Comparison of Generative and Discriminative Approaches for Speaker Recognition with Limited Data

Abstract

This paper presents a comparison of three different speaker recognition methods deployed in a broadcast news processing system. We focus on how the generative and discriminative nature of these methods affects the speaker recognition framework and we also deal with intersession variability compensation techniques in more detail, which are of great interest in broadcast processing domain. Performed experiments are specific particularly for the very limited amount of data used for both speaker enrollment (typically ranging from 30 to 60 seconds) and recognition (typically ranging from 5 to 15 seconds). Our results show that the system based on Gaussian Mixture Models (GMMs) outperforms both systems based on Support Vector Machines (SVMs) but its drawback is higher computational cost.