Non-Native Pronunciation Variation Modeling for Automatic Speech Recognition
Hong Kook, Mina Kim, Yoo Rhee · Sciyo eBooks · 2010
This chapter addressed issues associated with efficient pronunciation variation modeling for non-native automatic speech recognition (ASR), where non-native speech was mostly characterized by different pronunciations, speaking styles, and articulators of speakers from their native speech. The techniques for improving the performance of non-native ASR could then be classified into four approaches: acoustic modeling, language modeling, pronunciation modeling, and hybrid modeling approaches. We first reviewed these four approaches before proposing a new pronunciation model adaptation method. In particular, the proposed pronunciation adaptation method was based on a multiple pronunciation dictionary, designed using an indirect data-driven method. However, this approach resulted in an increased search space for ASR decoding due to the increase of the pronunciation dictionary size. Therefore, a method for optimizing the size of the multiple pronunciation dictionary was also proposed, where a confusability measure based on the Levenshtein distance was introduced in order to remove some confusable pronunciation variants from the dictionary. To investigate the effects of the proposed approach on ASR performance, English was selected as the target language and English utterances spoken by Koreans were considered as the non-native speech. Subsequently, it was shown from the continuous non-native ASR experiments that an ASR system using the optimized multiple pronunciation dictionary could achieve an average word error rate reduction of 15.30%, with a relative reduction in computational complexity of 21.10%, compared to that achieved using the multiple pronunciation dictionary without optimization.