Test of an automatic syllable peak detector

Abstract

It is a venerable and commonplace intuition that syllables are “peaks of sonority.” Previous work by Mermelstein [J. Acoust. Soc. Am. 58, 880–883 (1975)] suggests that this intuition is nearly true on the acoustic surface of speech, and that relatively crude algorithms based on amplitude variation can work fairly well at counting syllables and locating their approximate extent. We have devised a simple speaker‐independent algorithm for detecting sonority peaks in continuous speech, and tested its accuracy as a syllable finder on several paragraphs of casual reading by four speakers. Performance was encouraging: less than 10% misses, and less than 1% false triggers. The misses were about evenly divided among (1) vowel‐vowel transition, where no sonority minimum is expected, (2) intervocalic sonorants with insufficient amplitude dip, and (3) devoiced or deleted schwas. Methods for dealing with these cases will be discussed.