Speaker verification by computer using speech intensity for temporal registration

Abstract

A technique for automatic speaker verification is described in which voice pitch, low-frequency intensity, and the three lowest formant frequencies, all as functions of time, are the features used to represent an individual utterance. Verification consists of computing these features for a test utterance and comparing them with stored reference versions for the claimed identity. Before the test-versus-reference comparison is effected, the time dimension of the test utterance is warped to optimally register its intensity pattern onto the reference intensity pattern. Performance of the system is measured on a speaker population of moderate size. A variety of comparison formulas and various subsets of the five speech features are evaluated. The system responds either "accept" or "reject" to every utterance; "no decision" is not allowed. Automatic verification based solely upon voice pitch and intensity, both of which can be computed rapidly, yields average error rates below 1 percent.