Classification of Autistic and Non-Autistic Children's Speech: A Cross-Linguistic Study in Finnish, French, and Slovak

A new study by researchers Sofoklis Kakouros and Ida-Lotta Myllylä demonstrates that AI models can detect autism spectrum disorder from children's speech patterns, but with significant variation across different languages. The research, accepted for Speech Prosody 2026, analyzed speech samples from Finnish, French, and Slovak children using supervised classification with acoustic-prosodic features. The Finnish model performed best with 84% accuracy and 0.88 F1 score, while French showed the weakest results at 68% accuracy and 0.56 F1 score, highlighting how language-specific characteristics affect detection reliability.

Cross-language generalization experiments revealed moderate success when transferring models between languages, with Finnish-to-Slovak transfer achieving 0.70 F1 and Slovak-to-Finnish reaching 0.78 F1. However, transfer to French performed poorly at just 0.42 F1, suggesting French speech patterns present unique challenges for autism detection algorithms. Feature-importance analyses indicated that while some acoustic markers of autism appear to generalize across these typologically distinct languages, they are not fully language-invariant.

The researchers emphasize that their work serves as an analytical tool rather than seeking state-of-the-art performance, with findings suggesting future robust cross-linguistic classifiers will need language-aware modeling approaches and more homogeneous recording conditions. This study represents important progress toward developing culturally and linguistically sensitive diagnostic tools that could eventually support earlier autism identification across diverse populations.