Main Content


VGGish neural network

Since R2020b



    net = vggish returns a pretrained VGGish model.

    This function requires both Audio Toolbox™ and Deep Learning Toolbox™.


    collapse all

    Download and unzip the Audio Toolbox™ model for VGGish.

    Type vggish at the Command Window. If the Audio Toolbox model for VGGish is not installed, then the function provides a link to the location of the network weights. To download the model, click the link. Unzip the file to a location on the MATLAB path.

    Alternatively, execute these commands to download and unzip the VGGish model to your temporary directory.

    downloadFolder = fullfile(tempdir,'VGGishDownload');
    loc = websave(downloadFolder,'');
    VGGishLocation = tempdir;

    Check that the installation is successful by typing vggish at the Command Window. If the network is installed, then the function returns a SeriesNetwork (Deep Learning Toolbox) object.

    ans = 
      SeriesNetwork with properties:
             Layers: [24×1 nnet.cnn.layer.Layer]
         InputNames: {'InputBatch'}
        OutputNames: {'regressionoutput'}

    Load a pretrained VGGish convolutional neural network and examine the layers and classes.

    Use vggish to load the pretrained VGGish network. The output net is a SeriesNetwork (Deep Learning Toolbox) object.

    net = vggish
    net = 
      SeriesNetwork with properties:
             Layers: [24×1 nnet.cnn.layer.Layer]
         InputNames: {'InputBatch'}
        OutputNames: {'regressionoutput'}

    View the network architecture using the Layers property. The network has 24 layers. There are nine layers with learnable weights, of which six are convolutional layers and three are fully connected layers.

    ans = 
      24×1 Layer array with layers:
         1   'InputBatch'         Image Input         96×64×1 images
         2   'conv1'              Convolution         64 3×3×1 convolutions with stride [1  1] and padding 'same'
         3   'relu'               ReLU                ReLU
         4   'pool1'              Max Pooling         2×2 max pooling with stride [2  2] and padding 'same'
         5   'conv2'              Convolution         128 3×3×64 convolutions with stride [1  1] and padding 'same'
         6   'relu2'              ReLU                ReLU
         7   'pool2'              Max Pooling         2×2 max pooling with stride [2  2] and padding 'same'
         8   'conv3_1'            Convolution         256 3×3×128 convolutions with stride [1  1] and padding 'same'
         9   'relu3_1'            ReLU                ReLU
        10   'conv3_2'            Convolution         256 3×3×256 convolutions with stride [1  1] and padding 'same'
        11   'relu3_2'            ReLU                ReLU
        12   'pool3'              Max Pooling         2×2 max pooling with stride [2  2] and padding 'same'
        13   'conv4_1'            Convolution         512 3×3×256 convolutions with stride [1  1] and padding 'same'
        14   'relu4_1'            ReLU                ReLU
        15   'conv4_2'            Convolution         512 3×3×512 convolutions with stride [1  1] and padding 'same'
        16   'relu4_2'            ReLU                ReLU
        17   'pool4'              Max Pooling         2×2 max pooling with stride [2  2] and padding 'same'
        18   'fc1_1'              Fully Connected     4096 fully connected layer
        19   'relu5_1'            ReLU                ReLU
        20   'fc1_2'              Fully Connected     4096 fully connected layer
        21   'relu5_2'            ReLU                ReLU
        22   'fc2'                Fully Connected     128 fully connected layer
        23   'EmbeddingBatch'     ReLU                ReLU
        24   'regressionoutput'   Regression Output   mean-squared-error

    Use analyzeNetwork (Deep Learning Toolbox) to visually explore the network.


    Read in an audio signal to extract feature embeddings from it.

