Speech and Language Technology in Education (SLaTE 2013)
This study examines automatic detection of the words that will be unintelligible if they are spoken by Japanese speakers of English. In our previous study, 800 English utterances spoken by Japanese speakers, which contained 6,063 words, were presented to 173 American listeners and correct perception rate was obtained for each spoken word. By using the results, in this study, we define the words that will be very unintelligible through Japanese accented English pronunciation and also define the words that will be rather unintelligible. Then, by using Classification And Regression Tree (CART) with linguistic features and lexical features only, we examine automatic detection of these words. After that, we introduce an additional feature derived by considering phonological and phonotactic differences between Japanese and English. This additional feature is found to be very effective and our proposed method can detect very unintelligible words and rather unintelligible words automatically with F1-scores of 65.44 and 70.45 [%], respectively.
Index Terms: speech intelligibility, second language learning, foreign accent, ERJ database, CART
Bibliographic reference. Pongkittiphan, Teeraphon / Minematsu, Nobuaki / Makino, Takehiko / Hirose, Keikichi (2013): "Automatic detection of the words that will become unintelligible through Japanese accented pronunciation of English", In SLaTE-2013, 109-111.