ASA 129th Meeting - Washington, DC - 1995 May 30 .. Jun 06

1aSC37. Toward automatic segmentation and syllabification of dysarthric speech.

Guangsheng Zhang

H. Timothy Bunnell

Univ. of Delaware, Dept. of Linguist., and Speech Process. Lab, A. I. duPont Inst., 1600 Rockland Rd., Wilmington, DE 19899

A program has been developed which uses heuristic processing of the output of a neural network to locate syllable boundaries in connected speech. The neural network attempts a broad segmental (8 category) classification. This is followed by a program which locates syllable boundaries using phonological constraints and the pattern of network classifications over time. On test data consisting of 150 sentences (about 1800 syllables) from the TIMIT database, this program correctly identifies between 83% and 88% of the syllable boundaries (depending on the error criteria), including syllables bounded by sonorant consonants and syllables containing syllabic consonants. This program is now being applied to the identification of syllable boundaries in speech from talkers with cerebral palsy. In preliminary analyses of 20 sentences from ten dysarthric talkers (140 syllables) only 12 boundary omissions occurred. However, since dysarthric speech often contains hesitations, false starts, and extraneous noises, assessing boundary insertion errors is more difficult; about 50 boundary insertions occurred. Roughly 2/3 of these represented instances in which the talker inserted additional syllabic features into the speech. Further studies will examine methods for reducing the number of these ``unintended'' boundaries identified by the algorithm. [Work supported by the Nemours Foundation.]