LMU-Based Sequential Learning and Posterior Ensemble Fusion for Cross-Domain Infant Cry Classification

A research team from the University of Ottawa and Carleton University has published a novel AI framework designed to tackle the challenging problem of automatically classifying the causes of infant crying. The paper, "LMU-Based Sequential Learning and Posterior Ensemble Fusion for Cross-Domain Infant Cry Classification," addresses core issues in healthcare monitoring: short, non-stationary audio signals, limited annotated data, and significant domain shifts between different infants and datasets. The proposed system aims to move beyond lab conditions to a practical tool that can generalize across real-world scenarios.

The technical innovation centers on a compact acoustic model that fuses three types of audio features—MFCCs, STFT, and pitch—using a multi-branch convolutional neural network (CNN) encoder. For modeling the sequence of these features, the researchers employed an enhanced Legendre Memory Unit (LMU), a recurrent neural network variant that provides stable sequence modeling with substantially fewer parameters than traditional LSTMs, enabling efficient deployment. A key contribution is the "calibrated posterior ensemble fusion" technique, which uses entropy-gated weighting to intelligently combine predictions from domain-specific experts, mitigating dataset bias and improving cross-dataset generalization. Experiments demonstrated improved macro-F1 scores under rigorous cross-domain evaluation protocols, including leakage-aware data splits, confirming the model's robustness and its feasibility for real-time, on-device monitoring applications.