Modeling Pronunciation Variation for Universal Access to Speech Understanding

点击翻译按钮获取中文摘要

In order for speech-based interfaces to be universally accessible - that is, available to all citizens in all situations - they must deal successfully with massive variability in pronunciation, as studies have shown that this is by far the single largest source of error in machine speech recognition. This project will address that issue by building dynamic pronunciation models which use context to predict the most likely pronunciations of a word in a particular utterance. The PIs' previous work has established that speakers are more likely to delete certain sounds (e.g., the final "t" in "about") in a variety of situations: if the next word starts with certain consonants; if they are speaking quickly; if the word is unsurprising (has a high trigram probability); if they are young and male; if the word is not followed by a pause or a word-repetition; or if they are speakers of certain dialects. They have further shown how confidence metrics can be used to identify inaccurate dictionary pronunciations based on actual pronunciations in speech. This project will extend and combine both of these lines of research, by using phonetic-confidence models to identify words with incorrect pronunciations that contribute to word-error, and then building statistical (decision-tree) models of pronunciation variation for these words. The PIs' model provides a general engine which they will apply to several specific classes of pronunciation variation, including dialect and speaker type, embedded in the Sphinx-II speech recognizer. The application of computational models to the study of pronunciation variation is an important new research area in linguistics and cognitive modeling, with implications for and potential benefits to many areas of science and pedagogy, including universal access.