Hybrid Neuro-Acoustic Transductive Modeling of Infant Cry State Spaces for Early Physiological Condition Attribution

Abstract

The rapid advancement of intelligent audio-based monitoring systems has increased the demand for automated analysis of acoustic signals in healthcare-related applications, particularly in infant cry analysis where crying serves as the primary mode of communication. Accurately interpreting these signals is challenging due to variations in pitch, intensity, duration, and environmental noise, making reliable classification essential for effective assistance and early-stage monitoring. The problem lies in developing a system capable of distinguishing different categories of baby cries from unstructured audio data, where the existing system is primarily manual and depends on human observation and experience, leading to subjectivity, inconsistency, and lack of real-time applicability. These limitations highlight the need for an intelligent, automated, and scalable solution that minimizes human intervention while improving prediction reliability. To address this, the proposed system utilizes Mel-Frequency Cepstral Coefficients (MFCC) for feature extraction combined with multiple Machine Learning (ML) models, including Support Vector Machine (SVM), K-Nearest Neighbour (KNN), Decision Tree Classifier (DTC), Adaptive Boosting Classifier (ABC), and Linear Discriminant Analysis (LDA) for comparative evaluation, along with a Convolutional Neural Network (CNN) as the primary model for enhanced performance. The CNN enables the system to learn deeper feature representations from audio inputs, improving classification accuracy. The significance of this work lies in providing an automated and reliable method for baby cry classification, assisting caregivers in understanding infant needs, and demonstrating the effectiveness of integrating MFCC features with ML and CNN models for practical healthcare applications.