Frequency and wavelet filtering for robust speech recognition

Mel-frequency cepstral coefficients (MFCC) are the most widely used features in current speech recognition systems. However, they have a poor physical interpretation and they do not lie in the frequency domain. Frequency filtering (FF) is a technique that has been recently developed to design frequency-localized speech features that perform similar to MFCC in terms of recognition performances. Motivated by our desire to build time-frequency speech models, we wanted to use the FF technique as front-end. However, when evaluating FF on the Aurora-3 database we found some discrepancies in the highly mismatch case. This led us to put FF in another perspective: the wavelet transform. By doing so, we were able to explain the discrepancies and to achieve significant improvements in recognition in the highly mismatch case.