Automatic pronunciation assessment of Korean spoken by L2 learners using best feature set selection

Abstract

This paper proposes a method for automatic pronunciation assessment of Korean spoken by L2 learners by selecting the best feature set from a collection of the most well-known features in the literature. The L2 Korean Speech Corpus is used for assessment modeling, where the native languages of the L2 learners are English, Chinese, Japanese, Russian, and Mongolian. In our system, learners' speech is forced-aligned and recognized using a native Korean acoustic model. Based on these results, various features for pronunciation assessment are computed, and divided into four categories such as RATE, SEGMENT, SILENCE, and GOP. Pronunciation scores produced by combining categories of features by multiple linear regression are used as a baseline. In order to enhance the baseline performance, relevant features are selected by using Principal Component Regression (PCR) and Best Subset Selection (BSS), respectively. The results show that the BSS model outperforms the baseline and the PCR model, and that features corresponding to speech segment and rate are selected as the relevant ones for automatic pronunciation assessment. The observed tendency of salient features will be useful for further improvement of automatic pronunciation assessment model for Korean language learners.