Evaluating the accuracy of decoding in children who read aloud

Digital tools based on automatic speech recognition (ASR) could be a useful support for teachers in assessing the reading skills of the students. We focus on the evaluation of the decoding accuracy of children with grade level ranging from the 3rd to the 6th performing a reading aloud task on a narrative text displayed on an ordinary tablet using the ReadLet platform. On the basis of previously collected data, we built a gold dataset with sentences characterised by the audio data, the original text to be read, and the text actually spoken by the child. By using the open-source Kaldi toolkit an ASR system based on the GMM-HMM model was trained on the training portion of the gold dataset. The accuracy of the ASR system was calculated as the ability to correctly decode the test audio data with respect to the annotated text, and the decoding accuracy of the children was estimated by measuring the gap between the results obtained with the annotated text and the original text. A consistent trend with increasing grade level was found in terms of word correctness, substitutions and insertions, while the trained model appears to be significantly able to evaluate the children decoding accuracy.