Speaker verification—problems in small systems

Abstract

The problems encountered for speaker verification in small systems like laboratory PC-based experimental systems are: expensive implementation of complex algorithms and difficulty in obtaining satisfactory results using standard measures. For example, a system built by the authors following some successful works reported earlier failed to prove that the energy envelope is a good measure for verification. To cater to small-scale applications and to avoid uncertainities regarding the applicability of techniques used by others, a new approach is presented in this paper. The main thrust is to use a fixed set of phrases for verification and to use simple and easily computable measures for classification. For each speaker, training sessions are used to store master feature matrices, which are similar to templates. By appropriate weighting, vectors are derived from these matrices for easy comparisons with utterances during test sessions. The features chosen for the matrices are zero crossings, autocorrelation function, energy distribution, and pitch. In order to avoid complex time-warping and segmentation algorithms, the above features are evaluated for voiced sounds and over one or more pitch periods only. Also, the pitch period is determined by a new algorithm using symmetry check. The hardware needed and the highlights of the software and algorithms are presented in detail. Results for a small group of test speakers are also given. This new technique gives reasonably good performance compared to some of the standard techniques.