Efficient personalized mispronunciation detection of Taiwanese-accented English speech based on unsupervised model adaptation and dynamic sentence selection

Abstract

This study presents an efficient approach to personalized mispronunciation detection of Taiwanese-accented English. The main goal of this study was to detect frequently occurring mispronunciation patterns of Taiwanese-accented English instead of scoring English pronunciations directly. The proposed approach quickly identifies personalized mispronunciations of students, enabling English teachers to spend more time on teaching or rectifying student pronunciations. In this approach, an unsupervised model adaptation method was performed on the universal acoustic models to recognize the speech of a specific speaker with mispronunciations and a Taiwanese accent. A dynamic sentence selection algorithm that considers the mutual information of the related mispronunciations is proposed to select a sentence containing the most undetected mispronunciations to quickly detect personalized mispronunciations. The experimental results show that the proposed unsupervised adaptation approach obtains an accuracy improvement of approximately 2.1% in the recognition of Taiwanese-accented English speech.