Speaker Model and Decision Threshold Updating in Speaker Verification

Abstract

This paper deals with the problem of updating reference models and decision thresholds in speaker verification. In the real application of a speaker verification system, reference models of speakers and their decision thresholds should be already determined. Reference model and decision threshold updating can render a speaker verification system more robust against inter-session variability, due to changes in speaker’s voice and the diversity in transmission line characteristics. In this paper, different methods for the calculation of decision threshold, the updating of decision threshold and reference model, and BMDR decision criterion are compared and evaluated. In our experiments, Sphericity Measure (a Second Order Statistical Measure) was implemented and used for speaker modeling.