen*1/4)，1);En = max(round(chlen*3/4)，1);对空间轴(列)进行排序。Dim = sort(chunk,2);取内半部分的均值。Smoothed = repmat(mean(dim(:，st:en)，2)，1,chlen);结束

函数光滑的(块)平滑SED预测% smooened = smoothSED(chunk)使用模式平滑SED预测。smoothened = repmat(mode(chunk)，1,size(chunk,2));结束