火车生成对抗网络(GAN)合成声音

matFileName =“drumGeneratorWeights.mat”;loc = matlab.internal.examples.download金宝appSupportFile (“音频”,“GanAudioSynthesis /”+ matFileName);拷贝文件(loc pwd)

synthsound = synthesizePercussiveSound;fs = 16 e3;声音(synthsound fs)

t =(0:长度(synthsound) 1) / fs;情节(t, synthsound)网格在包含(“时间(s)”)标题(“合成敲击的声音”)

混响=反射镜(SampleRate = fs);parameterTuner(混响);

ts = timescope (SampleRate = fs,…TimeSpanSource =“财产”,…TimeSpanOverrunAction =“滚动”,…时间间隔= 10,…BufferLength = 10 * 256 * 64,…ShowGrid = true,…YLimits = [1]);

loopCount = 20;为2 = 1:loopCount synthsound = synthesizePercussiveSound;synthsound =混响(synthsound);ts (synthsound (: 1));soundsc (synthsound fs)暂停(0.5)结束

url1 =“https://zenodo.org/record/4687854/files/one_shot_percussive_sounds.zip”;url2 =“https://zenodo.org/record/4687854/files/licenses.txt”;downloadFolder = tempdir;percussivesoundsFolder = fullfile (downloadFolder,“one_shot_percussive_sounds”);licensefilename = fullfile (percussivesoundsFolder,“licenses.txt”);如果~ datasetExists percussivesoundsFolder disp (“下载Freesound一次性敲击的声音数据集(112.6 MB)……”)解压缩(url1 downloadFolder) websave (licensefilename url2);removeRestrictiveLicence (percussivesoundsFolder licensefilename)结束

广告= audioDatastore (percussivesoundsFolder IncludeSubfolders = true);

fftLength = 256;赢得=损害(fftLength,“周期”);overlapLength = 128;

如果canUseParallelPool池=质量;numPar = numpartitions(广告、池);其他的numPar = 1;结束

parfor2 = 1:numPar再分=分区(广告、numPar ii);应变= 0 (fftLength / 2 + 1128 1元素个数(subds.Files));为idx = 1:元素个数(subds.Files)%读取音频[x, xinfo] =阅读(再分);%进行预处理x = preprocessAudio(单(x), xinfo.SampleRate);% STFTS0 = stft (x窗口=赢,OverlapLength = OverlapLength FrequencyRange =“单向的”);%级S = abs (S0);应变(::,:,idx) = S;结束STrainC{2} =应变;结束

应变=猫(4,STrainC {:});

应变=日志(应变+ 1 e-6);

SMean =意味着(应变、[2 3 4]);SStd =性病(应变1 [2 3 4]);

应变= (STrain-SMean)。/ SStd;

应变=应变/ 3;Y =重塑(应变、元素个数(应变),1);Y (Y < 1) = 1;Y (Y > 1) = 1;应变=重塑(Y,大小(应变));

张力= (1:end-1,:,::);

应变=排列(应变、[2 1 3 4]);

maxEpochs = 1000;miniBatchSize = 64;

numIterationsPerEpoch =地板(大小(压力,4)/ miniBatchSize);

learnRateGenerator = 0.0002;learnRateDiscriminator = 0.0002;gradientDecayFactor = 0.5;squaredGradientDecayFactor = 0.999;

executionEnvironment =“汽车”;

generatorParameters = initializeGeneratorWeights;discriminatorParameters = initializeDiscriminatorWeights;

trailingAvgGenerator = [];trailingAvgSqGenerator = [];trailingAvgDiscriminator = [];trailingAvgSqDiscriminator = [];

doTraining =真正的;

saveCheckpoints =真正的;

numLatentInputs = 100;

