Adaptive management of speaker verification thresholds

Abstract

This paper describes an adaptive solution for threshold management of speaker verification systems. The reported algorithm estimates speaker-dependent thresholds based on successful verification sessions considering the minimization of a relation-based cost function. The optimum speaker verification threshold to authenticate a remotely located client is not achieved by only obtaining initial threshold estimations from enrollment data. Factors such as intra-speaker variability and the variability of communication channel impose dynamic online re-estimation of authentication thresholds. This paper describes a method that non-linearly maps speaker dependent scores to estimate a confidence measure for selecting reliable verification scores. The confidence measurement is combined with an adaptive algorithm to perform online re-estimation of speaker dependent thresholds. Different experiments were conducted on an ad-hoc database involving utterances from male and female speakers using cell and land-line phones. The reported algorithm provided a performance enhancement up to 36.2% when compared to traditional speaker independent or dependent threshold.