Individual-based accent clustering of World Englishes using the Speech Accent Archive

Abstract

In a language class of English, one specific type of pronunciation is often adopted as model pronunciation and learners try to imitate that pronunciation. Even though they acquire native-like pronunciation successfully, once they go out to the real world of spoken English, they surely encounter the pronunciation diversity of English and feel the necessity to survive it. As is well-known, English pronunciation is altered variously depending on the language background of speakers, called accents. This talk describes a technical attempt of automatic and individual-based clustering of these accents using the Speech Accent Archive. This technique may be used to realize a really individual-based map of World Englishes, which can be used to introduce learners to the real state of English. For clustering, the accent gap has to be quantitatively estimated between any two speakers of the archive, where non-linguistic factors such as age and gender have to be adequately cancelled in estimation. Pronunciation structure analysis is used for cancellation and experiments show that the estimated gaps have a better correlation to IPA-based manually defined gaps than phoneme-based gaps have to the IPA-based gaps.