Description of a Completely Automatic Speaker-Verification System

Abstract

This paper will discuss the results obtained to date with a real-time speaker-verification system being developed by ARL's Speech Systems Research Department. A speaker wishing to have his identity verified supplies his identity to a computer via a teletype. Reference data previously obtained for that speaker are read into the computer memory from a disk storage unit. The speaker then repeats the test phrase, “My code is _____,” into a microphone whose output is A/D converted and entered into the computer. A comparison is made between the test phrase and the reference data. On the basis of the comparison, the computer either verifies or challenges the purported identity of the speaker. The discussion will describe the “cepstral” measurements made on the test phrase, the nearest-neighbor decision procedure, and the problem of collecting adequate training samples. The system has been trained to recognize four speakers. When tested on new data for both the original speakers and 30 impostors, error rates of from 6% to 13% were obtained.