Automatic Estimation of a Priori Speaker Dependent Thresholds in Speaker Verification

Abstract

The selection of a suitable threshold is considered essential for the correct performance of automatic enrollment in speaker verification. Conventional methods have faced with the scarcity of data and the problem of an a priori decision, using biased client scores, impostor data, variances, a speaker independent threshold or some combination of them. Because of this lack of data, means and variances are estimated in most cases with very few scores. Noise or simply poor quality utterances, when comparing to the client model, can lead to some scores which produce a high variance in estimations. These scores are outliers and have an effect on the right estimation of mean and specially standard deviation. We propose here an algorithm to discard outliers. The method consists of iteratively selecting the most distant score with respect to mean. If this score goes beyond a certain threshold, the score is removed and mean and standard deviation estimations are recalculated. When there are only a few utterances to estimate mean and variance, this method leads to a great improvement. Text dependent and text independent experiments have been carried out by using a telephonic multisession database in Spanish with 184 speakers, that has been recently recorded by the authors.