Towards accurate recognition for children's oral reading fluency

Systems for assessing and tutoring reading skills place unique requirements on underlying ASR technologies. This paper presents VersaReader, a system automatically measuring children's oral reading fluency skills. Critical techniques that improve the recognition accuracy and make the system practical are discussed in detail. We show that using a set of linguistic rules learned from a collection of transcriptions, the proposed rule-based language model outperformed traditional n-gram language models. Combined with a specific acoustic model with explicit long silence modeling, plus adaptation, a WER 7.25% was achieved in our test set. The impact of different kinds of rules on performance is also discussed. We demonstrate that VersaReader can provide highly accurate Words Correct Per Minute scores automatically, which are virtually indistinguishable from scores provided by careful human analysis.