Open-set speaker identification using adapted Gaussian mixture models

This paper presents an investigation into the use of adapted Gaussian mixture models in the context of open-set, text-inde-pendent speaker identification (OSTI-SI). The study includes a scheme for using the fast-scoring method which has been proposed for speaker verification. Furthermore, it provides an evaluation of various score normalisation methods in the proposed OSTI-SI framework. The dataset used for the experimental investigation is based on NIST SRE2003 1-speaker detection task. It is shown that significant improvements can be achieved if only a single mixture is used in the fast-scoring technique. Furthermore, it is experimentally observed that comparable performance is obtained using unconstrained cohort normalisation, T-norm and TZ-norm. The paper provides a detailed description of the experimental set up, and presents an analysis of the results obtained.