An Analytical Study of Data Augmentation Across Audio Representations for Infant Cry Classification

Several multidisciplinary studies consider an infant’s cry as a valuable source of information, particularly for parents, caregivers, and medical professionals. From a signal processing viewpoint, infant cries can be represented either in the time domain (one-dimensional or 1D raw waveform) or in the time–frequency domain (two-dimensional or 2D spectrogram-based representations). However, the impact of these representations on classification performance, particularly under constrained and imbalanced dataset conditions, remains insufficiently explored. This study presents a comparative analysis of 1D and 2D convolutional neural networks applied to waveform and spectrogram representations of infant cries. Due to the significant class imbalance of the dataset, we employed data augmentation techniques. Experimental results show that the 1D CNN achieved 95% training accuracy and 91% validation accuracy, indicating a relatively small generalization gap. In contrast, 2D CNN reached 98% training accuracy but remained below 91% on the validation set, revealing a larger gap and suggesting potential overfitting to the augmented data.