迭代= 0;为时代= 1:maxEpochs%洗牌数据。idx = randperm(大小(压力,4));张力= (:,:,:,idx);%在mini-batches循环。为指数= 1:numIterationsPerEpoch迭代=迭代+ 1;% mini-batch读取的数据。dlX =应变(:,:,:,(索引1)* miniBatchSize + 1:指数* miniBatchSize);dlX = dlarray (dlX,“SSCB”);%为发电机网络生成潜在的输入。Z = 2 *(兰德(1,1,numLatentInputs miniBatchSize,“单身”)- 0.5);dlZ = dlarray (Z);%如果训练在GPU,然后将数据转换成gpuArray。如果(executionEnvironment = =“汽车”& & canUseGPU) | | executionEnvironment = =“图形”dlZ = gpuArray (dlZ);dlX = gpuArray (dlX);结束%对鉴频器使用dlfeval和梯度% modelDiscriminatorGradients helper函数。gradientsDiscriminator =…dlfeval (@modelDiscriminatorGradients discriminatorParameters、generatorParameters dlX, dlZ);%更新鉴别器网络参数。[discriminatorParameters, trailingAvgDiscriminator trailingAvgSqDiscriminator] =…adamupdate (discriminatorParameters gradientsDiscriminator,…trailingAvgDiscriminator trailingAvgSqDiscriminator,迭代,…learnRateDiscriminator、gradientDecayFactor squaredGradientDecayFactor);%为发电机网络生成潜在的输入。Z = 2 *(兰德(1,1,numLatentInputs miniBatchSize,“单身”)- 0.5);dlZ = dlarray (Z);%如果训练在GPU,然后将数据转换成gpuArray。如果(executionEnvironment = =“汽车”& & canUseGPU) | | executionEnvironment = =“图形”dlZ = gpuArray (dlZ);结束%对发电机使用dlfeval和梯度% | modelGeneratorGradients | helper函数。gradientsGenerator =…dlfeval (@modelGeneratorGradients discriminatorParameters generatorParameters, dlZ);%更新发电机网络参数。[generatorParameters, trailingAvgGenerator trailingAvgSqGenerator] =…adamupdate (generatorParameters gradientsGenerator,…trailingAvgGenerator trailingAvgSqGenerator,迭代,…learnRateGenerator、gradientDecayFactor squaredGradientDecayFactor);结束%每10时代,培训快照保存到垫文件。如果国防部(时代,10)= = 0 disp (“时代”+时代+“的”+ maxEpochs +“完成”。);如果saveCheckpoints%保存检查点训练中断。保存(“audiogancheckpoint.mat”,…“generatorParameters”,“discriminatorParameters”,…“trailingAvgDiscriminator”,“trailingAvgSqDiscriminator”,…“trailingAvgGenerator”,“trailingAvgSqGenerator”,“迭代”);结束结束结束

如果~ doTraining负载(matFileName“generatorParameters”,“SMean”,“SStd”);结束

dlZ = dlarray(2 *(兰德(1,1,numLatentInputs 1“单身”)- 0.5));

dlXGenerated = modelGenerator (dlZ generatorParameters);

S = dlXGenerated.extractdata;

=年代。”;

S =[年代;0 (1128)];

S = S * 3;S = (S . * SStd) + SMean;

S = exp(年代);

S = [S, S (end-1: 1:2,:));

S =[0(256100),年代,0 (256100)];

myAudio = stftmag2sig(年代,256,…FrequencyRange =“双侧”,…窗口=损害(256“周期”),…OverlapLength = 128,…MaxIterations = 20,…方法=“fgla”);myAudio = myAudio. / max (abs (myAudio), [],“所有”);myAudio = myAudio(128 * 100:结束- 128 * 100);

声音(收集(myAudio), fs)

t =(0:长度(myAudio) 1) / fs;情节(t, myAudio)网格在包含(“时间(s)”)标题(“合成氮化镓的声音”)

图stft (fs, myAudio窗口=损害(256年“周期”),OverlapLength = 128);

