An investigation of vowel substitution rules in the automatic evaluation system of English pronunciation

Abstract

We investigate the performance improvement of an automatic evaluation system of the English pronunciation of Japanese learners. In this system, Japanese and English acoustic models are used to detect mispronunciation at a phoneme level. Hidden Markov models (HMMs) are used as acoustic models. Mispronunciation is detected by comparing the output likelihoods of the two models. In order to improve the performance of this system, we investigate certain mispronunciation rules, which represent common mispronunciations among Japanese learners. We use four mispronunciation rules: vowel insertion (at the end of a word), vowel substitution, vowel insertion (between consonants), and consonant substitution. In this system, the accuracy of the mispronunciation rules is particularly important. The rules are determined on the basis of the knowledge of phonetics in our previous system. However, the effectiveness of the rules has not been analyzed quantitatively, and we do so in this work. A knockout procedure is used to select effective rules. By selecting effective rules, we found that the correlation coefficient between the subjective evaluation value and the system performance improved from 0.757 to 0.858.