A CNN-Based Method for Infant Cry Detection and Recognition

Abstract

Crying is the primary means of communication between the baby and the outside world. When a baby is crying, it is difficult for a novice parent to immediately understand the baby’s needs. If parents can accurately determine the cause of the baby’s cry, they can understand the baby’s emotional and physiological changes and needs. In real-world applications, recording devices may record sounds that are not produced by a baby. To reduce the burden on the recognition server and improve the accuracy of the classifier, this study proposes the conversion of the baby’s crying signal into a two-dimensional spectrogram. A convolutional neural network is used to determine if the input spectrum represents a baby’s cry. A baby’s cry is ultimately divided into four categories (including pain, hunger, sleepiness, and wet diaper) through additional one-dimensional convolutional neural networks. Experimental results showed that the proposed method achieves high crying detection and recognition rates.