انجمن کامپیوتر ایران

Threshold setting methods are one of the most important tasks in any speaker verification system. Equal Error Rate (EER) and Minimum Error Rate (MER) criteria are usually used for performance evaluation in speaker verification systems. It is important to know that EER and MER are two a posteriori methods for decision threshold estimation and can not be used in real world applications. In real world applications decision thresholds should be determined a priori, i.e. a validation phase is needed to conduct some inter-speaker and intra-speaker verification tests and to use obtained distances or likelihoods for estimation of a priori decision thresholds. In this paper, we present a least square error based method for calculation of parameters which we need to estimate decision thresholds. For this we use means and variances of inter-speaker and intra-speaker distances and estimate necessary parameters for calculating decision thresholds. The proposed method leads to an optimum a priori estimation of decision threshold values.