When Children Talk and Machines Listen: Toward an Interpretable Speech-Based Screener for Dutch Developmental Language Disorder

We investigate the automatic detection of Developmental Language Disorder (DLD) from semi-spontaneous Dutch child speech. We train classifiers exclusively on acoustic features, without relying on transcriptions. Using heterogeneous data from multiple sources, we address three research questions: (1) whether domain adaptation can reduce cross-domain variability; (2) whether self-supervised learning (SSL) speech representations outperform handcrafted acoustic features; and (3) which handcrafted features are consistently informative across datasets. Our results show that domain adaptation effectively mitigates domain effects. Moreover, handcrafted features achieve performance comparable to SSL features while providing greater transparency and interpretability. Finally, the most informative handcrafted features align closely with established clinical markers of DLD, supporting their use in interpretable and clinically meaningful early screening.