Analysis of type-II errors for VQ-distortion based speaker verification

Abstract

This paper investigates the distribution of interspeaker distances for the "training" speakers of the TIMIT speech database. The analysis is based on a speaker verification paradigm with 8 speakers serving as customers and the remaining speakers serving as impostors. The distance measure used is an average variance-weighted vector quantisation (VQ) distortion. It is found that the interspeaker distances correlate significantly with the differences of fundamental frequency (FO) between the speakers. Moreover, the shape of the distribution of impostor distances is largely determined by the customer's FO. A distinct asymmetry of VQ distances is observed between low-FO customers and high-FO impostors on the one hand and high-FO customers and low-FO impostors on the other. The type-II error function is estimated from a sample of impostors with similar FO to the customer. Dialect difference within TIMIT is not found to contribute significantly to VQ distance. Show more