Detection of intonation in L2 English speech of native Mandarin learners

Abstract

We aim to detect salient mispronunciations in intonation of English speech uttered by Mandarin speakers. The goal of our project is to detect intonation errors and provide corrective feedback to English second language (ESL) learners. An intonational event includes the pitch accent and edge tone, and the intonation is closely related to the nuclear tone of an intonational phrase (IP). Hence, we first develop a pitch accent detector to delineate the scope of analysis in an utterance. Then we develop a nuclear tone detector to classify the intonation of the IP as either rising or falling. The pitch accent detector is a Gaussian mixture model using the features based on energy, pitch contour and the duration of the vowels. The intonation detector is a Gaussian discriminator using three features derived from the pitch contour. Annotated L2 English speech from 40 Mandarin speakers is used in a 10-fold cross-validation setting. The pitch accent detector achieves an accuracy of 72.86%, while its EER is 33.00%. The average classification performance of the intonation detector is 91.17% in accuracy and the EER is 8.60%.