Infant Speech Emotion Recognition Based on Channel Attention Mechanism with ResNet-BiLSTM

Infants communicate their emotional needs through speech, which contains rich emotional information. In view of the fact that there is no publicly available infant speech emotion data set and the accuracy of existing infant speech emotion recognition methods is poor, in this paper, we collected 6100 infant speech samples from the Internet through crawling and other methods as the infant speech emotion recognition data set, and proposed a ResNet-BiLSTM model integrating channel attention mechanism for infant speech emotion recognition. The model first uses MFCC with dynamic and static feature fusion as input, then uses ResNet network embedded in SENet module to extract channel information of input features, and then sends it to BiLSTM model to extract speech context information. Finally, a fully connected layer is used for classification to obtain the final recognition result. The experimental results show that the proposed method can achieve an accuracy of 85% on the infant speech emotion dataset, which is 9% higher than the ordinary ResNet.