Automatic evaluation of reading aloud performance in children

Abstract

Evaluating children's reading aloud proficiency is typically a task done by teachers on an individual basis, where reading time and wrong words are marked manually. A computational tool that assists with recording reading tasks, automatically analyzing them and outputting performance related metrics could be a significant help to teachers. Working towards that goal, this work presents an approach to automatically predict the overall reading aloud ability of primary school children by employing automatic speech processing methods. Reading tasks were designed focused on sentences and pseudowords, so as to obtain complementary information from the two distinct assignments. A dataset was collected with recordings of 284 children aged 6–10 years reading in native European Portuguese. The most common disfluencies identified include intra-word pauses, phonetic extensions, false starts, repetitions, and mispronunciations. To automatically detect reading disfluencies, we first target extra events by employing task-specific lattices for decoding that allow syllable-based false starts as well as repetitions of words and sequences of words. Then, mispronunciations are detected based on the log likelihood ratio between the recognized and target words. The opinions of primary school teachers were gathered as ground truth of overall reading aloud performance, who provided 0–5 scores closely related to the expected performance at the end of each grade. To predict these scores, various features were extracted by automatic annotation and regression models were trained. Gaussian process regression proved to be the most successful approach. Feature selection from both sentence and pseudoword tasks give the closest predictions, with a correlation of 0.944 compared to the teachers' grading. Compared to the use of manual annotation, where the best models obtained give a correlation of 0.949, there was a relative decrease of only 0.5% for using automatic annotations to extract features. The error rate of predicted scores relative to ground truth also proved to be smaller than the deviation of evaluators’ opinion per child.