A robust sequential test for text-independent speaker verification

A robust speaker verification algorithm based on sequential hypothesis testing is presented. In speaker verification, the system performance is severely degraded by deviations from the nominal statistical speaker models caused by insufficient training data, varying microphone and transmission line characteristics, and different levels and types of background noise. Historically, this problem has been addressed by using an empirical method to determine an adequate decision threshold for a desired operating point. The robust detection algorithm presented here is based on a minimax criterion, such that the worst-case performance over a class of distributions is minimized. The sequential test provides further robustness by using additional data if a decision cannot be reached with the desired level of confidence. The sequential detector has been implemented and tested on a realistic database collected specifically for this purpose, and the performance has been shown to be comparable to or better than that of the corresponding heuristic detectors previously described in the literature.