Modeling of stressed syllables for their detection in English sentences to develop an English rhythm learning system

摘要

Current globalization requires Japanese people to speak English with correct pronunciation. To acquire this ability efficiently, they need to master generating English rhythm with positioning appropriate stressed syllables. With the aim of the developing an English rhythm learning system for Japanese learners, in this paper we proposed methods to model stressed/unstressed syllables in English sentences, where relative differences in prosodic features between two adjacent syllables and vowel-based modeling in addition to a syllable-based one were investigated. In experiments, with the previously proposed modeling method, about 91% accuracy was obtained as base-line in the speaker-closed/text-open condition. When using other speakers, approximately 80% and 70% accuracy was obtained respectively with native speakers and Japanese ones. The proposed methods showed that, in the speaker-open/text-open condition, 5% relative improvement was obtained by using the relative differences between two consecutive syllables and 8% improvement was observed for Japanese speakers by using the vowel-based modeling.