Accent reduction for computer-aided language learning

This paper studies accent reduction techniques which are used to provide English learners with their own converted speech as a reference for speaking skills training. Three kinds of modifications, namely prosodic modification, segmental modification and combined modification are compared and examined using objective measurements. Two different corpora, a prosody abundant corpus and a prosody flat corpus, are used in our study. Modified utterances show a clear reduction of accentedness and an acceptable acoustic quality when compared with the original speech, demonstrating the effectiveness of the proposed accent reduction techniques. Furthermore, differences between experimental results from the two corpora in terms of the reduction of accentedness are observed and explanations for this phenomenon are presented. This paper also discusses other issues in this area for future research.