Automatic detection of rhythmic patterns in native and L2 speech: Chinese, Japanese, and Japanese L2 Chinese

To explore possible contribution of speech rhythm to foreign accent, this study conducted statistical analysis and realized automatic detection of rhythmic patterns on Mandarin Chinese, Japanese and Japanese second language learners (L2) of Chinese using interval-based and amplitude-based measures. Classification models of Support Vector Machine (SVM) and Multilayer Perceptron (MLP) were trained and perceptual experiment was conducted to examine the effectiveness of the proposed method. Results showed: 1) Japanese L2 Chinese (JL2C) are different in rhythmic pattern from both Native Chinese (NC) and Native Japanese (NJ); 2) Correction rates of classification model SVM and MLP are 97.38% and 97.10%, respectively; 3) Average detection rate of five human experts is 89.9%. The high consistency between the statistical models and human experts indicates that measures we used are effective in characterizing rhythm difference between NC, NJ and JL2C and the framework we proposed is promising in exploring the possible contribution of speech rhythm to foreign accent.