원어민과 비원어민의 음성인식 음소 사후확률 비교 연구

Abstract

This study investigated the pronunciation differences in L1 and L2 English speech via the phonetic posteriorgram (PPG) vectors obtained from acoustic model trained for automatic speech recognition (ASR). PPG is the frame-wise phone-classification result in conventional ASR systems, which has been widely applied to Computer-Assisted Pronunciation Training (CAPT). More recently, PPG has been applied to foreign accent conversion, a field of voice conversion research that can be applied to user-adaptable pronunciation training. This study was aimed at investigating the L1-L2 PPG pattern differences between the speech data previously used for these research endeavors. PPG vectors of the 18 native and 24 non-native English speakers were extracted using a deep neural network (DNN) based ASR system. When analyzed in phone-level, results showed that L1 PPGs were more concentrated on correct phone dimension than L2 PPG, which appeared more dispersed across dimensions. The result is in line with the previous PPG-based CAPT research that identified distinct L2 patterns in English speech. The findings are further discussed in the context application to pronunciation training.