Incorporating Discriminating Observation Probabilities (DOP) into Semi-Continuous HMM for Speaker Verification

Abstract

This paper describes the use of a semi-continuous hidden Markov models for speaker verification. The system uses a technique for discriminative hidden Markov modelling known as discriminating observation probabilities (DOP). Results are presented for text-dependent experiments on isolated digits from 25 genuine speakers and 84 casual im-poster speakers, recorded over the public telephone network in the United Kingdom. Performance measures which are used to assess the DOP technique are equal error rate, zero false rejection rate, zero false acceptance rate and two measures of the distance between probability distributions for genuine and imposter speakers. The different performance measures are assessed with regard to their suitability for comparing speaker verification algorithms. This analysis further supports previous work which shows that the addition of DOP to an HMM system provides a significant advantage in speaker verification performance.