距跟踪使用多个音高估计,嗯

[x, fs] = audioread (“Counting-16-44p1-mono-15secs.wav”);负载TruePitch.mattruePitch

(f0、loc) =音高(x, fs);

isVoiced = ~ isnan (truePitch);f0 (~ isVoiced) =南;p = 0.1;GPE =意味着(abs (f0 (isVoiced) -truePitch (isVoiced)) > truePitch (isVoiced)。* p)。* 100;t =(0:长度(x) 1) / fs;t0 = (locs-1) / fs;声音(x, fs)图(1)tiledlayout (2, 1) nexttile情节(t, x) ylabel (“振幅”)标题(“估计清洁信号”)nexttile情节(t0, [truePitch, f0])传说(“参考”,“估计”位置=“西北”)ylabel (“F0 (Hz)”)包含(“时间(s)”)标题(" GPE = "+圆(GPE、2) +“(%)”)

desiredSNR = 5;x = mixSNR (x,兰德(大小(x)), desiredSNR);(f0、loc) =音高(x, fs);f0 (~ isVoiced) =南;GPE =意味着(abs (f0 (isVoiced)——truePitch (isVoiced)) > truePitch (isVoiced)。* p)。* 100;声音(x, fs)图(2)tiledlayout (2, 1) nexttile情节(t, x) ylabel (“振幅”)标题(“估计噪声信号”)nexttile情节(t0, [truePitch, f0])传说(“参考”,“估计”位置=“西北”)ylabel (“F0 (Hz)”)包含(“时间(s)”)标题(" GPE = "+ GPE +“(%)”)

帮助HelperPitchTracker

范围= [50400];HOPLENGTH =圆(fs。* 0.01);%缓冲成所需大小xBuff_SRH =缓冲区(x,圆(0.025 * fs),圆(0.02 * fs),“nodelay”);%定义距参数params_SRH = =结构(方法“SRH”,…范围=范围,…WindowLength =圆(fs * 0.06),…OverlapLength =圆(fs * 0.06 -hoplength),…SampleRate = fs,…NumChannels =大小(x, 2),…SamplesPerChannel =大小(x, 1));multiCandidate_params_SRH =结构(NumCandidates = 5, MinPeakDistance = 1);%得到音高估计和信心[f0_SRH, conf_SRH] = audio.internal.pitch.SRH (xBuff_SRH x,…params_SRH,…multiCandidate_params_SRH);

图(3)tiledlayout (2, 1) nexttile情节(t0, f0_SRH) ylabel (“F0候选人(Hz)”)标题(“从SRH音高估计多个候选人”conf_SRH) nexttile情节(t0) ylabel (“相对的信心”)包含(“时间(s)”)

xBuff_PEF =缓冲区(x,圆(0.06 * fs),圆(0.05 * fs),“nodelay”);params_PEF = =结构(方法“PEF”,…范围=范围,…WindowLength =圆(fs * 0.06),…OverlapLength =圆(fs * 0.06 -hoplength),…SampleRate = fs,…NumChannels =大小(x, 2),…SamplesPerChannel =大小(x, 1));multiCandidate_params_PEF =结构(NumCandidates = 5, MinPeakDistance = 5);[f0_PEF, conf_PEF] = audio.internal.pitch.PEF (xBuff_PEF,…params_PEF,…multiCandidate_params_PEF);xBuff_NCF =缓冲区(x,圆(0.04 * fs),圆(0.03 * fs),“nodelay”);xBuff_NCF = xBuff_NCF (: 2: end-1);params_NCF = =结构(方法“NCF”,…范围=范围,…WindowLength =圆(fs * 0.04),…OverlapLength =圆(fs * 0.04 -hoplength),…SampleRate = fs,…NumChannels =大小(x, 2),…SamplesPerChannel =大小(x, 1));multiCandidate_params_NCF =结构(NumCandidates = 5, MinPeakDistance = 1);f0_NCF = audio.internal.pitch.NCF (xBuff_NCF,…params_NCF,…multiCandidate_params_NCF);xBuff_CEP =缓冲区(x,圆(0.04 * fs),圆(0.03 * fs),“nodelay”);xBuff_CEP = xBuff_CEP (: 2: end-1);params_CEP = =结构(方法“CEP”,…范围=范围,…WindowLength =圆(fs * 0.04),…OverlapLength =圆(fs * 0.04 -hoplength),…SampleRate = fs,…NumChannels =大小(x, 2),…SamplesPerChannel =大小(x, 1));multiCandidate_params_CEP =结构(NumCandidates = 5, MinPeakDistance = 1);f0_CEP = audio.internal.pitch.CEP (xBuff_CEP,…params_CEP,…multiCandidate_params_CEP);BackupCandidates = [f0_NCF (: 1), f0_CEP (: 1)];

