Estimation of children's reading ability by fusion of automatic pronunciation verification and fluency detection

Pronunciation verification of children’s reading is a difficult task in itself, but automatic reading assessment software must also detect and evaluate other phenomena that influence human evaluators. Using an isolated word-reading task, we first show that humans use both pronunciation correctness (accuracy) and fluency information in their assessment of the reading ability of kindergarten to second grade children. Next, we used disfluency-specialized grammars and trained a Bayesian Network to automatically classify the fluency and accuracy of an utterance. Finally, we used these automatically determined scores to estimate evaluators’ scores of children’s reading ability with a 0.91 correlation.