Effect of window size and shift period in mel-warped cepstral feature extraction on GMM-Based Speaker Verification

Abstract

This paper investigates how window size and shift period affect the performance of a speaker verification system. Specifically, we investigate their effects on verification accuracy and computation time of speaker verification systems built using the mel-warped cepstral feature extraction and Guassian Mixture Model. Experiments show that window size should not be larger than a critical point, which is determined by testing with a set of registered speakers. Otherwise, the computation time increases while the verification accuracy decreases.