Automatically identifying characteristic features of non-native English accents

Abstract

In this work, we demonstrate the application of statistical measures from dialectometry to the study of accented English speech. This new methodology enables a more quantitative approach to the study of accents. Studies on spoken dialect data have shown that a combination of representativeness (the difference between pronunciations within the language variety is small) and distinctiveness (the difference between pronunciations inside and outside the variety is large) is a good way to identify characteristic features of a language variety. We applied this method from dialectology to transcriptions of the words from the Speech Accent Archive, while treating L2 English speakers with different L1s as ‘varieties’. This yields lists of words that are pronounced characteristically differently in comparison to native accents of English. We discuss English accent characteristics for French, Hungarian and Dutch. We compare the French and Hungarian results to phonological descriptions of those languages to identify the source of the difference. The Dutch results are compared to a Dutch accents judgement study (Van den Doel 2006) to evaluate the measure. Knowing about these characteristic features of accents has useful applicationsin teaching L2 learners of English, since potentially difficult sounds or sound combinations can be identified and addressed based on the learner’s native language.