In this chapter we have investigated the feasibility of applying automatic pronunciation assessment and pronunciation error detection into the CAPL system. Experimental results suggest that the automatic scoring of proficiency in L2 sounds is possible. Particularly, statistical analysis based on many speech samples has two advantages: (1) Detecting pronunciation errors caused by linguistic disparity, (2) Showing high reliability in speech recognition methods. We have presented the algorithm being developed to generate reliable pronunciation scores. In the experiments, the automatic scoring method based on HMM worked effectively and showed a high correlation with human ratings. With respect to the definition of proficiency, it is a very good proficiency indicator.
In order to decide pronunciation error, the use of thresholds calculated from the speech of native speakers was effective, although it is true that not all native speakers are completely fluent. We defined three kinds of thresholds for their respective tasks. It is not clear which thresholding method, absolute or relative, is desirable. However, our experiment of task 2 has shown that combined threshold function type of both absolute and relative method worked well with human ratings.
In conclusion, these findings suggest that the use of speech evaluation together with automatic speech recognition techniques may contribute to developing a practical CAPL system. This approach has enormous potentials for the future of automatic pronunciation assessment.
