Measuring Phoneme-Level Pronunciation Deviations in Japanese Learners of English Using Self-Supervised Speech Representations

Research suggests a part of the reason for the limited English Proficiency, or bilingualism, in Japan, is tied to a societal 'fear' of mispronouncing certain words in English. Accurate pronunciation feedback is therefore essential for Japanese learners of English, but traditional Computer-Assisted Language Learning (CALL) software typically provides general feedback, lacking the necessary regional nuance required for greater effectiveness. This paper provides insight into the phoneme-level deviations of Japanese speakers of English, compared to a baseline of American English speakers, conducted on the UME-ERJ corpus by the NII-SRC (Japan's National Institute of Informatics-Speech Resources Consortium) using a self-supervised speech representation (SSSR) model, which, to the best of our knowledge, represents the first attempt to quantify pronunciation deviations in Japanese-English using this state-of-the-art method.