This paper presented research results of ajoint project between UTokyo and UCL, whereautomated scoring of fluency was investigated.Since the L2 corpus prepared for developmentwas not large, we tested classical machinelearning techniques with recently proposedspeech representations such as posteriorgramwith variable granularity. Experimentsshowed a correlation of 0.925 to the perceivedfluency, which was higher than the maximuminter-rater (one-to-others) correlation (0.910).
展开▼