한국인 영어학습자 발화의 자동음성인식을 위한 영어 음성사전 구성

Abstract

This paper introduces a way of enhancing automatic speech recognition systems in processing non-native speakers utterances. The key procedure is to design and provide an English pronunciation dictionary which includes frequent error pronunciations by Korean learners as if they were regular pronunciation variations. First, knowledge-based rules are collected to generate as many nonnative variants as possible. Then the relative frequency of each variant is measured to select more frequent error-variants while discarding rare ones. In automatic recognition experiments with this dictionary, it is found that the word recognition performance increases considerably for Korean learners incorrect pronunciation. Although utterances by Korean learners of English are the main target data of the current study, the proposed method is expected to be further available in other language environments.