Auditory Filter Banks In Speech Recognition System

A major difficulty in speech recognition system is that the performance of system drops dramatically in noisy environment compared with the performance in clean environment. Human ear has strong capability in suppressing noise. How to make the recognition system behaves as selective as the human ear, in addition, to establish of the human auditory model which is in accordance with the characteristics of hearing system has been the research focus by the majority of researchers. Aiming at improving the robustness with respect to the noise, the auditory filter was applied to the Zero Crossing Peak Amplitude feature extraction to get the new features, and applied it to the speech recognition system in this thesis.This paper firstly introduced ZCPA feature extraction process in detail. This model uses the up zero-crossing interval to represent signal frequency information and amplitude to represent intensity information, and then combines the two together as the output of the feature.Moreover, the Gammatone filter was implemented in this paper, the filter is an ideal auditory filter which based on the characteristic of the basilar membrane. The16-channel Gammatone filter was used instead of FIR in extracting ZCPA to get GTZCPA feature. The experiments show that, the Gammatone filter didn’t achieve the expected results, which because the FIR filters were provided exactly by each channel.The Gammatone filter had the symmetrical frequency response, which did not fit for the human hearing property. In order to solve this problem, the Gammachirp filter was completed on the basis of Gammatone filter. The Gammachirp filter not only reflected the characteristics of the asymmetrical frequency response distribution of the basilar membrane which was as band-pass filter banks, but also implemented the level-dependent property. Using the16channels, frequency response asymmetrical distribution and level-independent Gammachirp filter banks for ZCPA extraction to obtain the AGCZCPA feature. The experiments show that, compared to the symmetric distribution frequency response Gammachirp filter banks, the recognition rates of the frequency response asymmetrical distribution Gammachirp filter banks was improved significantly. Then the16channels, frequency response asymmetrical distribution and level-dependent Gammachirp filter was used in ZCPA feature extraction, the result of GCZCPA feature were got. The experiments show that the system of level-dependent Gammachirp filter banks is more robust than of the level-independent Gammachirp filter banks.