Automatic Call Classification of Autism Model Marmosets by Deep Learning and Analysis of Their Vocal Development

In this study, we aim to automatically classify calls of autism model marmosets to analyze their vocal development. As a nonhuman primate model of autism spectrum disorder (ASD), valproic-acid-exposed (VPA) marmosets have recently been developed. This marmoset model has been shown to have abnormal vocal development similar to ASD symptoms compared with VPA-unexposed (UE) marmosets. Automatic classification of vocalizations and analysis of calls can be used to study the VPA marmosets. In the past, there have been no studies on the automatic classification of VPA marmoset calls. Therefore, we focus on classifying calls of both UE and VPA marmosets automatically and applying the automatic call classification to the analysis of their vocal development. We construct a call classifier based on deep learning, which distinguishes four types of marmoset call from audio signals. By experiment, we confirmed that the classifier trained on calls of UE marmosets could accurately classify two main classes of calls: 'phee' and 'twitter'. The classifier also worked well for VPA marmosets. Furthermore, we found that the classifier was also helpful for analyzing the vocal development of UE and VPA marmosets.