Text-independent speaker identification based on spectral weighting functions

Abstract

This paper introduces a novel approach to extract the speaker feature for textindependent speaker identification, the procedure to determine spectral weighting functions used in modeling the effect of the frequency selectivity and masking properties of human cochlea is systematically demonstrated. The modified LPC-derived cepstral coefficients based on spectral weighting functions are used in text-independent speaker identification to emphasize the individual information in the speech. The VQ technique is used as a classifier, experiment results obtained by the proposed approach in this paper are compared with those of LPC and PLP based method on a close set of 300 speakers, the results have shown that the proposed approach is robust against the intraspeaker and the interspeaker variations.