EFFICIENTAND ROBUST LANGUAGE MODELING INAN AUTOMATIC CHILDREN'S READINGTUTOR SYSTEM

Recently, there hasbeenarapidly increasing interest inusing ASR forchildren's language learning. An Automatic Reading Tutor system built withASR technologies cantrack children's oral reading against story texts, detect reading miscues, andmeasure thelevel ofreading fluency. Theymayevendiagnose thenature of themiscues andprovide feedback toimprove reading skills. In suchtasks, N-gramlanguage models (LM)maybetrained fromthe wholestory text, ormaybegenerated basedoncurrent story sentence withheuristic probabilities forbothregular wordsinthe sentence and explicitly predicted reading miscues. The disadvantages ofthose methods areeither they require arelatively large textandaretime-consuming, ora large-sized LM and complex processing areneeded toaccommodate all possible words inreading stories aswellasinreading miscues. Thispaper proposes anefficient androbust LM whichcanbeeasily built onthe-fly withcurrent reading sentences. Withanadditional parallel "garbage" model, theLM canalsodealeffectively withawide range ofreading miscues. Ourexperiments inastandard children's reading taskshowthat thenewLM reaches thestate-of-the-art performance indetecting reading miscues withafast speed while onlyarelatively simple children's acoustic modelofspeech was used. IndexTerms-Automatic Reading Tutor, Children's Speech Recognition, Language Model, ASR,Reading Miscues