A Cohort Methods for Score Normalization in Speaker Verification System, Acceleration of On-line Cohort Methods

This paper deals with several cohort methods for score normalization in speaker verification systems. At first, the reasons for score normalization are provided. Next, the principle of score normalization techniques based on Bayesian theorem are explained. The world, cohort, and unconstraint cohort normalization techniques are presented. A new normalization technique, unconstraint cohort extrapolated normalization, is introduced. Experiments on NIST 2002 corpus were performed in order to find which of the normalization methods give the best result. All experiments show that on-line cohort methods (especially unconstraint cohort extrapolated normalization) have outperformed the others. Finally, there is a discussion about time consumption of the on-line methods. An improvement for acceleration of these methods are proposed. The results of experiments on NIST 2002 data set showed same significant improvements of cohort and proposed cohort extrapolated normalization in comparison with the standard world normalization. The acceleration (4-times) on-line cohort methods is shown in another experiment.