Visualization of Mandarin Chinese Tone Production of Japanese L2 Learners for evaluation

Aiming at automatic characterisation and evaluation of second language (L2) learners production of Mandarin Chinese tones, we applied robust F0-features previously used to characterize and visualize the tone control of native Chinese speakers. They consist of the average height and the average slope of the contour, which form a minimal set of F0-features to efficiently separate the four Mandarin tones. The resulting two-dimensional scatterplots represent L2 learner’s characteristics of tone production very well, in particular the confusion between tone 2 and tone 3 which has been well known as a L2 learning problem of Chinese tone production. This analysis could be carried out completely automatically using open source tools. To further confirm that the information contained in those two features is sufficient to reflect native’s perception, we used a neural network (NN) to predict natives evaluation of utterances of tone 2 and tone 3 based on the two features. The experiment showed a reasonable high correlation to natives subjective rating, which confirmed the availability of those two features to evaluate L2 proficiency of Chinese tone production.