조음 기반의 음소 레벨 사후 확률을 이용한 한국인 영어 학습자의 자음 발음 오류 검출

Abstract

This paper proposes novel features for automatic pronunciation error detection of English consonants produced by Korean learners by using articulatory Goodness-Of-Pronunciation(aGOP). For calculating aGOP, articulatory acoustic models are trained for the corresponding articulatory attributes, such as nasal, sonorant, anterior, etc. For error detection modeling, we utilize 2,500 English sentences spoken by Korean learners. By comparing canonical and actual pronunciations, we select six consonant phonemes(/z, ð, θ, t, v, d/) which show high pronunciation error rate from 9.3% to 27.2%. Error detection modeling is performed in terms of each phoneme. As a baseline, the traditional GOP is applied for the modeling. In order to enhance the baseline performance, we append the proposed aGOPs as novel features. The results show that the proposed model outperforms the baseline for all consonant phonemes. It is noteworthy that integrating linguistic/phonetic knowledge is useful for automatic pronunciation error detection.