使用深度学习语音活动检测噪声

fs = 16 e3;[演讲,fileFs] = audioread (“Counting-16-44p1-mono-15secs.wav”);演讲=重新取样(演讲、fs、fileFs);演讲= speech. / max (abs(演讲));声音(演讲中,fs) detectSpeech(演讲、fs、窗口=汉明(0.04 * fs,“周期”),MergeDistance =圆(0.5 * fs))

(噪音、fileFs) = audioread (“洗衣机- 16 - 8 mono - 200 - secs.mp3”);噪音=重新取样(噪音、fs、fileFs);

信噪比=-10年;noisySpeech = mixSNR(演讲、噪声、信噪比);声音(noisySpeech fs)

detectSpeech (noisySpeech fs =汉明窗(0.04 * fs,“周期”),MergeDistance =圆(0.5 * fs))

downloadFolder = matlab.internal.examples.download万博1manbetxSupportFile (“音频”,“VoiceActivityDetection.zip”);dataFolder = tempdir;解压缩(downloadFolder dataFolder) netFolder = fullfile (dataFolder,“VoiceActivityDetection”);pretrainedNetwork =负载(fullfile (netFolder“voiceActivityDetectionExample.mat”));afe = pretrainedNetwork.afe;网= pretrainedNetwork.speechDetectNet;

afe

net.Layers

特点=提取(afe noisySpeech);特性=(功能-意味着(功能,1))。/性病(特性,[],1);特点=功能”;

decisionsCategorical =分类(净、特点);

决定=(双(decisionsCategorical) - 1) ';decisionsPerSample =[决定(1:圆形(元素个数(afe.Window) / 2)); repelem(决策,元素个数(afe.Window) -afe.OverlapLength, 1)];tiledlayout (2, 1) nexttile detectSpeech(演讲、fs、窗口=汉明(0.04 * fs,“周期”),MergeDistance =圆(0.5 * fs))标题(“地面实况和监督”)包含(”“)= signalMask nexttile面具(decisionsPerSample SampleRate = fs,类别=“活动”);plotsigroi(面具,noisySpeech,真正的)标题(“预测和监督”)

generateMATLABFunction (afe“featureExtractor”IsStreaming = true)

audiowrite (“Speech.wav”演讲中,fs) audiowrite (“Noise.wav”、噪音、fs)

信号=“Speech.wav”;噪音=“Noise.wav”;信噪比=-10年;% dBtestDuration =20.;%秒playbackSource =“吵”;

afe streamingDemo(净,…信号,噪声,信噪比,…testDuration playbackSource);

