Full Covariance Matrices based Gaussian Mixture Models

From LRDE

The printable version is no longer supported and may have rendering errors. Please update your browser bookmarks and please use the default browser print function instead.

Abstract

Universal Background Models combined with Gaussian Mixture Models (UBM - GMM) is a common approach to speaker verifcation systems. In general, we use diagonal covariance matrices. This simplification allows us to have faster training steps during speaker recognition. We will explore the case of full-covariance along with the additional complexity and the benefits in terms of speaker recognition performances. All experiments will be performed on NIST-SRE 2010 datasets.