On the Applications of Distinctive Features to Speech Recognition

Abstract

The value of a distinctive-feature description of the speech process at the level of linguistic studies involving grammar, syntax, morphology, phonology, etc., has already been demonstrated. The value of the distinctive-feature approach to automatic-speech recognition, compared to other descriptions of the speech process, has heretofore been only partially examined. A computer-recognition scheme, based on the goal of tracking the distinctive features of speech, is presented. The program provides a complete vowel-recognition logic, which compares favorably with the behavior of a panel of listeners when both are presented with the speech of a single speaker. Methods are presented by which single vowel, diphthong, and possible triphthong recognition is produced from the speech spectral data. Details are given on the recognition of the short or lax vowels for which formant transitions and variations are often extreme. Also discussed is the possible effect of speech tempo on a proposed tense/lax boundary and its implications. Preliminary evidence suggests that Fant's neutral vowel may take on fundamental phonetic as well as articulatory significance. [Research reported in this paper sponsored by the U. S. Air Force Cambridge Research Laboratories, Office of Aerospace Research, under contract AF 19(628)305.]