Automated Pronunciation Scoring for L2 English Learners

1. Introduction This study aims at developing an automated pronunciation scoring method for second language learners of English (Hereafter, L2 learners) using both confidence scoring and classifiers. The pronunciation errors have been detected using the confidence measure from speech recognition [Franco et al. However, the accuracy of the assessment based on the confidence scores is not always high. In contrast to the confidence scoring method, which has been the traditional method for pronunciation scoring, the classifier method can be trained using a small amount of data, and achieve high accuracy in the binary classification of phonetic features. Due to these advantages, [Truong et al., 2004] developed a classifier model which distinguishes the correct sounds from incorrect sounds. [Truong et al., 2004] used different acoustic feature vectors according to the target L1 phoneme, but implementing a per phone feature extraction algorithm is a difficult and time-consuming task. [Mark Hasegawa-Johnson and Wang, 2005] used the same acoustic features with speech recognition for all phonemes, but they could achieve high accuracy of a binary feature classification by extracting the feature vector from the appropriate time interval according to the phonemes. In this study, classifiers were trained for the specific English phonemes where L2 learners make frequent errors. Similar to [Mark Hasegawa-Johnson and Wang, 2005], the same acoustic features with speech recognition was used. The pronunciation scoring method based on the combination of two scores improve the accuracy of pronunciation error detection.