Towards a new model of speech evaluation system within the framework of World Englishes in Asia

In the field of speech recognition, attempts have been made to develop an automatic evaluation system for spoken language of second language (L2) learners. The models adopted in the previous studies are a multiple regression model where naturalness of speech is estimated by differences from the speech of the native speakers in the timing characteristics of overall sentence, individual content and function words, pauses, etc. The explained variable is the naturalness evaluation by a few native speakers of the target language. Although the estimation accuracy of the models was validated using open data, some need to be improved in selecting variables and the evaluation by raters. Furthermore, reflecting the situation where English is now used as an international language, it might be controversial that measuring the differences in speech between L2 learners and the native speakers. To revise the models, we collected the data from L2 learners of English with various first language backgrounds in Asia: Korean, Tagalog, Chinese, Malay, Thai, and Japanese. The learners read out a fairy tale, “The North Wind and the Sun”, and were digital-tape recorded. In the current study, we adopt not only the timing characteristics analyzed in the previous studies, but also prosodic features. We investigate the relationship between the existing and new explanatory variables, and propose optimized predictor variables.