A Study on Standard and Iterative Map Adaptation for Speaker Recognition
Jason W. Pelecanos, Robert Vogt, Subramanian Sridharan · 2002
ABSTRACT: This paper presents a study comparing two formulations of the Maximum a Posteriori (MAP) adaptation algorithm using Gaussian mixture models for text-independent speaker recognition applications. Fully coupled target and background models are used for hypothesis testing in this speaker model based adaptation algorithm. We contrast the standard single iteration adaptation algorithm to adaptation using multiple iterations. Depending on the assumptions made, both solutions may be derived within the expectation-maximisation theoretical framework. The advantage of the iterative approach is that it accommodates for the Gaussian mixture component inter-relationships in performing density function estimation given the target speaker’s adaptation speech. We examine the characteristics of three different feature post-processing techniques which provide an understanding of the applicability of the standard and iterative approaches.