    [audioIn,fs] = audioread("Ambiance-16-44p1-mono-12secs.wav");

    Plot and listen to the audio signal.

    t = (0:numel(audioIn)-1)/fs;
    xlabel("Time (s)")
    axis tight

    % To play the sound, call soundsc(audioIn,fs)

    VGGish requires you to preprocess the audio signal to match the input format used to train the network. The preprocesssing steps include resampling the audio signal and computing an array of mel spectrograms. To learn more about mel spectrograms, see melSpectrogram. Use vggishPreprocess to preprocess the signal and extract the mel spectrograms to be passed to VGGish. Visualize one of these spectrograms chosen at random.

    spectrograms = vggishPreprocess(audioIn,fs);
    arbitrarySpect = spectrograms(:,:,1,randi(size(spectrograms,4)));
    xlabel("Mel Band")
    title("Mel Spectrogram for VGGish")
    axis tight

    Create a VGGish neural network. Using the vggish function requires installing the pretrained VGGish network. If the network is not installed, the function provides a link to download the pretrained model.

    net = vggish;

    Call predict with the network on the preprocessed mel spectrogram images to extract feature embeddings. The feature embeddings are returned as a numFrames-by-128 matrix, where numFrames is the number of individual spectrograms and 128 is the number of elements in each feature vector.

    features = predict(net,spectrograms);
    [numFrames,numFeatures] = size(features)
    numFrames = 24
    numFeatures = 128

    Visualize the VGGish feature embeddings.

    view([90 -90])
    title("VGGish Feature Embeddings")
    axis tight

    In this example, you transfer the learning in the VGGish regression model to an audio classification task.

    Download and unzip the environmental sound classification data set. This data set consists of recordings labeled as one of 10 different audio sound classes (ESC-10).

    downloadFolder = matlab.internal.examples.downloadSupportFile("audio","");
    dataLocation = fullfile(tempdir,"ESC-10");

    Create an audioDatastore object to manage the data and split it into train and validation sets. Call countEachLabel to display the distribution of sound classes and the number of unique labels.

    ads = audioDatastore(dataLocation,IncludeSubfolders=true,LabelSource="foldernames");
    labelTable = countEachLabel(ads)
    labelTable=10×2 table
            Label         Count
        ______________    _____
        chainsaw           40  
        clock_tick         40  
        crackling_fire     40  
        crying_baby        40  
        dog                40  
        helicopter         40  
        rain               40  
        rooster            38  
        sea_waves          40  
        sneezing           40  

    Determine the total number of classes.

    numClasses = height(labelTable);

    Call splitEachLabel to split the data set into train and validation sets. Inspect the distribution of labels in the training and validation sets.

    [adsTrain, adsValidation] = splitEachLabel(ads,0.8);
    ans=10×2 table
            Label         Count
        ______________    _____
        chainsaw           32  
        clock_tick         32  
        crackling_fire     32  
        crying_baby        32  
        dog                32  
        helicopter         32  
        rain               32  
        rooster            30  
        sea_waves          32  
        sneezing           32  
    ans=10×2 table
            Label         Count
        ______________    _____
        chainsaw            8  
        clock_tick          8  
        crackling_fire      8  
        crying_baby         8  
        dog                 8  
        helicopter          8  
        rain                8  
        rooster             8  
        sea_waves           8  
        sneezing            8  

    The VGGish network expects audio to be preprocessed into log mel spectrograms. Use vggishPreprocess to extract the spectrograms from the train set. There are multiple spectrograms for each audio signal. Replicate the labels so that they are in one-to-one correspondence with the spectrograms.

    overlapPercentage = 75;
    trainFeatures = [];
    trainLabels = [];
    while hasdata(adsTrain)
        [audioIn,fileInfo] = read(adsTrain);
        features = vggishPreprocess(audioIn,fileInfo.SampleRate,OverlapPercentage=overlapPercentage);
        numSpectrograms = size(features,4);
        trainFeatures = cat(4,trainFeatures,features);
        trainLabels = cat(2,trainLabels,repelem(fileInfo.Label,numSpectrograms));

    Extract spectrograms from the validation set and replicate the labels.

    validationFeatures = [];
    validationLabels = [];
    segmentsPerFile = zeros(numel(adsValidation.Files), 1);
    idx = 1;
    while hasdata(adsValidation)
        [audioIn,fileInfo] = read(adsValidation);
        features = vggishPreprocess(audioIn,fileInfo.SampleRate,OverlapPercentage=overlapPercentage);
        numSpectrograms = size(features,4);
        validationFeatures = cat(4,validationFeatures,features);
        validationLabels = cat(2,validationLabels,repelem(fileInfo.Label,numSpectrograms));
        segmentsPerFile(idx) = numSpectrograms;
        idx = idx + 1;

