A text-independent method for estimating pronunciation quality of Chinese students

Abstract

In this paper, a novel approach is proposed for text-independent pronunciation quality assessment of Chinese students. We call the proposed method as double-models pronunciation scoring algorithm, which separates recognition from assessment stage. It can solve low recognition performance of standard method and score mismatch of nonstandard one. Applying the combination of Maximum Likelihood Linear Regression and Maximum A Posteriori adaptation achieves good recognition results for speech of Chinese students. Adjustment of scoring features signifies further improvement in correlation between machine scores and human judgment. The experimental results showed the proposed double-models technique reached good outcome for text-independent pronunciation quality assessment of Chinese students.