The basis of the automatic assessment of pronunciation is Hidden Markov Models (HMMs) that generate segmentation and scoring of the non-native learners' speech. The threshold functions calculated from the means and standard deviations of native speakers' speech samples are used to automatically detect pronunciation errors. Our detection methods focuses on the great degradation in scores that suggests pronunciation errors, because the conventional methods of automatic assessment have not always correlated well with human ratings due to the speaker variability. For automatically detected pronunciation errors, the evaluation by human judges was carried out to verify the reliability of our methods.
