0
$\begingroup$

I am working on pipelines for audio processing (audio features that will go into a neural net to create some embeddings).

We are using two features : Filter banks, or MFCC.

I was wondering what was the good approach with normalization :

Should we normalize those frames on each dimension (each frequency / mfcc coefficient), or normalize across all dimensions ?

This could lead to important differences in the relative importance of features .

$\endgroup$
0
$\begingroup$

Maybe a good starter would be to aggregate those features across your frames into one single feature vector, compute the mean and standard deviation and perform z-score normalization.

$\endgroup$
  • $\begingroup$ what do you mean by aggregate those features across frames into one single features ? $\endgroup$ – cdancette Mar 2 '18 at 9:21
  • $\begingroup$ By that I mean, I simply mean use your features across all the frames to calculate the mean and standard deviation to then perform z score normalization. Sorry if my earlier explanation was a little confusing. $\endgroup$ – Somesh Ganesh Mar 2 '18 at 15:03
  • $\begingroup$ So my question was exactly about this : Do we calculate mean and standard deviation across all dimension (features), or per dimension ? $\endgroup$ – cdancette Mar 2 '18 at 17:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.