Discriminating semi-continuous HMM for speaker verification

Abstract

This paper describes the use of a multiple codebook SCHMM speaker verification system, which uses a novel technique for discriminative hidden Markov modelling known as discriminative observation probabilities (DOP). DOP can easily be added to a multiple codebook HMM system and require minimal additional computation and no additional training. The DOP technique can be applied to both speech and speaker recognition. Results are presented for text-dependent experiments on isolated digits from 27 true speakers and 84 casual imposters, recorded over the public telephone network in the United Kingdom. DOP are shown to significantly improve speaker verification performance for several commonly used parameter sets.