ROBUSTNESS OF ACOUSTIC LANDMARKS IN SPONTANEOUSLY-SPOKEN AMERICAN ENGLISH

Stefanie Shattuck-Hufnagel1 & Nanette Veilleux2
1Speech Group, RLE, Massachusetts Institute of Technology; 2Speech Group, RLE, Massachusetts Institute of Technology, and Simmons College

ID 1584
[full paper]

Acoustic landmarks (abrupt changes associated with consonant closures and releases, vowels and glides) play an important role in some models of lexical access (e.g. Stevens 1998, 2002), so it is important to determine how often they survive articulatory overlap and weakening in spontaneous speech production. A corpus of spontaneous American English speech collected from 8 adult female speakers is hand labeled for the occurrence of landmarks. Preliminary results for one conversation (240 secs., 610 words, analysis completed for 1003 of 2750 predicted landmarks) show that 86% of landmarks were realized overall, with a sharply lower rate for coronal stops /t/ and /d/. These results suggest that the majority of landmarks are available for detection both by human listeners and automatic recognition algorithms. Ongoing analyses are comparing the rate of automatic detection of these acoustic events with the hand labels, and tabulating the contexts in which landmarks are lost or changed.