Analysis based on Japanese accented English characteristics categories for objective evaluation

Abstract

Aiming at objective evaluation of L2 (second language) learners' proficiency, mispronunciation characteristics of Japanese learners' English speech were analyzed based on phonetic knowledge. Through automatic forced alignment using a speech recognition tool kit, phone deletion, substitution and insertion were detected by substituting the pronunciation dictionary reflecting learners' phonetic error characteristics. For phonetic knowledge, typical mispronunciations, vowel insertion, vowel substitution and consonant omission and substitution were taken into account. These mispronunciation-pattern statistics were analyzed using correlation and segmented linear regression analysis method. The statistics showed consistent error characteristics along the learners' subjective evaluation scores and correlation relationship to the evaluation scores. Break points from the segmented linear regression analysis illustrated difference in the mispronunciations between vowel and consonant. Differential between the obtained break point and each learner's score in the subjective evaluation scores was thought as measurement of the learner's mispronunciation correctness and non-correctness. The correlation value of 0.511 was obtained between actual and estimated differential value by a multiple linear regression analysis, using the mispronunciation statistics as parameters.