Automatic lexical stress detection for English learning

We examine a method to automatically detect lexical stress of English, which is a part of a computer-assisted language learning (CALL) system for Chinese to learn English pronunciation. Acoustic differences between vowels in syllables that do or do not carry lexical stress are investigated. We use coefficients based on fundamental frequency, energy and duration to differentiate stressed vowels from unstressed ones by linear classification. Although the distributions of the coefficients overlap to some extent, the new normalized pitch-based coefficient obtains correct classification results of about 84%. Compared with previous detection methods, the results suggest that proper normalization of acoustic feature is much more valuable than the combination of different features.