Quick questions about mfcc and normalisation

Getting Started and Wayfinding
1

Hello,

I’mcurrently working on a patch where I analyse first 256 samples of slices so I can find the nearest neighbour with a kdtree when analysing audio input.

I used a dataset of 13 mfcc , means & skew , and 1 derivative.
It has a good success rate when testing with unseen material.

I wanted to add a few spectral shape descriptors and my guess was I should normalise the dataset I end up with ( mfcc + spectral shape )

first I observed the results of only my mfcc dataset normalised, and it was doing worse , could this be a mistake in my patch ? or could this be an usual effect when normalising mfcc ?

Cheers

I

1 Like
2

there is this other thread where we talk exactly about this. In short: normalising MFCCs can lead to much worse results, and we’re trying to find the latest literature to see how one can make it better…

If you feel geeky, there is a paper in that thread. I just had no time to dig and make it work… keep us posted if you find something in the meantime

2 Likes