Warped Filter Banks Used in Noisy Speech Recognition

The filter bank in the front-end of a speech recognition system mimics the function of the basilar membrane. It is believed that the closer the band subdivision is to human perception, the better the recognition results. This paper proposes the use of warped filter banks (WFBs) to replace traditional FIR filter banks and validates its use for the recognition of noisy speech. The WFBs bandwidths can be warped by using a first-order allpass transformation replacing the unit delay. Different warped factors in the allpass function can make the different scaled filter banks. Experiments carried on isolated words for speaker independent speech recognition show that the recognition rate with our proposed WFBs has been effectively increased.