A comparison of model estimation techniques for speaker verification

In this paper we address the problem of building speaker dependent hidden Markov models for a speaker verification system. A number of model building techniques are described and the comparative performance of a system using models built using each of these techniques is presented. Mean estimated models, models where the means of the HMMs are estimated using segmental K means but where the variances are taken from speaker independent models, out performed other techniques such as Baum-Welch re-estimation for training times of 120 s, 60 s and 15 s. Mean estimated models were also built with varying numbers of components in the state mixture distributions and a performance gain was again observed. The incorporation of transitional features into the system had degraded performance when the Baum-Welch algorithm was used for model estimation. However the inclusion of delta and delta-delta cepstra into the system using mean estimated models now gave a significant improvement in performance. Taken together these changes halved the equal error rate of the system from 15.7% to 7.8%