A method of estimating the equal error rate for automatic speaker verification

Abstract

In an automatic speaker verification (ASV) system, the equal error rate (EER) is a measure to evaluate the system performance. Usually it needs a large number of testing samples to calculate the EER. In order to estimate the EER without running the experiments using testing samples, a method of model-based EER estimation which computes likelihood scores directly from client speaker models and imposter models is proposed. However, the distribution of the computed likelihood scores is significantly biased against the distribution of likelihood scores obtained from testing samples. Here we propose a novel idea to manipulate the speaker models of the client speakers and the imposters so that the distribution of the computed likelihood scores is closer to the distribution of likelihood scores obtained from testing samples. Then a more reliable EER can be calculated by the speaker models. The experimental results show that the proposed method can properly estimate the EER.