Automatic Scoring of Pronunciation Quality with Hybrid Measure

In this paper, a hybrid measure for automatic scoring of Mandarin pronunciation quality is presented. Different to prevalent algorithms, mono-phone-based and tri-phone-based acoustic models are applied and two types of features are combined to get the score of ¿goodness of pronunciations¿ with support vector machine algorithm, which are the average logarithm of the frame-based posterior probability and the normalized logarithm of the phoneme-based posterior probability. With the hybrid measure, the average correlation coefficient between machine scores from automatic system and the rater scores is improved from 0.8379 to 0.8549 which almost reach the coefficient 0.8564 between different raters.