Speaker recognition with mixtures of gaussians with sparse regression matrices

Abstract

When estimating a mixture of Gaussians there are usually two choices for the covariance type of each Gaussian component. Either diagonal or full covariance. Imposing a structure though may be restrictive and lead to degraded performance and/or increased computations. In this work, several criteria to estimate the structure of regression matrices of a mixture of Gaussians are introduced and evaluated. Most of the criteria attempt to estimate a discriminative structure, which is suited for classification tasks. Results are reported on the 1996 NIST speaker recognition task and performance is compared with structural EM, a well-known, non-discriminative, structure-finding algorithm.