Classification of Child Vocal Behavior for a Robot-Assisted Autism Diagnostic Protocol

Abstract

Autism is a neurodevelopmental disorder affecting an increasing fraction of children, with severe social and economic consequences for affected persons and their families. Including robotic technologies in the diagnostic process could potentially increase its speed and reliability, opening the way towards earlier and more efficient therapy. The diagnostic process requires multimodal interaction, in which the vocal behavior of the child plays an important role. In this paper, we present a method for automatic classification of child vocal behavior, based on supervised learning, which is suitable for real-time execution on an autonomous robot with limited computational resources. The main contribution of the paper is an empirically determined minimal set of sound features, which allow efficient vocal behavior classification of preschool children, relevant in the context of autism diagnostics. The classifier is verified on a dataset combined from publicly accessible audio recordings and recordings collected during diagnostic and therapeutic sessions.