In this paper, to realize automatic phoneme error detection from learners' speech, their pronunciation variation is modeled with the G2P toolkit, Phonetisaurus. The trained model captures relationships between graphemes (visual words) and phonemes that will be actually generated by Japanese learners when they look at the words and read them aloud. For training the model, we collected phonemic transcriptions of 800 Japanese English utterances from two native speakers, who majored in phonetics. Using the built model, a Japanese English (JE) speech recognizer, and an American English (AE) speech recognizer, we detected phoneme errors automatically from approximately 24K unlabelled JE utterances. Using these automatically detected errors, we can train a new G2P model with a larger amount of training data in a semi-supervised way. We test the initial pronunciation variation model and the newly built variation model in the task of phoneme error detection and coverage of variation.
展开▼
