DCEPNet: Dual-Channel Emotional Perception Network for Speech Emotion Recognition

Speech emotion recognition has been widely used in many applications such as call centres and mental health monitoring. However, speech emotion recognition still faces great challenges due to the diversity of speech features and complexity of emotion, especially the problem of inadequate feature extraction. To enhance the ability to capture emotional features, a dual-channel emotional perception network (DCEPNet) is proposed: (i) For the first channel, a multi-branch time-domain perception (MBT) is proposed to capture key emotional segments in the speech signal. (ii) For the second channel, a multi-window transformer (MWFormer) is proposed to solve the problem of insufficient emotion multi-granularity information extraction. Experimental results demonstrate the proposed model outperforms the state-of-the-art models on the CASIA Chinese dataset.