I'm confused with the last step of extraction MFCC where we have to take DCT or IDCT.
Cepstrum is defined as IDFT of log(power spectrum). Since MFCC are cepstral coefficients so I think we have to take IDCT at the last step.
It's confirmed by several articles. For example this one (point 2.4 we see F^(-1)) and this one (scheme on the second page). In fact in first link formula below F(^-1) is DCT-2. Why? As wiki says DCT-3 is an inverse transform.
And if we look at implementations they take DCT-2 too. I checked two of them: pyAudioAnalysis (
stMFCC method) and implementation (
mfcc method) by the guy who writes for practicalcryptography.com.
And this is also confirmed by several links: here and here (second page).
Well.. who is right?