Infant Cry Scene Detection and Classification Using High-Frequency Wavelet Scattering Transform and Convolutional Recurrent Neural Network

India's high infant mortality rate (IMR) of 30.4 per 1000 live births is exacerbated primarily by delayed diagnoses. To address this, we propose a novel wavelet scattering transform initialized as a first layer in the convolutional recurrent neural network (referred as WST-CRNN) approach that enables early detection of infant health issues through scene detection and multiclass classification of cry sounds in continuous audio, helping to reduce IMR. The WST layer is nontrainable, providing fixed initializations that enhance model robustness; it is inherently stable to deformations and offers translation invariance, making the network more reliable. In contrast to prior studies that classify infant cries into single categories, our work focuses on event detection in continuous audio, enabling the identification of multiple simultaneous categories. To create a realistic dataset, we synthesize composite, multiclass audio files by combining individual single-class cry samples from the imbalanced Donateacry corpus. We leverage the high-frequency WST layer as a robust feature extractor, well suited to the high-frequency nature of infant cries, and achieve superior performance over conventional features, such as mel band energy. Furthermore, WST reduces the need for extensive labeled data, as its transformed-based features are informative even with limited training samples. Using a WST-CRNN, optimized with reduced layers and parameters, the proposed approach achieved an improved error rate of 0.0937 and F1-score of 0.9462, demonstrating effective infant cry scene detection with reduced computational complexity.