TriSense-AttentionNet: A CNN-LSTM With Multi-Head Attention Fusion Model for Infant Cry Classification

Deep learning has significant promise in automated infant cry analysis for early health assessment in neonatal care. However, existing approach faces two major limitations. First, most methods rely on single audio representations, missing complementary acoustic information across different domains. Second, temporal dependencies in cry sequences are often inadequately captured, reducing classification reliability. To overcome these challenges, this paper introduces Trisense-AttentionNet, a multimodal frame-work that combines parallel feature extraction with attention-based temporal modeling. The architecture simultaneously processes spectrograms, cochleagrams, and chomograms through dedicated CNN-LSTM pathways, capturing frequency, auditory, and harmonic characteristics respectively. Multihead attention layers identify diagnostically critical cry segments across time. Features from all three pathways undergo concatenation before final classification through dense layers. Compared to the existing deep learning models with single audio representations, experiments on the Baby Chillanto Dataset demonstrate superior performance, achieving 94.73% accuracy, 91.30% precision, 92.58% recall, and 91.52% F1-score across five cry types: Asphyxia, Deaf, Hunger, Normal, and Pain. Notably, the proposed approach outperforms the baseline CNN model by 3.74% and the CNN–LSTM model by 0.66%, highlighting the effectiveness of multi-representation feature fusion and attention mechanisms. The results demonstrate that the suggested model can effectively reduce the model overfitting problem caused by a small dataset. The proposed method fused features outperforms existing methods based on single domain features.