Text-Independent Speaker Verification Using Artificially Generated GMMs for Cohorts

Abstract

This paper presents a text-independent speaker verification method using Gaussian mixture models (GMMs), where only utterances of enrolled speakers are required. Artificial cohorts are used instead of those from speaker databases, and GMMs for artificial cohorts are generated by changing model parameters of the GMM for a claimed speaker. Equal error rates by the proposed method are about 60% less than those by a conventional method which also uses only utterances of enrolled speakers.

Artificial Cohort Generation Based on Statistics of Real Cohorts for GMM-Based Speaker Verification

“…In this section, after a brief introduction of GMM, we review the previous method to generate GMMs for artificial cohorts [6]. A mixture of K Gaussian distributions is described as…”

“…The number of mixtures for GMM is 16, and the number of cohorts used are 50 for both real and artificial cohorts, both of which follow [6]. Given the likelihood for a claimed speaker p s (Y) and that for an other speaker among 99 speakers p i (Y), the speaker i is selected as a member of 50 cohorts for the claimed speaker if the difference of likelihood |p s (Y) − p i (Y)| is within the smallest 50 among 99.…”

“…In such an SV system, it is desirable to perform SV using only utterances of enrolled speakers. This issue was already addressed in text-dependent SV using hidden Markov model [4] and text-independent SV using GMM [5], [6].…”

“…The difference between the two models is the number of Gaussian mixtures, where the smaller number was used for the background model. In [6], artificial cohorts were used instead of real cohorts from speaker databases. Considering that GMMs for cohorts for a claimed speaker are relatively close to that for the claimed speaker, GMMs for cohorts were generated by changing model parameters of the GMM for the claimed speaker.…”