函数海底= modelGenerator (dlX参数)海底= fullyconnect (dlX、parameters.FC.Weights parameters.FC.Bias, Dataformat =“SSCB”);海底=重塑(海底,[1024 4 4大小(海底,2)]);海底=排列(海底,[3 2 1 4]);海底= relu(海底);海底= dltranspconv(海底,parameters.Conv1.Weights parameters.Conv1.Bias,跨步= 2,裁剪=“相同”DataFormat =“SSCB”);海底= relu(海底);海底= dltranspconv(海底,parameters.Conv2.Weights parameters.Conv2.Bias,跨步= 2,裁剪=“相同”DataFormat =“SSCB”);海底= relu(海底);海底= dltranspconv(海底,parameters.Conv3.Weights parameters.Conv3.Bias,跨步= 2,裁剪=“相同”DataFormat =“SSCB”);海底= relu(海底);海底= dltranspconv(海底,parameters.Conv4.Weights parameters.Conv4.Bias,跨步= 2,裁剪=“相同”DataFormat =“SSCB”);海底= relu(海底);海底= dltranspconv(海底,parameters.Conv5.Weights parameters.Conv5.Bias,跨步= 2,裁剪=“相同”DataFormat =“SSCB”);海底=双曲正切(海底);结束

函数海底= modelDiscriminator (dlX参数)海底= dlconv (dlX、parameters.Conv1.Weights parameters.Conv1.Bias,跨步= 2,填充=“相同”);海底= leakyrelu(海底,0.2);海底= dlconv(海底,parameters.Conv2.Weights parameters.Conv2.Bias,跨步= 2,填充=“相同”);海底= leakyrelu(海底,0.2);海底= dlconv(海底,parameters.Conv3.Weights parameters.Conv3.Bias,跨步= 2,填充=“相同”);海底= leakyrelu(海底,0.2);海底= dlconv(海底,parameters.Conv4.Weights parameters.Conv4.Bias,跨步= 2,填充=“相同”);海底= leakyrelu(海底,0.2);海底= dlconv(海底,parameters.Conv5.Weights parameters.Conv5.Bias,跨步= 2,填充=“相同”);海底= leakyrelu(海底,0.2);海底= stripdims(海底);海底=排列(海底,[3 2 1 4]);海底=重塑(海底,4 * 4 * 64 * 16,元素个数(海底)/ (4 * 4 * 64 * 16));重量= parameters.FC.Weights;偏见= parameters.FC.Bias;海底= fullyconnect(海底,重量、偏见,Dataformat =“CB”);结束

函数gradientsDiscriminator = modelDiscriminatorGradients (discriminatorParameters generatorParameters X, Z)%计算实际数据与鉴别器网络的预测。Y = modelDiscriminator (X, discriminatorParameters);%计算生成的预测数据与鉴别器网络。Xgen = modelGenerator (Z, generatorParameters);Ygen = modelDiscriminator (dlarray (Xgen“SSCB”),discriminatorParameters);甘%计算损失。lossDiscriminator = ganDiscriminatorLoss (Y, Ygen);%为每个网络,计算梯度的损失。gradientsDiscriminator = dlgradient (lossDiscriminator discriminatorParameters);结束

函数gradientsGenerator = modelGeneratorGradients (discriminatorParameters generatorParameters, Z)%计算生成的预测数据与鉴别器网络。Xgen = modelGenerator (Z, generatorParameters);Ygen = modelDiscriminator (dlarray (Xgen“SSCB”),discriminatorParameters);甘%计算损失lossGenerator = ganGeneratorLoss (Ygen);%为每个网络,计算梯度的损失。gradientsGenerator = dlgradient (lossGenerator generatorParameters);结束

