Machine Recognition of Speech

Abstract

This paper summarizes the status of a continuing research program concerned with the recognition of phonemes in both isolated and continuous speech. The acoustic analysis technique employs analog-threshold logic to abstract features that provide the basis for the recognition decisions. These feature-abstraction networks have been organized in a hierarchy of processing levels, ranging from the recognition of broad classes of sounds to decisions on a phoneme-by-phoneme level. As part of this program, a detailed comparison has been made of the characteristics of phonemes in continuous speech as contrasted with well-articulated discrete speech. The development of a continuous speech recognition capability makes possible the realization of several specific applications. Among these are spoken ZIP-code recognition systems that will improve the efficiency of parcel-sorting operation. In this application, it is necessary to recognize continuous strings of digits in a very high-noise background. A speech communications system is also being developed utilizing a speech recognition system at one end of a communications link, a narrow bandwidth channel transmitting the recognition results, and a speech synthesizer at the receiver.