A two-pass architecture for detecting reading miscues

In this paper, we present a CALL system with novel two-pass architecture for sentence reading miscues detection. The research is concentrated on the effect of the language model (LM) of the system, which is necessary for recognizing what is actually spoken by the speaker. We compared the two situations of using LM or not in a one-pass baseline system at first, and found that LM can lead to relatively 60% improvement of miscue detection rate and 80% reduction of false alarm rate. However, the LM still has bad effect on detecting speech errors because the reading miscues are abnormal word sequences and can be easily depressed by it. So we propose to rescore these instances in a second-pass decoding without LM. By means of the second-pass, the miscue detection rate can be improved by 9.6% relatively and the false alarm rate can be reduced by 15.8% relatively.