Automatic detection of pronunciation errors in CAPT systems based on confidence measure

Computer aided pronunciation training (CAPT) systems aim at listening to a learner´s utterance, judging the overall pronunciation quality and pointing out any pronunciation errors in the utterance. However, the performance of the current error detection techniques can not satisfy the users´ expectation. In this paper, we introduce confidence measures and anti-models into the error detection algorithm. The basic theory and the roles of confidence measures and anti-models are discussed. We then propose a new algorithm, and evaluate it on an English learning system. Experiments are conducted based on the TIMIT database and an adaptive database, which involves 40 Chinese undergraduates. The results show that confidence measures can be utilized to effectively improve the performance of the CAPT system.