A STUDY OF SIGNIFICANT PARAMETERS OF SPEECH FOB APPLICATION IN AUTOMATIC SPEECH RECOGNITION SYSTEMS

摘要

The program of research described here Is directed toward the development of a procedure for the automatic recognition of a large vocabulary of English words as spoken by-many talkers under a wide range of conditions of speaking. As part of the program, high-resolution spectrum analyses were obtained as a function of time for each of 7500 spoken words, automatically converted to digital form, and recorded on magnetic tape. The recorded spectra were then processed by a digital computer for which a program was written for the extraction of parameters of speech. The selection of the speech data, the system for obtaining and recording the speech spectra, and the computer program for reduction of the spectra are described. Results of statistical analyses of parameters extracted from 500 of the words are presented and discussed. The results obtained have provided insight Into some of the statistical characteristics of basic param¬eters of speech and into some of the influences of the manner of speaking on the values assumed by these parameters.