Using Reference to Tune Language Model for Detection of Reading Miscues

Abstract

For a reading tutor, the reference content which the reader reads is known beforehand. This apriori information is very important in automatic detection of reading miscues. This paper proposed two methods to incorporate the reference information into LVCSR framework to improve the performance of miscue detection. The two methods both tune the n-gram Language Model (LM) probabilities dynamically in the decoding process based on the analysis of current reference sentence. The first method weighs the LM probability directly if current n-gram exists in the reference, and the second method takes a liner combination of the original LM probability and the reference probability. The experiments on a Chinese Mandarin reading corpus proved the effectiveness of both methods. The detection error rate and false alarm rate are decreased by 33.1 % and 35.5% respectively for the best method.