Combining acoustic and language model miscue detection methods for adult dyslexic read speech
Morten Højfeldt Rasmussen, Børge Lindberg, Zheng‐Hua Tan · 2011
One important feature of automatic reading tutors is their ability to detect miscues in order to provide feedback to the student and/or to automatically evaluate the student's reading proficiency.The focus of this paper is on improving accuracy of detecting miscues in dyslexic read speech.We present a miscue detection method that combines a specialized language model and the goodness of pronunciation (GOP) score.The language model is augmented with a subset of the real word substitutions that are observed in the training set.Experiments have been conducted on a corpus containing adult dyslexic read speech.At a miscue detection rate of 34% the false rejection rate (FRR) using only the specialized language model is 2.6%, for the GOP score it's 3.4%, whereas for a combination of the two miscue detection methods the FRR is only 1.8%, which is a 31% relative improvement of FRR when compared to the specialized language model method.