Speech Emotion Recognition using Channel Attention Mechanism

In order to improve the accuracy of speech emotion recognition, this paper proposes a speech emotion recognition method based on the channel attention mechanism. Firstly, Mel Frequency Ceptral Coefficient(MFCC), speech spectrograms and spectral envelopes are selected as the initial input features; then, multiple depth network models are used to extract feature maps from different angles in parallel; then, weights are assigned and fused to the feature maps output from each sub-depth network model by the channel attention mechanism; finally, the fused feature maps are used to predict emotion categories. The experimental results on CASIA, Emo-DB, and SAVEE emotion datasets show that the method achieves 88.3 ^% , 85.1%, and 64.5% recognition accuracy, respectively, with better recognition performance compared to recent comparative literature models.