Beyond the phoneme: A juncture-accent model of spoken language

Abstract:

Current-generation automatic speech recognition systems represent words as sequences of phonemes. This `phonemic beads on a string' approach provides satisfactory performance for scripted material and well-defined-vocabulary tasks, but performs poorly when applied to more casual (and realistic) styles of speaking. For large-vocabulary unscripted tasks, sophisticated phoneme-based pronunciation models are capable of improving recognition performance to only a modest degree. Why should this be so? A singular problem with phoneme-based models is their inability to capture much of the pronunciation variation associated with informal speaking styles. Pronunciation models generally assume that each phonemic entity is acoustically realized in comparable fashion across the syllable, conditioned solely on the surrounding phonological context. However, the phonological diversity of syllable structure makes it difficult for polyphone models to completely capture contextual effects conditioned at this level of linguistic organization if pronunciation is largely structured at the syllabic level. A five-hour subset of the Switchboard corpus (brief telephone dialogues by native speakers of American English) was phonetically and prosodically annotated. The presentation will discuss the relation of stress accent and syllable structure to pronunciation variation as well as the significance of pronunciation variation for lexical representation using a detailed statistical analysis of a spontaneous speech corpus (Switchboard) as the empirical base of support.

About the speaker:

Steven Greenberg is a scientist based in California, who has worked at the International Computer Science Institute, the University of California, Berkeley, and the University of Wisconsin. He received his A.B. in Linguistics from the University of Pennsylvania, and a Ph.D. in Linguistics from the University of California, Los Angeles. His recent research has focused on machine-learning methods for automatic phonetic and prosodic annotation of spontaneous American English dialogue material as well as statistical characterization of spoken language. A detailed description of his work is available at: Steven Greenberg's homepage

Parking: Visitors requiring a parking pass are asked to contact us at least one working day before the seminar.

Enquiries: sals@ics.mq.edu.au