Infant Cry Detection With Lightweight Wavelet Scattering Networks

Devices equipped with algorithms for detecting infant cries allow parents to respond immediately to their crying babies. However, improving detection performance often requires complex algorithms that consume more computing resources, leading to increased power consumption and prices. In this study, we propose a novel approach that leverages first-order Wavelet Scattering Coefficients (WSC) as translation-invariant and deformation-stable representations of infant crying sounds. Based on this approach, we introduce an end-to-end Deep Neural Network (DNN) architecture designed to detect crying using merely 17K parameters and 22.7M MACs. The accuracy results, with a 96.98% accuracy rate on open-source datasets, demonstrate the effectiveness and robustness of our model for detecting infant cries in real-world environments.