Attention with variational autoencoder enabled one-dimensional convolutional memory network for emotion recognition using EEG and speech signals

Abstract Emotion recognition is an emerging task that concentrates on recognizing the person’s diverse emotional states. However, several previous attempts have been made to acknowledge the person’s emotions and effectively classify them with better performance. However, several limitations have occurred during the detection of human emotions, such as high computational power consumption, minimal recognition accuracy, and diverse modality recognition. Therefore, the research proposes the Self-modular Attention with Variational Autoencoder-enabled one-dimensional Convolutional Neural Network and Long Short-Term Memory Network (SMAVA-1DCNN-LSTM) that overcomes the previous approach’s limitations and provides robust detection results with minimal computational complexity. In the research, emotions are recognized by internal physiological signals (EEG) and voice tone. During an evaluation, the significant features of speech and EEG signals are extracted through multiple feature extraction mechanisms that boost the classification accuracy and training process by minimizing overfitting problems. Specifically, the Self-modular Attention (SMA) mechanism aids in fusing the modality-specific informative features to concentrate on important parts of the input multimodal data. Experimental results show that the SMAVA-1DCNN-LSTM method demonstrates exceptional performance for emotion recognition, reporting accuracy of 97.70%, F1-score of 97.79%, sensitivity of 97.06%, MCC of 0.954, precision of 98.54%, and specificity of 98.34% for 90% of training with the EAV dataset.