Towards Explainable Infant Cry Reasoning Model

Infant crying provides crucial insights into their physiological and psychological status, aiding in early-warning and daily caregiving. However, most studies in infant cry reasoning face challenges like poor generalization caused by infant voiceprint difference and poor interpretability caused by black-box operation of deep learning. To address these two issues, we propose an explainable infant cry reasoning model (EICR) that emphasizes the representation consistency of samples sharing the same label but originating from different infants in a CNN model, and employs a frequency- and energy-based fuzzy decision tree to investigate how the model distinguishes different cry reasons. Specifically, EICR applies the fuzzy decision tree, with energy features from different frequency bands as node attributes, to approximate the CNN's predictions through knowledge distillation. Extensive cross-subject experiments are conducted on a public dataset, showing that EICR improves the accuracy by 2.39% and the F1-score by 1.48% compared to models without EICR. More importantly, the decision tree displays decision rules of different cry reasons step by step, providing a new basis for better understanding of infant cries.