Speech Emotion Recognition Using U-Net

Recognizing emotions from speech is of great significance in enhancing human-machine interaction. Convolutional neural networks (CNN) continuously compress size and stack weighted values when capturing temporal dynamic features, resulting in the loss of important dynamic features in different channels and depths during the extraction process, which reduces recognition accuracy. To address this issue, the U-Net architecture is employed in this study for speech emotion recognition, and an extended version of the U-Net structure is proposed. The specific method involves extracting the temporal dynamic features of the audio signal through rectangular convolution to generate the Mel-spectrogram dynamic feature map. Then, the U-Net architecture is utilized to establish connections between feature maps of varying scales, while channel selection attention is employed to assess dissimilarities among dynamic features across different channels. Experimental findings on the combined CER dataset reveal that the enhanced U-Net effectively filters essential temporal dynamic features, resulting in a 4.29 percentage point improvement in recognition accuracy compared to the baseline model.