comparison with librosa

[aashish-kumar]

Hi,

I was trying to compare your mfcc w.r.t librosa mfcc and I found lot of difference. Here are my arguments for librosa. I was able to match the other arguments to generate same feature maps but the values were very different and not just scaled versions. k = librosa.feature.mfcc(y,sr,window="hamming",center=False) Can you please help?

Thanks

[dspavankumar]

Hi Aashish,

MFCC implementations can be a lot different. Mine closely follows that of HTK; I am not aware of how exactly librosa implements it. Some general implementation differences are:

• using magnitude spectrum versus power spectrum
• differences in the filterbank cutoff frequencies
• differences in the cepstral lifter implementation (Kaldi's version is different from HTK's version)

So MFCCs given by two toolkits are usually different; it's mostly to do with the convention followed by the research lab that developed it. And there is no clear answer to which one is better across all the applications. Hope this helps.

Pavan.