Computer aided evaluation of intonation for language learning based on prosodic unit segmentation

Abstract

This paper considers computer aided intonation evaluation based on prosodic unit segmentation for the learning of English. To evaluate the intonation of utterances of a foreign English learner, the learner's intonation pattern is normally compared with that of the teacher's utterance. The evaluation score can be obtained by measuring the "distance" between the two intonation patterns. One salient feature of most current computer-aided language learning (CALL) systems is segmentation of an utterance according to word or syllable boundaries for distance measurement in the evaluation process. This method may lead to inaccuracy of the evaluation, since the prosody of natural speech always corresponds to the boundaries of prosodic units, such as foot and intonation unit, rather than the word boundaries in an utterance. In this paper, the pronounced sentences are segmented according to prosodic unit boundaries. Dynamic Time Warping (DTW) and Mahalanobis Distance are then used to measure the difference between the learner's and the teacher's intonation, leading to a machine score.