An Application of Mispronunciation Detecting Network for Computer Assisted Language Learning System

This paper put forth a technique called Error-Detecting Network of Pronunciation (EDNP) that is applied to Computer Assisted Language Learning (CALL) system. By comparison with state-of-the-art CALL systems, the application of this kind of network is to insert mispronunciation detection routes into task-specific Finite State Grammar (FSG) network and avoid constructing complex mispronounced models. The detailed procedures of how to construct mispronunciation detection network and how to perform an error callback strategy are introduced in this paper. The algorithm is simply to be implemented and is independent to any speech toolkit. The experiments show that the application of this network achieves a False Acceptance Rate (FAR) of 7.38%, as well as a False Rejection Rate (FRR) of 12.25% for the deletion errors and achieves a FAR of 4.94%, as well as a FRR of 26.17% for the insertion errors. Furthermore, compared to traditional forced alignment, there is 4.29% improvement to correlation rate between the objective and the subjective pronunciation quality evaluation by using EDNP.