Paediatric-HGNN uses graph neural networks to detect stuttering in children's speech

Researchers from multiple institutions have developed Paediatric-HGNN, a novel framework that uses a context-aware part-whole interaction network (CaPIN) to detect stuttering and disfluency in children's speech. The key innovation is replacing conventional 1D signal processing with a heterogeneous graph that models hierarchical relationships between lexical units (word nodes) and fine-grained acoustic segments (frame nodes). This approach helps distinguish between pathological stuttering and typical developmental disfluencies—a challenge that plagues automated stuttering detection systems due to high acoustic variability in developing voices. The model was trained on curated paediatric corpora UCLASS and FluencyBank, achieving 82.4% weighted accuracy and a Typical Disfluency F1-score of 0.386. By capturing developmental 'searching' behavior, the system offers a more robust and interpretable tool for clinicians.

The paper was accepted at INTERSPEECH 2026 and is available on arXiv (paper ID: 2606.08210). The work highlights how graph-based architectures can outperform traditional sequential models in handling the subtle acoustic cues present in children's speech. As early intervention for stuttering can significantly improve long-term communication outcomes, Paediatric-HGNN represents a step toward more accurate, scalable screening tools that could be deployed in schools or paediatric clinics. The hybrid nature of the network—combining lexical and acoustic insights—also points to a broader trend in speech pathology AI: moving beyond raw signal analysis toward hierarchical, context-aware modeling that mirrors how human listeners process disfluent speech.