Laura Mayfield Tomokiyo
School of Computer Science
Language Technologies Institute
Carnegie Mellon University
Low-proficiency non-native speakers represent a significant challenge for large-vocabulary continuous speech recognition (LVCSR). Acoustic models are confused by a heavy accent; language models are confused by poor grammar and unconventional word choice. Lack of comfort with the spoken language affects the fundamenal properties of connectded speech that have been a focus of L VCSR research; cross-word and interword coarticulation, disfluency, and prosody are among the features that differ in native and non-native speech. In this dissertation, I address the problems of characterizing, adapting to and detecting non-native speech, specifically for the population of Japanese-native speakers of English who have had less than one year's immersion in an English-language environment. Some extensions to other speaker groups are also explored.
Dissertation summary available here.
Full text available here.