Infant Cry Classification using Transfer Learning

Infants use cry as their main tool for communication as they cannot speak. It is imperative to understand that not being able to comprehend these cries can affect the health of an infant. Parents or caretakers generally find it very difficult to interpret these cries and understand what the infant is feeling. Many approaches have been explored to tackle the problem of infant cry classification using different datasets. The intrinsic obstacle with infant cry classification is that the datasets available are small and in real-time prone to noise as the audio samples are collected in different environments. Also, to the best of our knowledge, there is no real-time system deployed for infant cry classification. In this view, this paper aims at addressing these issues and design a real-time embedded system running a deep learning model based on the Transfer Learning approach that classifies the infant cry. The Dunstan Baby Language dataset is used for this research. Various features for cry classification and potential approaches such as CNN, CNN+LSTM, Hybrid Mixed Deep Learning model and models based on Transfer Learning were considered. The best performance was attained by using the finetuned VGG16 model with an accuracy of 0.92 and F1 score of 0.92.