Benoit Verpeaux
Weighted Sub-band MFCCs for speech recognition
on the AURORA2 database
Abstract:In my work I try to develop a new technique for the calculation of the MFCC features.
The basic idea is that, very often, the energy of "real-world" noises is not equally
distributed along the frequency axis. The technique I present here is based on the
division of the signal spectrum into sub-bands and on the weighting of these sub-bands
before applying the IDCT. The evaluation on the AURORA2 database shows that the
use of this technique, by enhancing the useful parts of the signal and by attenuating the
noisy ones, results in an improvement of the recognition rate.