hr = harmonicRatio (xBuff_PEF fs,…=汉明(窗口大小(xBuff_NCF, 1),“周期”),…OverlapLength = 0);图(4)tiledlayout (2, 1) nexttile情节(t, x) ylabel (“振幅”)nexttile情节(t0、人力资源)ylabel (“谐波比率”)包含(“时间(s)”)

idxToKeep =逻辑(movmedian(人力资源>((3/4)*马克斯(人力资源),3));longTermMedian =值([f0_PEF (idxToKeep 1); f0_SRH (idxToKeep 1)]);低= max ((2/3) * longTermMedian范围(1));上= min ((4/3) * longTermMedian范围(2));图(5)tiledlayout (1,1) nexttile情节(t0, [f0_PEF f0_SRH])在情节(t0, longTermMedian。*的(大小(f0_PEF, 1)),r:“线宽= 3)情节(t0,上层。*的(大小(f0_PEF, 1)),“r”线宽= 2)情节(t0,低。*的(大小(f0_PEF, 1)),“r”线宽= 2)从包含(“时间(s)”)ylabel (“频率(赫兹)”)标题(“长期值”)

无效的= f0_PEF >低| f0_PEF >上;conf_PEF(无效)= conf_PEF(无效)/ 10;[conf_PEF,顺序]=排序(conf_PEF 2“下”);为i = 1:尺寸(f0_PEF, 1) f0_PEF(我:)= f0_PEF(我订单(我,:));结束无效的= f0_SRH >低| f0_SRH >上;conf_SRH(无效)= conf_SRH(无效)/ 10;[conf_SRH,顺序]=排序(conf_SRH 2“下”);为i = 1:尺寸(f0_SRH, 1) f0_SRH(我:)= f0_SRH(我订单(我,:));结束候选人= [f0_PEF (:, 1:2), f0_SRH (: 1:2)];信心= [conf_PEF (:, 1:2), conf_SRH (: 1:2)];图(6)情节(t0,候选人)包含(“时间(s)”)ylabel (“频率(赫兹)”)标题(“减候选人”)

跨度= 5;confidenceFactor = 1;为r = 1:尺寸(候选人,1)为c = 1:尺寸(候选人,2)tf = abs(候选人(r, c)候选人(r,:)) <跨度;候选人(r、c) =意味着(候选人(r, tf));信心(r、c) =总和(信心(r, tf)) * confidenceFactor;结束结束候选人= max (min(候选人,400),50);

负载EmissionMatrix.matemissionMatrixemissionMatrixRange现状后=150年;图(7)情节(emissionMatrixRange (1): emissionMatrixRange (2), emissionMatrix(现状后,emissionMatrixRange(1) 1:))标题(“排放PDF”+现状后+“赫兹”)包含(“下一个音高值(Hz)”)ylabel (“概率”)

