From Cries to Answers: A Comprehensive CNN+DNN Hybrid Model for Infant Cry Classification with Enhanced Data Augmentation and Feature Extraction

Infant cry classification involves utilizing machine learning techniques to automatically recognize and categorize various types of infant cries. The goal of infant cry classification using deep learning is to develop a robust and accurate system capable of automatically recognizing and categorizing the different types of infant cries. This technology can potentially assist healthcare professionals or caregivers in identifying the underlying reasons for the crying, such as hunger, discomfort, pain, etc., aiding in prompt and appropriate responses to the infant's needs. This paper presents an innovative approach to infant cry classification using deep learning technique, data augmentation and feature extraction methods. Our primary model is based on popular infant cry classification dataset named Donateacry and Infants Cry Sounds corpus. We use data augmentation techniques like noise, stretch, shift, pitch, time masking and frequency masking to enhance generalizability of our model. Then we use SMOTE to balance all the classes of dataset. Furthermore, we use feature extraction techniques like Chroma STFT, Root Mean Square, Mel-frequency Cepstral Coefficients (MFCC), and Zero Crossing Rate to capture distinctive patterns in children's cry signals. To validate our model, we employed k-fold cross-validation, yielding a mean accuracy of 0.9384 and a mean loss of 0.0652.