“Total Voice”—Speaker verification goes deluxe!

Abstract

Practical use of speaker verification implies a preliminary task of user identification. The identification mode is selected to be a spoken sequence of six digits. Verification may then be performed on the identifying speech data. This “total voice” system must recognize connected digits independent of speaker with very high reliability. Three sequence constraints aid recognition: (1) two parity checks must be satisfied, (2) all digits must be different, and (3) “difficult” digit pairs are disallowed. The speech processing strategy features highly reliable time registration and accommodates multiple concurrent hypotheses at various processing levels. Operational performance appears feasible from preliminary results: sequence recognition-< 1% rejection, < 0.2% error; speaker verification-< 1% user rejection, < 2% impostor acceptance. This system has been simulated in real time on a minicomputer. [Research was sponsored by the Air Force Systems Command, Rome Air Development Center.]