Lightweight deep learning system for infant-cry recognition with real-time notification in resource-constrained environments

• Lightweight deep learning pipeline for baby cry recognition on embedded devices. • Transfer learning from YAMNet embeddings combined with CNN, MLP, and Random Forest. • CNN achieved 95.2% accuracy, F1 = 0.93, ROC-AUC = 0.96 on 9,000 audio segments. • Real-time notification via Flask-WhatsApp API with < 0.8 s end-to-end latency. • Robust performance in noisy, resource-constrained environments (Raspberry Pi). Ensuring infant safety is a major challenge, especially when constant supervision is not possible. Crying is the main acoustic cue that reveals an infant’s needs. However, most baby monitors perform poorly in noisy or low-resource environments. The authors propose a lightweight deep-learning system that links YAMNet transfer embeddings with a compact convolutional neural network (CNN). A Flask microservice connects the model to WhatsApp, sending alerts to caregivers in real time. The framework runs smoothly on a Raspberry Pi 4B and was trained on 9,000 audio clips drawn from Kaggle and home recordings. The CNN reached 95.2% accuracy, 0.93 F1-score, and 0.96 ROC-AUC, surpassing both MLP and Random Forest models. Latency from audio capture to message delivery stays below 0.8 s, even with background noise. By combining deep-audio transfer learning, IoT-based communication, and instant messaging, this work delivers a novel, reproducible, and low-cost intelligent monitoring solution for infant-cry detection in resource-limited settings.