Bio-inspired graph attention network for intelligent neonatal cry classification

Life Cycle Reliability and Safety Engineering · Springer Nature · 2026-10-02 · DOI: 10.1007/s10772-026-10291-8

Аннотация

The study of neonatal vocalisation has received growing interest in speech processing research because it can be used in monitoring infant distress for early health evaluation. Nevertheless, the consistent classification of neonatal signals of crying is not easy due to high intra-class variability, overlapping acoustic patterns, and temporal dependencies of cry segments. Traditional speech classification techniques using Cepstral features and deep-network systems tend to handle segmentation separately and do not support the analysis of the contextual relationship between segments of vocalisation. This paper proposes a Bio-Inspired Optimized Graph Attention Network (B-GAT) to learn context-related neonatal vocalisation recognition. An auditory-inspired Gammatone Frequency Cepstral Coefficient (GFCC) representation is extracted from each vocalisation recording. The resulting segment-level representations are organized into a relational graph, where nodes represent vocalisation segments and edges encode temporal continuity and acoustic relationships. A multi-head graph attention mechanism is employed to learn context-dependent representations by adaptively aggregating information from neighbouring segments. Particle Swarm Optimization is integrated to search for an effective dataset-specific configuration of the graph-attention framework, and the optimized model is subsequently used for neonatal vocalisation classification. The experimental results demonstrate competitive performance across the evaluated datasets, with particularly strong results on the multi-class vocalisation datasets.

Читать у издателя →

Похожие статьи