Speaker verification using passwords

Abstract

The use of passwords embedded in free text for speaker verification increases system integrity through privileged knowledge (of the passwords). Two problems arise, however: the stability of the password speech data suffers from changing context, and the password must be detected reliably with few false alarms. Sensitivity to context is minimized by choosing passwords to be content words that begin and end with “hard” consonants. Password detection is based upon finding key reference points within each password through speaker-specific data comparison. An evaluation was performed using ten 3-min sessions of read text from each of eight subjects. A total of 160 tokens of each of five passwords was processed. The results, using filter bank spectra, are word false alarms/hour SV error, % (at 10% miss) (TS = IMP) congress 1.5 6.7 candidate 2.5 9.1 pentagon 0.5 8.8 constitution 0 3.5 secretary of state 0 4.4 [Research was sponsored by the Air Force Systems Command, Rome Air Development Center.]