Subband approach for automatic speaker recognition: Optimal division of the frequency domain

Abstract

This paper presents a new method for automatic speaker recognition. The principle is to split the whole spectral domain into partial frequency subbands on which recognizers are independently applied and then recombined to yield a global recognition decision. In this article, we particularly discuss the selection of the most critical subbands for the speaker recognition task and the choice of an optimal division of the frequency domain. Speaker recognition experiments are conducted on different subbands for a 630 population on TIMIT using second-order statistical methods. Large differences in identification between subbands are observed. In particular, the low-frequency subbands (under 600Hz) and the high-frequency subbands (over 2000Hz) are more speaker specific than middle-frequency ones. An appropriate selection of the most critical subbands shows that very good performances are still obtained with only half of the frequency domain. Finally experiments on different subband system architectures show that the correlations between frequency channels are of prime importance for the speaker recognition task. Some of these correlations are lost when the frequency domain is divided into subbands. Consequently efficient recombination procedures need to be investigated to perform enhanced speaker identification results.