Pronunciation Modeling for Large Vocabulary Speech Recognition by Arthur

The large pronunciation variability of words in conversational speech is one of the major causes of low accuracy for automatic speech recognition (ASR). Many pronunciation modeling approaches have been developed to address this problem. Some explicitly manipulate the pronunciation dictionary as well as the set of the units used to define the pronunciations of words. Others model the pronunciation implicitly by using the long duration acoustical context to more accurately classify the spoken pronunciation unit. This thesis is a study of the relative ability of the acoustic and the pronunciation models to capture pronunciation variability in a nearly state of the art conversational telephone speech recognizer. Several methods are tested, each designed to improve modeling accuracy of the ASR. Some of the experiments result in a lower word error rate, but many do not, apparently because, in different ways, the accuracy gained by one part of the recognizer comes at the expense of accuracy lost or transferred from another part of the recognizer. Pronunciation variability is modeled with two approaches: from above with explicit pronunciation modeling and from below with implicit pronunciation modeling within the acoustic model. Both approaches make use of long duration context, explicitly by considering long-duration pronunciation units and implicitly by having the acoustic model consider long-duration speech segments. Some pronunciation models address the pronunciation variability problem by introducing multiple pronunciations per word to cover more variants observed in conversational speech. However, this can potentially increase the confusability between words. This thesis studies the relationship between pronunciation perplexity and the lexical ambiguity. The design of explicit pronunciation models presented here is informed by this study.