Music Genre Classification Using Wavelet Time Scattering

This example shows how to classify the genre of a musical excerpt using wavelet time scattering and the audio datastore. In wavelet scattering, data is propagated through a series of wavelet transforms, nonlinearities, and averaging to produce low-variance representations of the data. These low-variance representations are then used as inputs to a classifier.

GTZAN Dataset

The data set used in this example is the GTZAN Genre Collection [7][8]. The data is provided as a zipped tar archive which is approximately 1.2 GB. The uncompressed data set requires about 3 GB of disk space. Extracting the compressed tar file from the link provided in the references creates a folder with ten subfolders. Each subfolder is named for the genre of music samples it contains. The genres are: blues, classical, country, disco, hiphop, jazz, metal, pop, reggae, and rock. There are 100 examples of each genre and each audio file consists of about 30 seconds of data sampled at 22050 Hz. In the original paper, the authors used a number of time-domain and frequency-domain features including mel-frequency cepstral (MFC) coefficients extracted from each music example and a Gaussian mixture model (GMM) classification to achieve an accuracy of 61 percent [7]. Subsequently, deep learning networks have been applied to this data. In most cases, these deep learning approaches consist of convolutional neural networks (CNN) with the MFC coefficients or spectrograms as the input to the deep CNN. These approaches have resulted in performance of around 84% [4]. An LSTM approach with spectrogram time slices resulted in 79% accuracy and time-domain and frequency-domain features coupled with an ensemble learning approach (AdaBoost) resulted in 82% accuracy on a test set [2][3]. Recently, a sparse representation machine learning approach achieved approximately 89% accuracy [6].

Download Data

The first step is to download the GTZAN Genre Collection [7][8]. The instructions in this example assume you download the data set to your temporary directory, tempdir, in MATLAB®. If you choose to download the data in a folder different from tempdir, change the directory name in the subsequent instructions. Use gunzip to download and uncompress the data set. Then use untar to extract the contents of the tar file. The folder genres is created in tempdir. Inside genres are ten subfolders, one for each music genre.

dataURL = 'http://opihi.cs.uvic.ca/sound/genres.tar.gz';
gunzip(dataURL,tempdir)                       % creates genres.tar in tempdir
untar(fullfile(tempdir,'genres.tar'),tempdir) % creates genres folder

Wavelet Scattering Framework

The only parameters to specify in a wavelet time scattering network are the duration of the time invariance, the number of wavelet filter banks, and the number of wavelets per octave. For most applications, cascading the data through two wavelet filter banks is sufficient. In this example, we use the default scattering network which uses two wavelet filter banks. The first filter bank has 8 wavelets per octave and the second filter bank has 1 wavelet per octave. For this example, set the invariant scale to be 0.5 seconds, which corresponds to slightly more than 11,000 samples for the given sampling rate. Create the wavelet time scattering decomposition network.

sn = waveletScattering('SignalLength',2^19,'SamplingFrequency',22050,...
    'InvarianceScale',0.5);

To understand the role of the invariance scale, obtain and plot the scaling filter in time along with the real and imaginary parts of the coarsest-scale wavelet from the first filter bank. Note that the time-support of the scaling filter is essentially 0.5 seconds as designed. Further, the time support of the coarsest-scale wavelet does not exceed the invariant scale of the wavelet scattering decomposition.

[fb,f,filterparams] = filterbank(sn);
phi = ifftshift(ifft(fb{1}.phift));
psiL1 = ifftshift(ifft(fb{2}.psift(:,end)));
dt = 1/22050;
time = -2^18*dt:dt:2^18*dt-dt;
scalplt = plot(time,phi,'linewidth',1.5);
hold on
grid on
ylimits = [-3e-4 3e-4];
ylim(ylimits);
plot([-0.25 -0.25],ylimits,'k--');
plot([0.25 0.25],ylimits,'k--');
xlim([-0.6 0.6]);
xlabel('Seconds'); ylabel('Amplitude');
wavplt = plot(time,[real(psiL1) imag(psiL1)]);
legend([scalplt wavplt(1) wavplt(2)],...
    {'Scaling Function','Wavelet-Real Part','Wavelet-Imaginary Part'});
title({'Scaling Function';'Coarsest-Scale Wavelet First Filter Bank'})
hold off

