A fundamental challenge for current research on speech science and technology is understanding and modeling individual variation in spoken language. Individuals have their own speaking styles, depending on many factors, such as their dialect and accent as well as their socioeconomic background. These individual differences typically introduce modeling difficulties for large-scale speaker-independent systems designed to process input from any variant of a given language. This dissertation focuses on automatically identifying the dialect or accent of a speaker given a sample of their speech, and demonstrates how such a technology can be employed to improve Automatic Speech Recognition (ASR). In this thesis, we describe a variety of approaches that make use of multiple streams of information in the acoustic signal to build a system that recognizes the regional dialect and accent of a speaker. In particular, we examine frame-based acoustic, phonetic, and phonotactic features, as well as high-level prosodic features, comparing generative and discriminative modeling techniques. We first analyze the effectiveness of approaches to language identification that have been successfully employed by that community, applying them here to dialect identification. We next show how we can improve upon these techniques. Finally, we introduce several novel modeling approaches -- Discriminative Phonotactics and kernel-based methods. We test our best performing approach on four broad Arabic dialects, ten Arabic sub-dialects, American English vs. Indian English accents, American English Southern vs. Non-Southern, American dialects at the state level plus Canada, and three Portuguese dialects. Our experiments demonstrate that our novel approach, which relies on the hypothesis that certain phones are realized differently across dialects, achieves new state-of-the-art performance on most dialect recognition tasks. This approach achieves an Equal Error Rate (EER) of 4% for four broad Arabic dialects, an EER of 6.3% for American vs. Indian English accents, 14.6% for American English Southern vs. Non-Southern dialects, and 7.9% for three Portuguese dialects. Our framework can also be used to automatically extract linguistic knowledge, specifically the context-dependent phonetic cues that may distinguish one dialect form another. We illustrate the efficacy of our approach by demonstrating the correlation of our results with geographical proximity of the various dialects. As a final measure of the utility of our studies, we also show that, it is possible to improve ASR. Employing our dialect identification system prior to ASR to identify the Levantine Arabic dialect in mixed speech of a variety of dialects allows us to optimize the engine's language model and use Levantine-specific acoustic models where appropriate. This procedure improves the Word Error Rate (WER) for Levantine by 4.6% absolute; 9.3% relative. In addition, we demonstrate in this thesis that, using a linguistically-motivated pronunciation modeling approach, we can improve the WER of a state-of-the art ASR system by 2.2% absolute and 11.5% relative WER on Modern Standard Arabic.
展开▼
