Getting NaN and Inf values after extracting features from Audio files

Hello,

I am trying to extract audio features from audio files that range from 1 - 30 seconds. When I extract the features using the code below, I get some NaN and Inf valuses which causing errors when I try to classify my files. Can someone help me to find out why this is happending? what causes this issue? or how to slove it? or any other thoughts.

I checked the files and there is nothing wrong with them."played the audio and was working fine"

Thank you everyone

trainingFeatures = cell(1,numel(adsTrain.Files));
windowLength = 512;
overlapLength = 0;

aFE = audioFeatureExtractor('SampleRate',16e3, ...
    'Window',hamming(windowLength,'periodic'),...
    'OverlapLength',overlapLength,...
    'spectralCentroid',true, ...
    'spectralCrest',true, ...
    'spectralDecrease',true, ...
    'spectralEntropy',true,...
    'spectralFlatness',true,...
    'spectralFlux',false,...                
    'spectralKurtosis',true,...
    'spectralRolloffPoint',true,...
    'spectralSkewness',true,...
    'spectralSlope',true,...
    'spectralSpread',true);

reset(adsTrain);
index = 1;

while hasdata(adsTrain)
    data = read(adsTrain);
    trainingFeatures{index} = extract(aFE,data); % After extracting the features, some of the valuse are NaN and Inf 
    index = index + 1;
end

[Ibrahim A]

Not sure what causes it but by applying fillmissing function, it should solve the problem by replacing the NaN values with valid values.

[Greg Heath]

You might also have to deal with Inf!

Hope this helps

Greg

Answer 1

[Brian Hemmat]

The features you are extracting (basically statistics about a spectrum) are either not defined or poorly defined for an all-zero input, which is mostly likely what you have at the beginning of your audio signal. Audio signals are sometimes padded with zeros to make they a consistent length. As you mention, you can replace the first couple feature vectors with some placeholder that won't error downstream for your system. Or, you can just disregard the feature vectors that correspond to all-zero input, since there is no relevant information there anyway.

Here is an illustration of what you are encountering:

>> windowLength = 512;
overlapLength = 0;
aFE = audioFeatureExtractor('SampleRate',16e3, ...
    'Window',hamming(windowLength,'periodic'),...
    'OverlapLength',overlapLength,...
    'spectralCentroid',true, ...
    'spectralCrest',true, ...
    'spectralDecrease',true, ...
    'spectralEntropy',true,...
    'spectralFlatness',true,...
    'spectralFlux',false,...                
    'spectralKurtosis',true,...
    'spectralRolloffPoint',true,...
    'spectralSkewness',true,...
    'spectralSlope',true,...
    'spectralSpread',true);

input = zeros(512,1);
features = extract(aFE,input)

features =

   NaN   NaN   NaN     0   Inf   NaN     0   NaN     0   NaN
   

The equations for the spectral descriptors can be found on the individual reference pages in the Audio Toolbox documentation, or summarized here:

https://www.mathworks.com/help/audio/ug/spectral-descriptors.html

If you look at the algorithms, you will see that many will result in a divide by zero (Inf) or a 0/0 (NaN).