函数streamingDemo(净,afe、信号、噪声、信噪比、testDuration, playbackSource)% streamingDemo(净、afe、信号、噪声、信噪比、testDuration, playbackSource)%实时VAD演示。%创建dsp。AudioFileReader对象读取语音和噪声文件框架%的框架。如果语音信号被指定为麦克风,使用一个% audioDeviceReader作为源。如果strcmpi(信号,“麦克风”)speechReader = audioDeviceReader (afe.SampleRate);其他的speechReader = dsp.AudioFileReader(信号,PlayCount =正);结束noiseReader = dsp.AudioFileReader(噪音、PlayCount =正无穷,SamplesPerFrame = speechReader.SamplesPerFrame);fs = speechReader.SampleRate;%创建一个dsp。MovingStandardDeviation对象和一个dsp.MovingAverage%的对象。您将使用这些来确定标准差和均值%的音频功能标准化。统计应该改善随着时间的推移%。movSTD = = dsp.MovingStandardDeviation(方法“指数权重”ForgettingFactor = 1);movMean = = dsp.MovingAverage(方法“指数权重”ForgettingFactor = 1);%创建一个dsp。MovingMaximum对象。你会用它来规范%的音频。movMax = dsp.MovingMaximum (SpecifyWindowLength = false);%创建一个dsp。MovingRMS对象。您将使用这个来确定信号%和噪声混合在所需的信噪比。这个对象仅仅是有用的%的目的你在哪里人为添加噪音。movRMS = = dsp.MovingRMS(方法“指数权重”ForgettingFactor = 1);%创建三个dsp。AsyncBuffer对象。一个缓冲区输入音频,一个%缓冲提取的特征,一个缓冲输出音频% VAD决定对应的音频信号。输出缓冲区%只有实时可视化决策所必需的。audioInBuffer = dsp.AsyncBuffer (2 * speechReader.SamplesPerFrame);featureBuffer = dsp.AsyncBuffer(装天花板(2 * speechReader.SamplesPerFrame /(元素个数(afe.Window) -afe.OverlapLength)));audioOutBuffer = dsp.AsyncBuffer (2 * speechReader.SamplesPerFrame);%创建时间范围可视化原始语音信号,噪声网络应用于%信号,决定输出%网络。范围= timescope (SampleRate = fs,…TimeSpanSource =“财产”,…时间间隔= 3,…BufferLength = f * 3 * 3,…TimeSpanOverrunAction =“滚动”,…AxesScaling =“更新”,…MaximizeAxes =“上”,…AxesScalingNumUpdates = 20,…NumInputPorts = 3,…LayoutDimensions = (3,1),…ChannelNames = [《吵闹的演讲》,“干净的演讲(原始)”,“发现演讲”),……ActiveDisplay = 1,…ShowGrid = true,……ActiveDisplay = 2,…ShowGrid = true,……ActiveDisplay = 3,…ShowGrid = true);% #好< DUPNAMEARG >设置(范围、{1 1 1})%创建一个audioDeviceWriter对象原始或吵闹%音频扬声器。deviceWriter = audioDeviceWriter (SampleRate = fs);%初始化变量在循环中使用。windowLength =元素个数(afe.Window);hopLength = windowLength - afe.OverlapLength;%流示范运行。loopTimer =抽搐;而toc (loopTimer) < testDuration%读取一帧语音信号和噪声信号的帧speechIn = speechReader ();noiseIn = noiseReader ();%将语音和噪声在指定的信噪比能量= movRMS ([speechIn noiseIn]);noiseGain = 10 ^(信噪比/ 20)*(1),/能源(,2);noisyAudio = speechIn + noiseGain * noiseIn;%更新运行最大范围内音频myMax = movMax (abs (noisyAudio));noisyAudio = noisyAudio / myMax(结束);%写吵闹的音频和语音缓冲区写(audioInBuffer [noisyAudio speechIn]);%如果有足够多的样本在音频缓冲区计算功能%向量,看样品,规范化,提取特征%向量,并编写最新的特性向量缓冲器的特性。而(audioInBuffer。NumUnreadSamples > = hopLength) x =阅读(audioInBuffer,元素个数(afe.Window) afe.OverlapLength);写(audioOutBuffer x (end-hopLength + 1:,:));noisyAudio = x (: 1);特点= featureExtractor (noisyAudio);写(featureBuffer、特点);结束如果featureBuffer。NumUnreadSamples > = 1%读取音频数据对应数量的未读%的特征向量。audioHop =阅读(audioOutBuffer, featureBuffer.NumUnreadSamples * hopLength);%阅读所有未读的特征向量。特点=阅读(featureBuffer);%只使用新功能更新标准差和%的意思。规范化的特点。rmean = movMean(特性);rstd = movSTD(特性);特性=(功能- rmean(最终,:))。/ rstd(最终:);%网络推理[净,决定]= classifyAndUpdateState(净,功能);%将决定每个特征向量决定样本决定= repelem(决定hopLength 1);%面具应用到可视化的嘈杂的演讲vadResult = audioHop (: 1);vadResult(决定= =分类(0))= 0;% +噪音听演讲或演讲开关playbackSource情况下“清洁”deviceWriter (audioHop (:, 2));情况下“吵”deviceWriter (audioHop (: 1));情况下“detectedSpeech”deviceWriter (vadResult);结束%可视化语音+噪音,原来的演讲,和声音%活动检测。范围(audioHop (: 1), audioHop (:, 2), vadResult)结束结束结束

函数[noisySignal, requestedNoise] = mixSNR(信号、噪声、比例)% (noisySignal requestedNoise] = mixSNR(信号、噪音比)返回一个吵了%的信号,noisySignal。吵闹的信号混合了%指定比率的噪音分贝。numSamples =大小(信号,1);%将噪声转化为mono噪音=意味着(噪音,2);%修剪或扩大噪声与信号的大小如果(噪音,1)> = numSamples大小%选择一个随机指数,这样你还有numSamples开始%索引后的噪音。开始=兰迪(大小(噪音,1)- numSamples + 1);噪音=噪音(启动:启动+ numSamples-1);其他的numReps =装天花板(numSamples /大小(噪音,1));temp = repmat(噪音numReps 1);开始=兰迪(大小(temp, 1) - numSamples - 1);噪音= temp(启动:启动+ numSamples-1);结束signalNorm =规范(信号);noiseNorm =规范(噪音);goalNoiseNorm = signalNorm /(10 ^(比率/ 20));因素= goalNoiseNorm / noiseNorm;requestedNoise =噪音。*因素;noisySignal = + requestedNoise信号;noisySignal = noisySignal. / max (abs (noisySignal));结束