On the use of decoupled and adapted Gaussian mixture models for open-set speaker identification

Abstract

This paper presents a comparative analysis of the performance of decoupled and adapted Gaussian mixture models (GMMs) for open-set, text-independent speaker identification (OSTISI). The analysis is based on a set of experiments using an appropriate subset of the NIST-SRE 2003 database and various score normalisation methods. Based on the experimental results, it is concluded that the speaker identification performance is noticeably better with adapted-GMMs than with decoupled- GMMs. This difference in performance, however, appears to be of less significance in the second stage of OSTISI where the process involves classifying the test speakers as known or unknown speakers. In particular, when the score normalisation used in this stage is based on the unconstrained cohort approach, the two modelling techniques yield similar performance. The paper includes a detailed description of the experiments and discusses how the OSTI-SI performance is influenced by the characteristics of each of the two modelling techniques and the normalisation approaches adopted.