Audio Datastore

The audio datastore enables you to manage collections of audio data files. For machine or deep learning, the audio datastore not only manages the flow of audio data from files and folders, the audio datastore also manages the association of labels with the data and provides the ability to randomly partition your data into different sets for training, validation, and testing. In this example, use the audio datastore to manage the GTZAN music genre collection. Recall each subfolder of the collection is named for the genre it represents. Set the 'IncludeSubFolders' property to true to instruct the audio datastore to use subfolders and set the 'LabelSource' property to 'foldernames' to create data labels based on the subfolder names. This example assumes the top-level directory is inside your MATLAB tempdir directory and is called 'genres'. Ensure that location is the correct path to the top-level data folder on your machine. The top-level data folder on your machine should contain ten subfolders each named for the ten genres and must only contain audio files corresponding to those genres.

location = fullfile(tempdir,'genres');
ads = audioDatastore(location,'IncludeSubFolders',true,...
    'LabelSource','foldernames');

Run the following to obtain a count of the musical genres in the data set.

countEachLabel(ads)

ans=10×2 table
      Label      Count
    _________    _____

    blues         100 
    classical     100 
    country       100 
    disco         100 
    hiphop        100 
    jazz          100 
    metal         100 
    pop           100 
    reggae        100 
    rock          100 

As previously stated, there are 10 genres with 100 files each.

Training and Test Sets

Create training and test sets to develop and test our classifier. We use 80% of the data for training and hold out the remaining 20% for testing. The shuffle function of the audio datastore randomly shuffles the data. Do this prior to splitting the data by label to randomize the data. In this example, we set the random number generator seed for reproducibility. Use the audio datastore splitEachLabel function to perform the 80-20 split. splitEachLabel ensures that all classes are equally represented.

rng(100)
ads = shuffle(ads);
[adsTrain,adsTest] = splitEachLabel(ads,0.8);
countEachLabel(adsTrain)

ans=10×2 table
      Label      Count
    _________    _____

    blues         80  
    classical     80  
    country       80  
    disco         80  
    hiphop        80  
    jazz          80  
    metal         80  
    pop           80  
    reggae        80  
    rock          80  

countEachLabel(adsTest)

ans=10×2 table
      Label      Count
    _________    _____

    blues         20  
    classical     20  
    country       20  
    disco         20  
    hiphop        20  
    jazz          20  
    metal         20  
    pop           20  
    reggae        20  
    rock          20  

You see that there are 800 records in the training data as expected and 200 records in the test data. Additionally, there are 80 examples of each genre in the training set and 20 examples of each genre in the test set.

To obtain the scattering features, use a helper function, helperbatchscatfeatures, that obtains the natural logarithm of the scattering features for $$2^{19}$$ samples of each audio file and subsamples the number of scattering windows by 6. The source code for helperbatchscatfeatures is listed in the appendix. Wavelet scattering features are computed using a batch size of 64 signals.

If you have Parallel Computing Toolbox™ and a supported GPU, set useGPU to true in the following code and the scattering transform will be computed using the GPU. Using an NVIDIA Titan V GPU with a batch size of 64, the scattering features in this example were computed approximately 9 times faster than using the CPU.

N = 2^19;
batchsize = 64;
scTrain = [];
useGPU = false; % Set to true to use the GPU

while hasdata(adsTrain)
    sc = helperbatchscatfeatures(adsTrain,sn,N,batchsize,useGPU);
    scTrain = cat(3,scTrain,sc);
end

Record the number of time windows in the scattering transform for label creation.

numTimeWindows = size(scTrain,2);

In this example, there are 43 time windows, or frames, for each scattering path.

Repeat the same feature extraction process for the test data.

scTest = [];

while hasdata(adsTest)
   sc = helperbatchscatfeatures(adsTest,sn,N,batchsize,useGPU);
   scTest = cat(3,scTest,sc); 
end

Determine the number of paths in the scattering network and reshape the training and test features into 2-D matrices.

