Improved speaker verification through probabilistic subspace adaptation

Abstract

In this paper we propose a new adaptation technique for improved text-independent speaker verification with limited amounts of training data using Gaussian mixture models (GMMs). The technique, referred to as probabilistic subspace adaptation (PSA), employs a probabilistic subspace description of how a client?s parametric representation (i.e. GMM) is allowed to vary. Our technique is compared to traditional maximum a posteriori (MAP) adaptation, or relevance adaptation (RA), and maximum likelihood eigen-decomposition (MLED), or subspace adaptation (SA) techniques. Results are given on a subset of the XM2VTS databases for the task of textindependent speaker verification.