Toward Explainable Automatic Classification of Children’s Speech Disorders

Abstract

Early and adequate diagnosis of speech disorders can contribute to the quality of the treatment and thus to treatment success rates. Using acoustic analysis of the speech of children with speech disorders may aid therapists in the diagnostic process by identifying the acoustic characteristics that are unique to a specific disorder and that distinguish it from normal speech development. The purpose of this work is to investigate the feasibility of the automatic detection of speech disorders based on children’s voices. In this preliminary study, using a dataset of utterance recordings of 24 children whose mother tongue is Hebrew, we propose an automatic system that may facilitate accurate speech assessment by therapists by providing a preliminary diagnosis and explainable insights about the model’s predictions. We built a serial, two-step network that is both powerful and possibly interpretable. The first step can model the complex relations between acoustic features and the speech disorder while the second can shed light on the utterances that make the greatest contribution to the final classification. Our preliminary results focus on the broad spectrum of speech disorders. In future work, we plan to design a system that will be able to detect childhood apraxia of speech (CAS) specifically and shed light on the differences in the speech of individuals with CAS and those with other speech disorders.