Automatic prediction of intelligibility of English words spoken with Japanese accents — Comparative study of features and models used for prediction

Abstract

This study investigates automatic prediction of the words in given sentences that will be unintelligible to American listeners when they are pronounced with Japanese accents. The ERJ intelligibility database [1] contains results of a large listening test, where 800 English sentences read with Japanese accents were presented to 173 American listeners and correct perception rate was obtained for each spoken word. By using this database, in our previous study [8], an intelligibility predictor was built for each word of input texts or utterances. For prediction, lexical and linguistic features were extracted from texts and pronunciation distance and word confusability were calculated from utterances. CART was used as prediction model. In this paper, new features that are related to speech prosody and three new prediction models of ensemble methods (Adaboost, Random Forest and Extremely Randomized Trees) are tested and compared to the old features and model. Finally, our new system can predict very unintelligible words and rather unintelligible words with F1-scores of 72.74% and 84.78%, respectively.