Understanding MFCC

I am trying to understand how the MFCC block works. If I understood correctly, after the Discrete Cosine Transformation, we have de cepstral coefficients. I have defined the number of coefficients in 13. After that, how does edge impulse obtain the characteistics to plot?

And, are these characteristics the input for the learning block, or learning block´s inputs are the spectrograms obtained after FFT?


Hi @Txomin,

If you need more information on how the MFCCs are calculated, I highly recommend this blog post: Practical Cryptography

If you take a look at the output coefficients plot here:

Each column represents a window of time, which is determined by the “Normalization window size” parameter. Each row in that column is the cepstral coefficient for that window of time.

The entire 2D array of MFCCs are used as inputs to the learning block.

Hope that helps!

1 Like

If I am not mistaken, the number of time windows is determined by the length of the frame and not by the normalization window size.

1 Like

Oops, yes–you are correct.