A DEEP LEARNING BASED APPROACH WITH DATA AUGMENTATION FOR INFANT CRY SOUND VERIFICATION

The way newborn infants communicate is by crying. There have been studies aimed at deriving specific analyses from an infant's cry. Identification of the infant from the cry is one such research area. In this work, the CryCeleb2023 dataset is used to perform experiments. To identify the crying infant, the state-of-the-art deep learning system ECAPA-TDNN, originally designed for adult speaker identification, is used. One of the major constraints in working with infant cry identification is the scarcity of data. To address this prob-lem, we have performed time-spectral domain augmentation, speed perturbation, and noise addition techniques to increase the dataset size. The experimental results show a significant reduction in the Equal Error Rate (EER) for both development and test sets, demonstrating the model's strong potential for infant cry identification.