Quantifying surface phonetic variation using acoustic landmarks as feature cues

Abstract

Acoustic landmarks, which are abrupt spectral changes associated with certain feature sequences in spoken utterances, are highly informative and have been proposed as the initial analysis stage in human speech perception, and for automatic speech recognition (Stevens, JASA 111(4), 2002, 1872–1891). These feature cues and their parameter values also provide an effective tool for quantifying systematic context-governed surface phonetic variation (Shattuck-Hufnagel and Veilleux, ICPhS XVI, 2007, 925–928). However, few studies have provided landmark-based information about the full range of variation in continuous communicative speech. The current study examines landmark modification patterns in a corpus of maptask-elicited speech, hand annotated for whether the landmarks were realized as predicted from the word forms or modified in context. Preliminary analyses of a single conversation (400 s, one speaker) show that the majority of landmarks (about 84%) exhibited the canonical form predicted from their lexical specifications, and that modifications were distributed systematically across segment types. For example, 90% of vowel landmarks (at amplitude/F1 peaks) were realized as predicted, but only 70% of the closures for non-strident fricatives /v/ and /dh/, and 40–50% of /t/ closures and releases. Further quantification of landmark modification patterns will provide useful information about the processing of surface phonetic variation.