Accession Number : AD0638380

Title :   A STUDY OF CRITICAL-INSTANT SAMPLING OF SPEECH PARAMETERS FOR AUTOMATIC RECOGNITION OF SPOKEN WORDS.

Descriptive Note : Final rept., Jul 64-Aug 65.

Corporate Author : FEDERAL SCIENTIFIC CORP NEW YORK

Personal Author(s) : Weiss,Mark

Report Date : JUL 1966

Pagination or Media Count : 146

Abstract : A study has been made to advance the development of a general and practical procedure for automatic recognition of spoken words. Two major problems were studied: variations in parameter values of speech due to variations in speech characteristics of different talkers, and due to variations in the duration of spoken words. Included in this study was an analysis of the means and standard deviations of formant frequencies made for a group of twenth talkers, each saying a list of 30 words two times. A range of standard deviations of from 5 percent of the mean formant frequency to 50 percent was observed. For the first time, statistical information has been obtained on the spectra of fricative sounds. One of the serious problems inherent in most techniques of automatic recognition of speech is that of time normalization, since words or parts of words normally are spoken with different durations. A new technique has been investigated which avoids many of the problems of time normalization. Word patterns which are independent of word duration were generated using critical-instant sampling of the speech data. Recognition tests using these patterns were made to evaluate the relative significance of different critical instants. The results obtained indicate that critical-instant sampling provides effective and efficient patterns on which a general spoken-word recognition procedure may be based. (Author)

Descriptors :   (*SPEECH RECOGNITION, *STATISTICAL ANALYSIS), AUTOMATIC, PATTERN RECOGNITION, LEARNING, DECISION MAKING, SOUND PITCH, ACOUSTIC SIGNALS, COMPUTERS

Subject Categories : Voice Communications

Distribution Statement : APPROVED FOR PUBLIC RELEASE