%预先配置numPaths = 4;numHops =大小(候选人,1);日志= 0 (numHops 3 numPaths);suspectHops = 0 (numHops, 1);%与octave-smoothing迭代为hopNumber = 1: numHops nowCandidates =候选人(hopNumber:);nowConfidence =信心(hopNumber:);%去除八度跳跃如果hopNumber > 100 numCandidates =元素个数(nowCandidates);%短期加权值medianWindowLength = 50;aTemp =日志(max (hopNumber-min hopNumber medianWindowLength) (1): hopNumber-1, 1:);shortTermMedian =值(aTemp (:));previousM =意味着([longTermMedian shortTermMedian]);lowerTight = max ((4/3) * previousM emissionMatrixRange (1));upperTight = min ((2/3) * previousM emissionMatrixRange (2));numCandidateOutside =总和([nowCandidates < lowerTight nowCandidates > upperTight]);%如果至少1候选人不在的八度为中心%短期中,添加备份的候选人归一化相关函数生成的%和倒频谱%确定算法作为潜在的候选人。如果numCandidateOutside > 0%应用备份音高估计nowCandidates = [nowCandidates BackupCandidates (hopNumber:)];% #好< AGROW >nowConfidence = [nowConfidence repmat(意味着(nowConfidence), 1, 2)];% #好< AGROW >%使独特的跨度= 10;confidenceFactor = 1.2;为r = 1:尺寸(nowCandidates, 1)为c = 1:尺寸(nowCandidates 2) tf = abs (nowCandidates (r, c) -nowCandidates (r,:)) <跨度;nowCandidates (r、c) =意味着(nowCandidates (r, tf));nowConfidence (r、c) =总和(nowConfidence (r, tf)) * confidenceFactor;结束结束结束结束%建立信心矩阵confidenceMatrix = 0(元素个数(nowCandidates),大小(日志,3));为pageIdx = 1:尺寸(日志,3)如果hopNumber ~ = 1 pastPitch =地板(日志(hopNumber-1 1 pageIdx))——emissionMatrixRange (1) + 1;其他的pastPitch =南;结束为candidateNumber = 1:元素个数(nowCandidates)如果hopNumber ~ = 1%获取当前的音高和转换为索引的范围currentPitch =地板(nowCandidates (candidateNumber))——emissionMatrixRange (1) + 1;confidenceMatrix (candidateNumber pageIdx) =…emissionMatrix (currentPitch pastPitch) *日志(hopNumber-1 2 pageIdx) * nowConfidence (candidateNumber);其他的confidenceMatrix (candidateNumber pageIdx) = 1;结束结束结束%分配估计为每一个路径为pageIdx = 1:尺寸(日志,3)%确定当前最自信的过渡从过去的[~,idx] = max (confidenceMatrix (:));%信心指数矩阵转换为音高和日志页面[chosenPitch, pastPitchIdx] = ind2sub([元素个数(nowCandidates),大小(日志,3)],idx);日志(hopNumber: pageIdx) =…[nowCandidates (chosenPitch),…confidenceMatrix (chosenPitch pastPitchIdx),…pastPitchIdx];%删除当前所选的音高的信心矩阵confidenceMatrix (chosenPitch:) =南;结束%正常化的信心日志(hopNumber 2:) =日志(hopNumber 2:) /笔(日志(hopNumber 2:));结束

numHops =大小(日志,1);keepLooking = true;指数= numHops + 1;而keepLooking指数=指数- 1;如果abs (max(日志(指数2:))分钟(日志(指数2:)))~ = 0 keepLooking = false;结束结束[~,bestPathIdx] = max(日志(指数2:));bestIndices = 0 (numHops, 1);bestIndices(指数)= bestPathIdx;为2 =指数:1:2 bestIndices (ii-1) =日志(2、3、bestIndices (ii));结束bestIndices (bestIndices = = 0) = 1;f0 = 0 (numHops, 1);为2 = (numHops): 1:2 f0 (ii) =日志(二1 bestIndices (ii));结束f0toPlot = f0;f0toPlot (~ isVoiced) =南;GPE =意味着(abs (f0toPlot (isVoiced)——truePitch (isVoiced)) > truePitch (isVoiced)。* p)。* 100;图(8)情节(t0, [truePitch f0toPlot])传说(“参考”,“估计”)ylabel (“F0 (Hz)”)包含(“时间(s)”)标题(" GPE = "+圆(GPE、2) +“(%)”)

f0 = movmedian (f0, 3);f0 (~ isVoiced) =南;GPE =意味着(abs (f0 (isVoiced)——truePitch (isVoiced)) > truePitch (isVoiced)。* p)。* 100;图(9)情节(t0, [truePitch, f0])传说(“参考”,“估计”)ylabel (“F0 (Hz)”)包含(“时间(s)”)标题(" GPE = "+圆(GPE、2) +“(%)”)

downloadFolder = matlab.internal.examples.download万博1manbetxSupportFile (“音频”,“ptdb-tug.zip”);dataFolder = tempdir;解压缩数据集(downloadFolder dataFolder) = fullfile (dataFolder,“ptdb-tug”);

