INFANT CRY DETECTION USING DEEPLEARN ING: EXPLORING DEEP LEARNING ARCHITECTURES AND OVERSAMPLING TECHNIQUES

Abstract

Infant crying is one of the earliest and most important non-verbal signals used to communi cate physiological and emotional needs such as hunger, discomfort, and pain etc. Infant cry classification can help caregivers and healthcare professionals in early diagnosis and timely in tervention. Traditional classifiers and handcrafted feature-based methods often fight with class imbalance, noise, and variability in real-world audio dataset. Recent advances in DL(Deep Learning) have demonstrated the effectiveness of models such as CNN (Convolutional Neu ral Network), RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), and GRU (Gated Recurrent Unit) in learning temporal and spectral property of infant cries. This study presents a comparative study of these four DL architectures using five oversampling tech niques:SMOTE (Synthetic Minority Oversampling Technique), Borderline-SMOTE, ADASYN (Adaptive Synthetic Sampling), and RMO-DA (Random Oversampling with Data Augmenta tion). The combined dataset contains 873 audio samples across eight different cry classes, that comprises hungry, burping, belly pain, discomfort, crying, laughing, noise, and silence. The models were trained using the same preprocessing, augmentation, and evaluation procedures to ensure fairness in their comparison. The results reveal that the CNN model performed the best among all models, attaining a classification accuracy rate of 93%, which can be attributed to its ability to extract spectral information efficiently. On the other hand, the RNN and LSTM models were better at recalling minority classes owing to their learning capabilities in time rela tionships. Use of oversampling always led to better detection of minority classes; but according to Wilcoxon signed-rank test results, this difference was statistically insignificant due to small sample sizes. These results highlight the significance of choosing proper deep learning archi tectures combined with adequate methods of dealing with data imbalance to create infant cry detection systems.