Towards Cross-lingual Automatic Diagnosis of Autism Spectrum Condition in Children's Voices

Automatic diagnosis of Autism Spectrum Conditions (ASC) from the voice is still in its infancy. The comparably few studies up to now focus mostly on the relevance of acoustic features and optimal learning algorithms. However, cross-lingual studies with a higher number of speakers are a white spot in the literature. The present contribution thus focusses on extensive cross-lingual evaluations based on four databases collected in English, French, Hebrew, and Swedish. The datasets contain speech of children with ASC and typically developing (TD) children matched in both age and gender. Their speech further intentionally varies in emotion. This introduces an additional challenge besides the change of languages. We demonstrate automatic ASC vs TD classification to be feasible despite such variation with a remaining error.