An approach for infant cry sound classification

Abstract

Infant cry analysis is a promising application of artificial intelligence and audio signal processing, as crying represents the primary means of communication during the first months of life. This work aims to develop an intelligent system capable of automatically classifying infant cries according to the need expressed by the infant. The study is based on the Donate-a-Cry dataset, which contains five cry categories: hungry, discomfort, belly pain, burping, and tiredness. This dataset reflects real-world infant crying conditions, where hungry is one of the most common causes of crying during early infancy, naturally resulting in a larger representation of this category compared to the others. Consequently, classification challenges arise that require models capable of maintaining balanced performance across all cry types. To address these challenges, a Hierarchical assification Framework (HCF) combining Support Vector Machines (SVM) and Random Forest (RF) is roposed, together with acoustic feature extraction, data augmentation, and SMOTE-based oversampling techniques to improve model generalization. In addition, a Streamlit based web application was developed to enable real-time infant cry classification and provide an accessible tool for parents and caregivers.