The role of dynamic features in text dependent and -independent speaker verification

A segmental hidden Markov model (SHMM) is a hidden Markov model (HMM) whose states are associated with sequences of acoustic feature vectors (or segments), rather than individual vectors. By treating segments as homogeneous units it is possible, for example, to develop better models of speech dynamics. This paper considers the potential benefits of a trajectory-based segmental HMM for speaker recognition. Text-dependent speaker verification (TDSV) results obtained on YOHO and text-independent speaker verification (TI-SV) results on Switchboard are presented. The YOHO results show a 44% reduction in false acceptances using the segmental model compared with a conventional HMM, while the Switchboard results do not show any improvement relative to a conventional Gausian Mixture Model (GMM) system. Further experiments were conducted to explain these results. They indicate that the priority of a "segmental GMM" is to model stationary regions and shed light on the role of delta parameters in conventional TI-SV.