[~,npaths] = sn.paths();
Npaths = sum(npaths);
TrainFeatures = permute(scTrain,[2 3 1]);
TrainFeatures = reshape(TrainFeatures,[],Npaths,1);
TestFeatures = permute(scTest,[2 3 1]);
TestFeatures = reshape(TestFeatures,[],Npaths,1);

Each row of TrainFeatures and TestFeatures is one scattering time window across the 334 paths in the scattering transform of each audio signal. For each music sample, we have 43 such time windows. Accordingly, the feature matrix for the training data is 34400-by-334. The number of rows is equal to the number of training examples (800) multiplied by the number of scattering windows per example (43). Similarly, the scattering feature matrix for the test data is 8600-by-334. There are 200 test examples and 43 windows per example. Create a genre label for each of the 43 windows in the wavelet scattering feature matrix for the training data.

trainLabels = adsTrain.Labels;
numTrainSignals = numel(trainLabels);
trainLabels = repmat(trainLabels,1,numTimeWindows);
trainLabels = reshape(trainLabels',numTrainSignals*numTimeWindows,1);

Repeat the process for the test data.

testLabels = adsTest.Labels;
numTestSignals = numel(testLabels);
testLabels = repmat(testLabels,1,numTimeWindows);
testLabels = reshape(testLabels',numTestSignals*numTimeWindows,1);

In this example, use a multi-class support vector machine (SVM) classifier with a cubic polynomial kernel. Fit the SVM to the training data.

template = templateSVM(...
    'KernelFunction', 'polynomial', ...
    'PolynomialOrder', 3, ...
    'KernelScale', 'auto', ...
    'BoxConstraint', 1, ...
    'Standardize', true);
Classes = {'blues','classical','country','disco','hiphop','jazz',...
    'metal','pop','reggae','rock'};
classificationSVM = fitcecoc(...
    TrainFeatures, ...
    trainLabels, ...
    'Learners', template, ...
    'Coding', 'onevsone','ClassNames',categorical(Classes));

Test Set Prediction

Use the SVM model fit to the scattering transforms of the training data to predict the music genres for the test data. Recall there are 43 time windows for each signal in the scattering transform. Use a simple majority vote to predict the genre. The helper function helperMajorityVote obtains the mode of the genre labels over all 43 scattering windows. If there is no unique mode, helperMajorityVote returns a classification error indicated by 'NoUniqueMode'. This results in an extra column in the confusion matrix. The source code for helperMajorityVote is listed in the appendix.

predLabels = predict(classificationSVM,TestFeatures);
[TestVotes,TestCounts] = helperMajorityVote(predLabels,adsTest.Labels,categorical(Classes));
testAccuracy = sum(eq(TestVotes,adsTest.Labels))/numTestSignals*100

testAccuracy = 87.5000

The test accuracy, testAccuracy, is approximately 88 percent. This accuracy is comparable with the state of the art of the GTZAN dataset.

Display the confusion matrix to inspect the genre-by-genre accuracy rates. Recall there are 20 examples in each class.

confusionchart(TestVotes,adsTest.Labels)

The diagonal of the confusion matrix plot shows that the classification accuracies for the individual genres is quite good in general. Extract these genre accuracies and plot separately.

cm = confusionmat(TestVotes,adsTest.Labels);
cm(:,end) = [];
genreAccuracy = diag(cm)./20*100;
figure;
bar(genreAccuracy)
set(gca,'XTickLabels',Classes);
xtickangle(gca,30);
title('Percentage Correct by Genre - Test Set');

Summary

This example demonstrated the use of wavelet time scattering and the audio datastore in music genre classification. In this example, wavelet time scattering achieved an classification accuracy comparable to state of the art performance for the GTZAN dataset. As opposed to other approaches requiring the extraction of a number of time-domain and frequency-domain features, wavelet scattering only required the specification of a single parameter, the scale of the time invariant. The audio datastore enabled us to efficiently manage the transfer of a large dataset from disk into MATLAB and permitted us to randomize the data and accurately retain genre membership of the randomized data through the classification workflow.

References

1. Anden, J. and Mallat, S. 2014. Deep scattering spectrum. IEEE Transactions on Signal Processing, Vol. 62, 16, pp. 4114-4128.

