Intel® Integrated Performance Primitives
Deliberate problems developing high-performance vision, signal, security, and storage applications.

MFCC -> ippsEvalFBank


I am experimenting with IPP & Speech coding using the Aurora sample.
1. I have synthisized a 11KHz wav file with a clean 500Hz tone.
2. Using the aurora sample I extract the features for the 500Hz tone ( saved as a WAV file )

Evaluating the23 coeffcient output of ippsEvalFBank_32fI would expect the 5th or 6th filter tohave value dramatically higher than the rest of the filters, surprisingly, none of the filters was dramatically dominant than the other, moreover, the resulting coeeficients were growing in directrelation to the filter number, thus the 23rd filter had the highest value while the 1st filter had the lowest value, inspecting the filters generated by ippsMelFBankInitAlloc_32f,it was clear thatthe integral of the Mel filters is growing and is not normalaized.

A. Shouldn't the integral of the Mel filters be similar with all filters?
B. Having in mind the aurora sample provided with IPP, using the syntetically generated 500Hz file, why wasn't there a dramatically dominant filter for the ippsEvalFBank output ?

Any help would be appreciated

0 Kudos
0 Replies