Normalization of GOP for Chinese Mispronunciation Detection
Wenwei Dong, Yanlu Xie · 2019
Goodness of Pronunciation (GOP) is a kind of Computer-Assisted Pronunciation Training (CAPT) technique that can provide language learners with scoring feedback, and its accuracy easily suffers from the performance of model alignment and phone classification. In order to reduce the influence of those aspects, this paper proposes two ways to normalize GOP scores. The first is to separate the GOP calculation of Chinese Initials and those of Chinese Finals. The second is to use the corresponding native pronunciation score as a template to scale the non-native one. In 2-hours test set of Japanese speaking Chinese corpus, the experiment results show the average relative improvement of Diagnose Accuracy (DA) in the approach one is 16.9%, and 28.7% in scaling approach comparing to the traditional scoring method. The combination of those two methods achieves the best performance. The result is 35.9% of average relative improvement. Experimental results demonstrate the effectiveness of the two methods.