函数lossDiscriminator = ganDiscriminatorLoss (dlYPred dlYPredGenerated)假= dlarray(0(1、大小(dlYPred, 2)));真正的= dlarray(的(1、大小(dlYPred 2)));D_loss =意味着(sigmoid_cross_entropy_with_logits (dlYPredGenerated、假));D_loss = D_loss +的意思(sigmoid_cross_entropy_with_logits (dlYPred,实际));lossDiscriminator = D_loss / 2;结束

函数真正lossGenerator = ganGeneratorLoss (dlYPredGenerated) = dlarray(的(1、大小(dlYPredGenerated, 2)));lossGenerator =意味着(sigmoid_cross_entropy_with_logits (dlYPredGenerated,实际));结束

函数discriminatorParameters = initializeDiscriminatorWeights filterSize = 5 [5];昏暗的= 64;% Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2) 1暗]);偏见= 0(1,- 1,昏暗,“单身”);discriminatorParameters.Conv1。重量= dlarray(重量);discriminatorParameters.Conv1。偏见= dlarray(偏差);% Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2)暗2 *暗]);偏见= 0(1 1 2 *昏暗,“单身”);discriminatorParameters.Conv2。重量= dlarray(重量);discriminatorParameters.Conv2。偏见= dlarray(偏差);% Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2) 2 *暗4 *暗]);偏见= 0(1、1、4 *昏暗,“单身”);discriminatorParameters.Conv3。重量= dlarray(重量);discriminatorParameters.Conv3。偏见= dlarray(偏差);% Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2) 4 *昏暗的8 *暗]);偏见= 0(1 1 8 *昏暗,“单身”);discriminatorParameters.Conv4。重量= dlarray(重量);discriminatorParameters.Conv4。偏见= dlarray(偏差);% Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2) 8 *暗16 *暗]);偏见= 0(1,1,16 *昏暗,“单身”);discriminatorParameters.Conv5。重量= dlarray(重量);discriminatorParameters.Conv5。偏见= dlarray(偏差);%完全连接重量= iGlorotInitialize([1, 4 * 4 *的* 16]);偏见= 0 (1,1,“单身”);discriminatorParameters.FC。重量= dlarray(重量);discriminatorParameters.FC。偏见= dlarray(偏差);结束

函数generatorParameters = initializeGeneratorWeights昏暗= 64;%的1重量= iGlorotInitialize((暗* 256100));偏见= 0(暗* 256 1“单身”);generatorParameters.FC。重量= dlarray(重量);generatorParameters.FC。偏见= dlarray(偏差);filterSize = 5 [5];%反式Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2) 8 *暗16 *暗]);偏见= 0(1,- 1,昏暗的* 8,“单身”);generatorParameters.Conv1。重量= dlarray(重量);generatorParameters.Conv1。偏见= dlarray(偏差);%反式Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2) 4 *昏暗的8 *暗]);偏见= 0(1,1,暗淡的* 4“单身”);generatorParameters.Conv2。重量= dlarray(重量);generatorParameters.Conv2。偏见= dlarray(偏差);%反式Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2) 2 *暗4 *暗]);偏见= 0(1,1,昏暗的* 2,“单身”);generatorParameters.Conv3。重量= dlarray(重量);generatorParameters.Conv3。偏见= dlarray(偏差);%反式Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2)暗2 *暗]);偏见= 0(1,- 1,昏暗,“单身”);generatorParameters.Conv4。重量= dlarray(重量);generatorParameters.Conv4。偏见= dlarray(偏差);%反式Conv2D重量= iGlorotInitialize ([filterSize (1) filterSize(2) 1暗]);偏见= 0 (1,1,1,“单身”);generatorParameters.Conv5。重量= dlarray(重量);generatorParameters.Conv5。偏见= dlarray(偏差);结束