2. Bergstra, J., Casagrande, N., Erhan, D., Eck, D., and Kegl, B. Aggregate features and AdaBoost for music classification. Machine Learning, Vol. 65, Issue 2-3, pp. 473-484.

3. Irvin, J., Chartock, E., and Hollander, N. 2016. Recurrent neural networks with attention for genre classification. https://www.semanticscholar.org/paper/Recurrent-Neural-Networks-with-Attention-for-Genre-Irvin/6da301817851f19107447e4c72e682e3f183ae8a

4. Li, T., Chan, A.B., and Chun, A. 2010. Automatic musical pattern feature extraction using convolutional neural network. International Conference Data Mining and Applications.

5. Mallat. S. 2012. Group invariant scattering. Communications on Pure and Applied Mathematics, Vol. 65, 10, pp. 1331-1398.

6. Panagakis, Y., Kotropoulos, C.L., and Arce, G.R. 2014. Music genre classification via joint sparse low-rank representation of audio features. IEEE Transactions on Audio, Speech, and Language Processing, 22, 12, pp. 1905-1917.

7. Tzanetakis, G. and Cook, P. 2002. Music genre classification of audio signals. IEEE Transactions on Speech and Audio Processing, Vol. 10, No. 5, pp. 293-302.

8. GTZAN Genre Collection. http://marsyas.info/downloads/datasets.html

Appendix -- Supporting Functions

helperMajorityVote -- This function returns the mode of the class labels predicted over a number of feature vectors. In wavelet time scattering, we obtain a class label for each time window. If no unique mode is found a label of 'NoUniqueMode' is returned to denote a classification error.

type helperMajorityVote

function [ClassVotes,ClassCounts] = helperMajorityVote(predLabels,origLabels,classes)
% This function is in support of wavelet scattering examples only. It may
% change or be removed in a future release.

% Make categorical arrays if the labels are not already categorical
predLabels = categorical(predLabels);
origLabels = categorical(origLabels);
% Expects both predLabels and origLabels to be categorical vectors
Npred = numel(predLabels);
Norig = numel(origLabels);
Nwin = Npred/Norig;
predLabels = reshape(predLabels,Nwin,Norig);
ClassCounts = countcats(predLabels);
[mxcount,idx] = max(ClassCounts);
ClassVotes = classes(idx);
% Check for any ties in the maximum values and ensure they are marked as
% error if the mode occurs more than once
tmpsum = sum(ClassCounts == mxcount);
ClassVotes(tmpsum > 1) = categorical({'NoUniqueMode'});
ClassVotes = ClassVotes(:);

helperbatchscatfeatures - This function returns the wavelet time scattering feature matrix for a given input signal. In this case, we use the natural logarithm of the wavelet scattering coefficients. The scattering feature matrix is computed on $$2^{19}$$ samples of a signal. The scattering features are subsampled by a factor of 6. If useGPU is set to true, the scattering transform is computed on the GPU.

function sc = helperbatchscatfeatures(ds,sn,N,batchsize,useGPU)
% This function is only intended to support examples in the Wavelet
% Toolbox. It may be changed or removed in a future release.

% Read batch of data from audio datastore
batch = helperReadBatch(ds,N,batchsize);
if useGPU
    batch = gpuArray(batch);
end

% Obtain scattering features
S = sn.featureMatrix(batch,'transform','log');
gather(batch);
S = gather(S);
    
% Subsample the features
sc = S(:,1:6:end,:);
end

helperReadBatch - This function reads batches of a specified size from a datastore and returns the output in single precision. Each column of the output is a separate signal from the datastore. The output may have fewer columns than the batchsize if the datastore does not have enough records.

function batchout = helperReadBatch(ds,N,batchsize)
% This function is only in support of Wavelet Toolbox examples. It may
% change or be removed in a future release.
%
% batchout = readReadBatch(ds,N,batchsize) where ds is the Datastore and
%   ds is the Datastore
%   batchsize is the batchsize

kk = 1;

while(hasdata(ds)) && kk <= batchsize
    tmpRead = read(ds);
    batchout(:,kk) = cast(tmpRead(1:N),'single'); %#ok<AGROW>
    kk = kk+1;
end

end

×

Open Example

You have a modified version of this example. Do you want to open this example with your edits?

No, overwrite the modified version Yes