Evaluation of Japanese manners of generating word accent of English based on a stressed syllable detection technique

Di erent languages sometimes use di erent acoustic manners to transmit the same kind of linguistic information. This fact lets us easily suppose that learners tend to transmit the information in a manner of not a language to learn but their native language. While English word accent is linguistically almost the same as Japanese one, the word accent acoustically di ers between the two languages. This paper aims at automating the detection of the generated word accent and the evaluation of how it is generated. By using context-sensitive HMMs, stressed and unstressed syllables were modeled separately for their structure and for their position in a word. Results showed that 90% of the stressed syllables were correctly detected in open experiments. In the matching process, each of the likelihood scores derived from di erent parameters was multiplied with its weighting factor. The optimal combination of the factors for the detection can be thought to re ect each speaker's own manner of the accent generation. Di erences of the quasi-optimal combinations between Japanese and native speakers mainly accorded with ndings reported in previous studies on English teaching to foreign learners.