How can I extract mfcc features for audio and pass it to the cnn to train the model?

Question

How can I extract audio features using MFCC algorithm and use it with the Convolutional Neural Network to train the model?

I have extracted features of audio using MFCC and the file contained floating point columns but I am unable to distinguish between the columns?

for filename in os.listdir(directoryName):
    if filename.endswith('.wav'): # only get MFCCs from .wavs

        (rate,sig) = wav.read(directoryName + "/" +filename)


        mfcc_feat = mfcc(sig,rate)


        fbank_feat = logfbank(sig,rate)


        outputFile = resultsDirectory + "/" + os.path.splitext(filename)[0] + ".csv"
        file = open(outputFile, 'w+')
        numpy.savetxt(file, fbank_feat, delimiter=",")
        file.close() # close file

The values contained in the csv file like this.

7.01E+00    5.94E+00    5.28E+00    5.25E+00    5.24E+00
5.87E+00    3.53E+00    3.61E+00    2.32E+00    2.13E+00
5.68E+00    8.36E-01    1.75E-01    -8.48E-01   1.77E+00
7.96E+00    6.12E+00    5.47E+00    4.66E+00    4.34E+00
6.29E+00    4.34E+00    3.51E+00    3.15E+00    2.30E+00
6.37E+00    5.34E+00    4.76E+00    3.98E+00    3.77E+00
4.72E+00    1.62E+00    3.09E+00    1.66E+00    1.37E+00
6.14E+00    5.82E+00    5.12E+00    4.11E+00    3.76E+00
7.49E+00    3.79E+00    2.25E+00    5.03E+00    5.69E+00
5.89E+00    4.88E+00    5.88E+00    6.22E+00    6.19E+00

score 4 · Answer 1 · answered Sep 29 '19 at 16:45

The MFCC features of an audio signal is a time-series. If your input audio is 10 seconds at 44100 kHz and a 1024 samples hop-size (approx 23ms) for the MFCC, then you will get 430 frames, each with MFCC coefficients (maybe 20).

In order to classify this with a Convolutional Neural Network, you need to split it into fixed-size analysis windows of a practical size. For example a 43 MFCC frames window would correspond to approximately 1 second. Input to CNN is then of shape 43x20x1. If you want overlapping analysis windows (can improve performance, at cost of increased compute time) - then jump less than 43 frames ahead when computing the next window.

Here is an answer with example Python code. It is shown for mel-spectrogram, but can be adapted to MFCC by just replacing the call to librosa.feature.melspectrogram() with librosa.feature.mfcc().

How can I extract mfcc features for audio and pass it to the cnn to train the model?

1 Answers1