Speech Emotion Recognition based on Multi-Level Residual Convolutional Neural Networks

Speech emotion recognition, using the convolutional neural networks (CNN) model, is challenging due to the problem of features loss and the decrease of recognition accuracy. To address this issue, a Multi-level residual CNN model is proposed is this paper. In this model, the speech signals are converted into spectrogram, then the multi-level residual identity maps are introduced to compensate the missing features in the CNN during the convolution process, so as to improve the recognition accuracy of speech emotion. The research results show that the Multi-level residual CNN can achieve 74.36% recognition accuracy on the EMO-DB dataset, which has better performance than traditional deep CNN method.