Dual-TBNet: Improving the Robustness of Speech Features via Dual-Transformer-BiLSTM for Speech Emotion Recognition

Speech emotion recognition has always been one of the topics that have attracted a lot of attention from many researchers. In traditional feature fusion methods, the speech features used only come from the data set, and the weak robustness of features can easily lead to overfitting of the model. In addition, these methods often use simple concatenation to fuse features, which will cause the loss of speech information. In this paper, to solve the above problems and improve the recognition accuracy, we utilize self-supervised learning to enhance the robustness of speech features and propose a feature fusion model(Dual-TBNet) that consists of two 1D convolutional layers, two Transformer modules and two bidirectional long short-term memory (BiLSTM) modules. Our model uses 1D convolution to take features of different segment lengths and dimension sizes as input, uses the attention mechanism to capture the correspondence between the two features, and uses the bidirectional time series module to enhance the contextual information of the fused features. We designed a total of four fusion models to fuse five pre-trained features and acoustic features. In the comparison experiments, the Dual-TBNet model achieved a recognition accuracy and F1 score of 95.7% and 95.8% on the CASIA dataset, 66.7% and 65.6% on the eNTERFACE05 dataset, 64.8% and 64.9% on the IEMOCAP dataset, 84.1% and 84.3% on the EMO-DB dataset and 83.3% and 82.1% on the SAVEE dataset. The Dual-TBNet model effectively fuses acoustic features of different lengths and dimensions with pre-trained features, enhancing the robustness of the features, and achieved the best performance.