The Likelihood Ratio Decision Criterion for Nuisance Attribute Projection in GMM Speaker Verification

Abstract

We propose a way of integrating likelihood ratio (LR) decision criterion with nuisance attribute projection (NAP) for Gaussian mixture model- (GMM-) based speaker verification. The experiments on the core test of the NIST speaker recognition evaluation (SRE) 2005 data show that the performance of the proposed approach is comparable to that of the standard approach of NAP which uses support vector machines (SVMs) as a decision criterion. Furthermore, we demonstrate that the two criteria provide complementary information that can significantly improve the verification performance if a score-level fusion of both approaches is carried out.