广告= audioDatastore ([fullfile(数据集,“语音数据”,“女性”,“麦克风”),fullfile(数据集,“语音数据”,“男性”,“麦克风”)),…IncludeSubfolders = true,…FileExtensions =“wav”);wavFileNames = ads.Files;ads.Labels =取代(wavFileNames, (“麦克风”,“麦克风”,“wav”]、[“REF”,“ref”,“f0”]);idxToRemove =包含(ads.Files, (“F10”,“M10”]);ads1 =子集(广告,idxToRemove);ads9 =子集(广告,~ idxToRemove);

ads1 = shuffle (ads1);ads9 = shuffle (ads9);

emissionMatrixRange = [50400];emissionMatrix = [];为i = 1:元素个数(ads9.Files) x = getReferencePitch (ads9.Labels{我});emissionMatrix = HelperUpdateEmissionMatrix (x, emissionMatrixRange emissionMatrix);结束emissionMatrix = emissionMatrix + sqrt (eps);emissionMatrix = emissionMatrix. /规范(emissionMatrix);

noiseType = [“白色”,“氛围”,“引擎”,“喷气机”,“街道”];numNoiseToTest =元素个数(noiseType);desiredFs = 16 e3;whiteNoiseMaker = dsp.ColoredNoise(颜色=“白色”SamplesPerFrame = 40000, RandomStream =“与种子mt19937ar”,BoundedOutput = true);噪音{1}= whiteNoiseMaker ();[氛围,ambianceFs] = audioread (“Ambiance-16-44p1-mono-12secs.wav”);噪音{2}=重新取样(氛围、desiredFs ambianceFs);(引擎,engineFs) = audioread (“Engine-16-44p1-stereo-20sec.wav”);噪音{3}=重新取样(引擎、desiredFs engineFs);(飞机,jetFs) = audioread (“JetAirplane-16-11p025-mono-16secs.wav”);噪音{4}=重新取样(飞机,desiredFs jetFs);[街,streetFs] = audioread (“MainStreetOne-16-16-mono-12secs.wav”);噪音{5}=重新取样(街道、desiredFs streetFs);snrToTest = [10 5 0, -10);numSNRtoTest =元素个数(snrToTest);

比较=真正的;numFilesToTest = 20;p = 0.1;GPE_pitchTracker = 0 (numSNRtoTest numNoiseToTest numFilesToTest);如果比较GPE_swipe = GPE_pitchTracker;结束为i = 1: numFilesToTest [cleanSpeech信息]=阅读(ads1);cleanSpeech =重新取样(cleanSpeech desiredFs info.SampleRate);truePitch = getReferencePitch (info.Label {:});isVoiced = truePitch ~ = 0;truePitchInVoicedRegions = truePitch (isVoiced);为j = 1: numSNRtoTest为k = 1: numNoiseToTest noisySpeech = mixSNR (cleanSpeech,噪音{k}, snrToTest (j));f0 = HelperPitchTracker (noisySpeech、desiredFs EmissionMatrix = EmissionMatrix EmissionMatrixRange = EmissionMatrixRange);f0 = [0; f0];%手动对齐与数据库。GPE_pitchTracker (j, k,我)=意味着(abs (f0 (isVoiced) - truePitchInVoicedRegions) > truePitchInVoicedRegions。* p)。* 100;如果比较f0 = swipep (noisySpeech desiredFs, [50400], 0.01);f0 = f0(3:结束);%手动对齐与数据库。GPE_swipe (j, k,我)=意味着(abs (f0 (isVoiced) - truePitchInVoicedRegions) > truePitchInVoicedRegions。* p)。* 100;结束结束结束结束GPE_pitchTracker =意味着(GPE_pitchTracker, 3);如果比较GPE_swipe =意味着(GPE_swipe, 3);结束

为2 = 1:元素个数(噪音)图(9 + 2)情节(snrToTest, GPE_pitchTracker(:,(二),“b”)举行在如果比较图(snrToTest GPE_swipe(:,(二),“g”)结束情节(snrToTest GPE_pitchTracker(:,(二),“波”)如果比较图(snrToTest GPE_swipe(:,(二),“全球之声”)结束标题(noiseType (ii))包含(“信噪比”(dB))ylabel (“周节误差总值(p =”+圆(p, 2) +“)”)如果比较传奇(“HelperPitchTracker”,“刷”)其他的传奇(“HelperPitchTracker”)结束网格在持有从结束