Infant Cry Pattern Analysis and Classification Using Machine Learning Techniques

Abstract

This paper suggests an integrated infant state monitoring system that involves both an image-based model and an audio-based cry classification model with a fusion model that combines their outputs for steady, real-time inference. The audio branch employs time–frequency representations (Mel-frequency cepstral coefficients, MFCCs, and Mel-spectrograms) of infant cries to predict four classes (discomfort, hunger, pain, fatigue) using a convolutional neural network. The image branch makes predictions on facial frames and predicts the same states. A late-fusion approach averages probabilities of classes computed from both branches to decide. The models are trained and run in Colab notebooks ('image', 'cry²') and merged in 'full'. We present data preprocessing, model architectures, training procedures, and deployment strategies for real-time deployment. Empirical analysis demonstrates that the integration offers greater stability compared to individual-modality models, projecting the promise of multimodal learning to infant-care solutions.