Speech Emotion Recognition Based on Feature Fusion and Residual Graph Convolutional Network

Applying Graph Convolutional Networks (GCN) to Speech Emotion Recognition (SER) tasks is a novel idea, but the selection of effective features and deepening of the GCN layers remain unresolved issues. To address this, this paper uses two acoustic features as inputs and improves the model structure. Firstly, AlexNet that integrates Convolutional Block Attention Module (CBAM) is used to extract log Mel-spectrogram features, and the CBAM attention output is concatenated with Mel-frequency cepstral coefficients (MFCC). Then, the concatenated features are input into a 4-layer residual graph convolutional network and followed by a fully connected classification layer. Experimental results on the CASIA, RAVDESS, and IEMOCAP datasets demonstrate the superiority of the proposed method with weighted accuracies of 92.2%, 79.8%, and 70.1 % respectively.