EXPLOITING MULTISPECTRAL FEATURES FOR NEONATAL CRY ANALYSIS

Abstract

Infant cries are the primary means of communication for newborns, conveying needs such as hunger, pain, sleepiness, or discomfort. Globally, it is estimated that caregivers misinterpret up to 40% of baby cries, which can lead to delayed responses and potential health risks. Traditionally, the identification of cry types relies on manual listening and observation, requiring caregivers or medical staff to interpret acoustic patterns, which is highly subjective and prone to error. The proposed system combines audio preprocessing, feature extraction using MFCCs, traditional machine learning models such as Support vector machine (SVM) classifier, K-nearest neighbours classifier (KNN), Decision tree classifier (DTC), Adaptive Boosting classifier (Adaboost), and Linear discriminant analysis (LDA), and a 1D Convolutional Neural Network (CNN) to provide a robust framework for categorizing infant cries. By integrating a user-friendly Tkinter-based graphical interface, the system allows seamless dataset uploading, preprocessing, model training, evaluation, and prediction on unseen audio data. Performance metrics including accuracy, precision, recall, F1-score, confusion matrices, and ROC curves ensure rigorous evaluation, while visualizations such as waveforms and training graphs enhance interpretability. This approach not only addresses the limitations of manual systems by providing objective and scalable classification but also enables real-time monitoring and early health intervention. The proposed automated system demonstrates the advantages of combining classical machine learning with deep learning, offering a reliable, efficient, and practical solution for caregivers and healthcare providers to interpret infant needs accurately, ultimately contributing to better infant care and timely medical response.