An Application of Modified Confusion Network for Improving Mispronunciation Detection in Computer- aided Mandarin Pronunciation Training

In this paper, we propose an application of confusion network for Mandarin mispronunciation detection. Compared to former published works, which are proven to work effectively and robustly in detecting mispronunciation in word level and only successfully detect mispronunciation in sentence level in strictly small constrained search space, our modified confusion network based Computer-aided Pronunciation Training (CAPT) system is designded for exploring mispronunciation detections in sentence level with less constrained search space. Our benchmark test based on this improved CAPT system shows that in sentence level the mispronunciation detecting precision rate in strict constrained search space is up to 91.8% for the substituted cases, 93.6% for the deleted and 90.6 for the inserted, while the Recall rates for all three cases are above 90%, which has achieved much more improvement than system without this technique.