Emotion Recognition of Infant Crying Sounds Using Convolutional Recurrent Neural Network with Multi-Scale Joint Attention Mechanism

Abstract

Infant cries contain rich emotional information, and recognizing them can help us better understand and meet the emotional needs of infants. However, there is currently no publicly available dataset for general infant cry emotion recognition research. To address this issue, we first collected infant cry data through the internet and filtered it using the Dan's Infant Language Theory, resulting in the construction of an infant cry emotional dataset. In addition, we propose a novel emotion recognition method for infant cries using a combination of Convolutional Neural Network (CNN), Convolutional Block Attention Module (CBAM), and Bidirectional Long Short-Term Memory (BLSTM) with a Multi-Scale Joint Attention Mechanism. Experimental results show that our proposed model achieves an accuracy of 85.01%, which is 13.39% higher than the CNN model without multi-scale and attention mechanisms, and 9.2% higher than the Multi-Scale CNN model without attention mechanisms.