Comparison of VQ and DTW classifiers for speaker verification

Abstract

An investigation into the relative speaker verification performance of various types of vector quantisation (VQ) and dynamic time warping (DTW) classifiers is presented. The study covers a number of algorithmic issues involved in the above classifiers, and examines the effects of these on the verification accuracy. The experiments are based on the use of a subset from the Brent (telephone quality) speech database. This subset consists of repetitions of isolated digit utterances 1 to 9 and zero. The paper describes the experimental work, and presents an analysis of the results.