Robust speaker recognition using glottal information-based cepstral mean subtraction

Abstract

Channel distortion and background noise often severely degrade the performance of automatic speaker recognition (ASR) system. In this paper, a new compensation method called glottal information-based cepstral mean subtraction (GIBCMS), which improves upon the conventional cepstral mean subtraction (CMS) method, is presented. Besides the cepstral information, GIBCMS has utilized the speaker’s glottal information, which also holds speaker-dependent characteristics, but is less vulnerable to environment than the cepstral one. In order to test its robustness under channel distortion, even with high level of background noise, we applied this method to the SRMC corpus which is added by noise at different SNRs. The experimental results show that GIBCMS gains better performance over other improved CMS methods on it.