The Comparative Study On Acoustic Model Training Criterion For Automatic Mispronunciation Detection

In this thesis,we research on the performance of automatic mispronunciation detection technology on Xinjiang non-native speaker's pronunciation,it is well know that mispronunciation detection technology is an important manifestation of computer-assisted language learning,Therefore,how to improve the performance of the Xinjiang non-native speaker Mandarin pronunciation error detection system is significant for the masses of ethnic minority learners.In this paper,taking Xinjiang Uygur and Kazakh speech database as an example,using Xinjiang non-native speakers small corpus,By improving and training on traditional pronunciation error detection acoustic model,the discriminative training algorithm can effectively improve the detection performance.The main work of this thesis as follows:Firstly,this paper introduces the calculation method of important index of pronunciation error detection of commonly used measurement GOP algorithm and F1-score which is an important indicator of error-detecting capability evaluation system calculation.Also introduces the maximum likelihood estimation(MLE)training criterion,and use MLE training criterion on Xinjiang non-native speech corpus,calculate the F1-score of the mispronunciation detection system,proposed for the comparison in this article laid the Foundation for the effectiveness of training criterion.Secondly,the traditional speech recognition of discriminative training criteria applied to the mispronunciation detection system,through the analysis of commonly used algorithm for discriminative training,we choose the widely used MPE discriminative training criteria,And verify the MPE discriminative training criterion can effectively improve the phoneme recognition error rate of data of experiments,but is not helpful to improve mispronunciation detection system performance,it shows that the existing discriminative training criteria can not effectively to improve mispronunciation detection system performance.Finally,based on the analysis of MLE and MPE acoustic training model,a discriminative acoustic model training method is proposed,through the establishment of Xinjiang non-native speaker speech database annotated by experts,to construct the objective function of maximize F1-score training criterion,and using to construct the weak-sense auxiliary function to update and optimization of parameters on the objective function,Through the use of threshold value iteration update method to ensure objective function improvement.The discriminative training criterion can automatically to learn standard of mispronunciation on Xinjiang non-native speakers small corpus,and the experimental results show that this criterion can effectively improve the performance of the mispronunciation detection system.