Our most successful Automatic Speech Recognition (ASR) systems remain too tied to specific domains (`air travel' or `business news') and to carefully pronounced, slow speech, because our statistical algorithms do not deal well with the extreme variation, in pronunciation, word-choice, and grammar, that characterizes natural, conversational, speech. The research goal of this CAREER project is to build better speech recognizers by taking human psychological results on lexical processing, giving them explicit probabilistic models, and integrating them into ASR algorithms. This project is incorporating part-of-speech and word frequency into probabilistic models of word pronunciation in ASR. This project is using distributional-vector models of word meaning in an ASR system to increase the probability of a word occurring near similar words, compiling probabilistic dictionaries of verb-argument structure to help ASR systems. The educational goal of this CAREER investigation is to incorporate computational, corpus-based, and probabilistic training into the undergraduate and graduate curriculum in phonetics, psycholinguistics, computational linguistics, and cognitive science. Through its research and educational facets, the CAREER project will help draw talented and knowledgeable linguists and psychologists into ASR and related fields, and produce research with a direct impact on ASR and its technical and commercial implications.