函数y = synthesizePercussiveSound持续的pGeneratorParameters pMean pSTD如果isempty (pGeneratorParameters)%如果垫文件不存在,下载它文件名=“drumGeneratorWeights.mat”;负载(文件名,“SMean”,“SStd”,“generatorParameters”);pMean = SMean;pSTD = SStd;pGeneratorParameters = generatorParameters;结束%生成随机向量dlZ = dlarray (2 * (rand (1100 1“单身”)- 0.5));%生成谱图dlXGenerated = modelGenerator (dlZ pGeneratorParameters);%从dlarray转换为单身S = dlXGenerated.extractdata;=年代。”;%在消除边缘效应S =[年代;0 (1128)];%反向从训练步骤S = S * 3;S = (S . * pSTD) + pMean;S = exp(年代);%让它两面S =[年代;S (end-1: 1:2,)];%垫与零结束和开始S = [0 (256100) 0 (256100)];%使用快速Griffin-Lim算法重构信号。myAudio = stftmag2sig(年代,256,…FrequencyRange =“双侧”,…窗口=损害(256“周期”),…OverlapLength = 128,…MaxIterations = 20,…方法=“fgla”);myAudio = myAudio. / max (abs (myAudio), [],“所有”);y = myAudio(128 * 100:结束- 128 * 100);结束

函数= sigmoid_cross_entropy_with_logits (x, z) = max (x, 0) - x。* z +日志(1 + exp (abs (x)));结束函数w = iGlorotInitialize(深圳)如果元素个数(深圳)= = 2 numInputs =深圳(2);numOutputs =深圳(1);其他的numInputs = prod(深圳(1:3));numOutputs = prod(深圳([1 2 4]));结束乘数=√2 / (numInputs + numOutputs));w =乘数* sqrt(3) *(2 *兰特(深圳,“单身”)- 1);结束函数= preprocessAudio (fs),%确保mono=意味着(2);%重新取样16赫兹x =重新取样(16 e3, fs);%削减或垫有大约秒的长度加填充物,以确保% 128分析windows的STFT窗口256点和128点%重叠。y = trimOrPad (x, 16513);%正常化出= y / max (abs (y));结束函数y = trimOrPad (x, n)% trimOrPad修剪或垫音频%% y = trimOrPad (x, n)修剪或垫输入x n沿着第一个样本%的维度。如果x是修剪,修剪同样在正面和背面。%如果x是填充,填充同样与零正面和背面。%为奇数长的削减或填充,修剪多余的样品或垫%从后面。一个=大小(x, 1);如果< n frontPad =地板((一)/ 2);挤压垫= n - - frontPad;y = [0 (frontPad、大小(x, 2),如= x); x; 0(挤压垫、大小(x, 2),如= x));elseif> n frontTrim =地板((一)/ 2)+ 1;backTrim = a - n - frontTrim;y = x (frontTrim: end-backTrim:);其他的y = x;结束结束函数removeRestrictiveLicence (percussivesoundsFolder licensefilename)% removeRestrictiveLicense消除限制性的许可%解析id映射到许可的许可文件。创建一个表来保存信息。f = fileread (licensefilename);K = jsondecode (f);fn =字段(K);T =表(大小=[元素个数(fn), 4],…VariableTypes = [“字符串”,“字符串”,“字符串”,“字符串”),…VariableNames = [“ID”,“文件名”,“用户名”,“许可证”]);为2 = 1:元素个数(fn) fn =字符串(k . (fn {2}) . name);李=字符串(k . (fn {2}) .license);id = extractAfter(字符串(fn {2}),“x”);联合国=字符串(k . (fn {2}) .username);T (ii):) = {id、fn,李};结束%删除任何文件,禁止商业用途。发现里面的文件%合适的文件夹中,然后删除它。un金宝appsupportedLicense =“http://creativecommons.org/licenses/by-nc/3.0/”;fileToRemove = T.ID (strcmp (T.License 金宝appunsupportedLicense));为2 = 1:元素个数(fileToRemove) fileInfo = dir (fullfile (percussivesoundsFolder,“* *”fileToRemove (ii) +“wav”));删除(fullfile (fileInfo.folder fileInfo.name))结束结束