    Load the VGGish model and convert it to a layerGraph (Deep Learning Toolbox) object.

    net = vggish;
    lgraph = layerGraph(net.Layers);

    Use removeLayers (Deep Learning Toolbox) to remove the final regression output layer from the graph. After you remove the regression layer, the new final layer of the graph is a ReLU layer named 'EmbeddingBatch'.

    lgraph = removeLayers(lgraph,"regressionoutput");
    ans = 
      ReLULayer with properties:
        Name: 'EmbeddingBatch'

    Use addLayers (Deep Learning Toolbox) to add a fullyConnectedLayer (Deep Learning Toolbox), a softmaxLayer (Deep Learning Toolbox), and a classificationLayer (Deep Learning Toolbox) to the graph. Set the WeightLearnRateFactor and BiasLearnRateFactor of the new fully connected layer to 10 so that learning is faster in the new layer than in the transferred layers.

    lgraph = addLayers(lgraph,[ ...

    Use connectLayers (Deep Learning Toolbox) to append the fully connected, softmax, and classification layers to the layer graph.

    lgraph = connectLayers(lgraph,"EmbeddingBatch","FCFinal");

    To define training options, use trainingOptions (Deep Learning Toolbox).

    miniBatchSize = 128;
    options = trainingOptions("adam", ...
        MaxEpochs=5, ...
        MiniBatchSize=miniBatchSize, ...
        Shuffle="every-epoch", ...
        ValidationData={validationFeatures,validationLabels}, ...
        ValidationFrequency=50, ...
        LearnRateSchedule="piecewise", ...
        LearnRateDropFactor=0.5, ...
        LearnRateDropPeriod=2, ...
        OutputNetwork="best-validation-loss", ...
        Verbose=false, ...

    To train the network, use trainNetwork (Deep Learning Toolbox).

    [trainedNet, netInfo] = trainNetwork(trainFeatures,trainLabels,lgraph,options);

    Figure Training Progress (07-Jul-2022 14:34:35) contains 2 axes objects and another object of type uigridlayout. Axes object 1 contains 12 objects of type patch, text, line. Axes object 2 contains 12 objects of type patch, text, line.

    Each audio file was split into several segments to feed into the VGGish network. Combine the predictions for each file in the validation set using a majority-rule decision.

    validationPredictions = classify(trainedNet,validationFeatures);
    idx = 1;
    validationPredictionsPerFile = categorical;
    for ii = 1:numel(adsValidation.Files)
        validationPredictionsPerFile(ii,1) = mode(validationPredictions(idx:idx+segmentsPerFile(ii)-1));
        idx = idx + segmentsPerFile(ii);

    Use confusionchart (Deep Learning Toolbox) to evaluate the performance of the network on the validation set.

    figure(Units="normalized",Position=[0.2 0.2 0.5 0.5]);
    confusionchart(adsValidation.Labels,validationPredictionsPerFile, ...
        Title=sprintf("Confusion Matrix for Validation Data \nAccuracy = %0.2f %%",mean(validationPredictionsPerFile==adsValidation.Labels)*100), ...
        ColumnSummary="column-normalized", ...

    Output Arguments

    collapse all

    Pretrained VGGish neural network, returned as a SeriesNetwork (Deep Learning Toolbox) object.


    [1] Gemmeke, Jort F., Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. 2017. “Audio Set: An Ontology and Human-Labeled Dataset for Audio Events.” In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 776–80. New Orleans, LA: IEEE.

    [2] Hershey, Shawn, Sourish Chaudhuri, Daniel P. W. Ellis, Jort F. Gemmeke, Aren Jansen, R. Channing Moore, Manoj Plakal, et al. 2017. “CNN Architectures for Large-Scale Audio Classification.” In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 131–35. New Orleans, LA: IEEE.

    Extended Capabilities

    Version History

    Introduced in R2020b