A field study of performance improvements in HMM-based speaker verification

Abstract

This study reports our findings on speaker verification (SV) performance improvements using random 4-digit utterances collected over a single microphone type. The databases used in this study are the result of an ongoing field trial of SV access to automatic teller machines (ATMs) for secure unattended banking services. The SV system uses continuous density HMM models trained on 18 connected 4-digit utterances and has a baseline equal-error-rate (EER) of between 5.5 and 11% for different sets of data. Because of the limited training data, estimates for the mixture variances are most often poor. By calculating average mixture variances using all of the training data for a given speaker and then setting all of the model variances for that speaker to these speaker dependent values and using cohort normalization, the EER decreases consistently to between 2.5 and 6.5%.<>