Performance analysis of a matcher in a lexical access system based on landmarks and distinctive features

Abstract

Performance characteristics of a prototype matcher in a lexical access system based on landmarks and distinctive features are analyzed. A database of 16 CONV files containing spontaneous American English utterances produced by 8 female speakers is annotated with words, and phone sequences derived from the word sequences are generated using the CMU phone-based dictionary. Predicted landmark and distinctive feature sequences are then generated using context-dependent rules from the phone sequences. These labels are used to map back to a lexicon which is also represented in terms of landmarks and distinctive features. The results for using core lexicons consisting of words within a CONV file show an average match rate of about 23% using only manner-class-related landmarks, and about 93% using the distinctive feature labels. Using an expanded lexicon combining all core lexicons lowers average match rates, by about 7% using landmark labels, and by 4% using the distinctive feature labels. These results provide characteristic rates for using linguistically motivated features to match to a lexicon, for both the landmark labels and for the more detailed distinctive feature labels.