Automatic pronunciation error detection in non-native speech: the case of vowel errors in Dutch

This research is aimed at analyzing and improving automatic pronunciation error detection in a second language. Dutch vowels spoken by adult non-native learners of Dutch are used as a test case. A first study on Dutch pronunciation by L2 learners with different L1s revealed that vowel pronunciation errors are relatively frequent and often concern subtle acoustic differences between the realization and the target sound. In a second study automatic pronunciation error detection experiments were conducted to compare existing measures to a metric that takes account of the error patterns observed to capture relevant acoustic differences. The results of the two studies do indeed show that error patterns bear information that can be usefully employed in weighted automatic measures of pronunciation quality. In addition, it appears that combining such a weighted metric with existing measures improves the equal error rate by 6.1 percentage points from 0.297, for the Goodness of Pronunciation (GOP) algorithm, to 0.236.