Automatic Recognition of Spoken Digits

Abstract

No necessary-and-sufficient stimulus, or stimuli, has been discovered for the identification of any given word, yet people can reliably recognize speech under a wide variety of conditions One implication of this fact might be that speech recognition involves a complex data processing which takes advantage of linguistic redundancy. Such a system could make possible a reliable identification on the basis of a number of discriminating conditions, no one of which was in itself dependable. An attempt was made to realize such a data processing system in a program written for an experimental computer at Lincoln Laboratory. The program input consisted of the spoken digits processed through the Haskins Laboratories' resonance vocoder. Ninety-eight percent recognition was achieved on a sample of ten voices, mixed male and female. [The research reported in this talk was supported jointly by the U. S. Army, Navy, and Air Force under contract with the Massachusetts Institute of Technology.]