New techniques for automatic speaker verification

An interactive automatic speaker verification system has been augmented to include linear prediction parameters in addition to the already existing pitch and intensity analysis of sentence-long utterances. This improved system has been evaluated on a new and enlarged speaker population. A method for selecting optimum speaker-dependent features has been incorporated in this system which significantly improves its performance. The evaluation indicates that verification error rate is approximately 1 percent with respect to casual impostors and 4 percent with respect to well-trained mimics.