Improving Infant Cry Recognition Using MFCC And CNN-Based Audio Augmentation

Recognizing infant cries is essential for understanding a baby's needs; however, previous research has struggled with imbalanced datasets and limited feature extraction techniques. Conventional methods utilizing CNN without data augmentation often failed to accurately classify minority classes such as belly pain, burping, and discomfort, resulting in biased models that predominantly recognized majority classes. This study proposes an MFCC-based data augmentation pipeline, incorporating time stretching, pitch scaling, noise addition, polarity inversion, and random gain adjustments to increase dataset diversity and enhance model generalization. By applying this approach, the dataset size was expanded from 457 to 8,683 samples, and a CNN model with three convolutional layers, ReLU activation, and max pooling was trained for cry pattern classification. The results indicate a substantial accuracy improvement from 78% to 98%, with F1-scores for minority classes rising from 0.00 to above 0.90, confirming that augmentation effectively addresses dataset imbalance. This research advances computer science and artificial intelligence, particularly in audio signal processing and deep learning for healthcare applications, by demonstrating the role of data augmentation in improving cry classification performance. Future directions include integrating multimodal data (visual and physiological signals), exploring advanced deep learning architectures, and developing real-time applications for smart baby monitoring systems to further enhance infant cry recognition technology.