New face representation using compressive sensing

Abstract

GMM supervectors are among the most popular feature sets used in SVM-based text-independent speaker verification. Most of the studies represent speaker characteristics obtained from a long recording with a single supervector in the SVM space. Working on the NIST SRE’10 dataset, this study compares the effect of two sampling methods to increase the number of supervectors, on the verification performance. Dominance of positive and negative classes on model construction is investigated.