In this paper, we propose a statistical evaluation method of pronunciation proficiency and intelligibility for presentations made in English by native Japanese speakers. We statistically analyzed the actual utterances of speakers to find combinations of acoustic and linguistic features with high correlation between the scores estimated by the system and native English teachers. Our results showed that the best combination of acoustic features produced correlation coefficients of 0.929 and 0.753 for pronunciation and intelligibility scores, respectively, using open data for speakers at the 10-sentence level. In an offline test, we evaluated possibly-confusing pairs of phonemes that are often mispronounced by Japanese speakers of English. In addition, we developed an online real-time score estimation system for Japanese learners of English using offline techniques to evaluate the pronunciation and intelligibility scores in real-time with almost the same ability as English teachers. Finally, we show that both the objective and subjective evaluations improved after learning with our system.
