Relative Effectiveness of Score Normalisation Methods in Open-Set Speaker Identification

This paper presents an investigation into the relative effectiveness of various well-known score normalisation methods in the context of open-set, text-independent speaker identification. The scope of the study includes a thorough experimental analysis of the performance of the methods considered. The experimental investigations are based on the use of the dataset proposed for the 1-speaker detection task of the NIST Speaker Recognition Evaluation 2003. The results clearly demonstrate that significant benefits can be achieved by using score normalisation in open-set identification, and that the level of this depends highly on the type of the approach adopted. Based on the experimental results, it is found that amongst the various normalisation methods considered, those which are based on the Bayesian solution provide the best performance. In particular, the unconstrained cohort method with a small cohort size appears to outperform all other approaches. The paper provides a detailed description of the experimental set up, and presents an analysis of the results obtained.