An Automatic Method for Computer Recognition of Continuous Speech

Abstract

An automatic speech recognition system for syntactically acceptable units of continuous speech has been implemented on a digital computer. The input speech to the system is carefully spoken by a single cooperative male speaker. A primary segmentation is attempted by delineating the boundaries of fricatives, stops, and nasals. The remaining vocalic sections are subsegmented on the basis of short-time phoneme identification. The principle identification procedure uses cepstrum matching; however, auxilliary use is made of zero crossing and slope change rates for fricatives, and apparent place of articulation for stops. Three phoneme choices are assigned to each segment. In order to recognize the utterance, successive phoneme strings are generated from these choices until one is found which satisfies the lexical and syntactic constraints. The lexical constraint requires that the phoneme string consist only of “phonemicon” (phonetic dictionary) entries and the syntactic constraint requires that the word string satisfy a simplified English syntax. A relatively simple syntax checking algorithm has been devised for this purpose, which, although it admits some non-sense structures, imposes a severe constraint on allowed word strings.