Audio-Visual Fusion Network Based on Conformer for Multimodal Emotion Recognition

Audio-visual emotion recognition aims to integrate audio and visual information for accurate emotion prediction, which is widely used in real application scenarios. However, most existing methods lack fully exploiting complementary information within modalities to obtain rich feature representations related to emotions. Recently, Transformer and CNN-based models achieve remarkable results in the field of automatic speech recognition. Motivated by this, we propose a novel audiovisual fusion network based on 3D-CNN and Convolution-augmented Transformer (Conformer) for multimodal emotion recognition. Firstly, the 3D-CNN is employed to process face sequences extracted from the video, and the 1D-CNN is used to process MFCC features of audio signals. Secondly, the visual and audio features are fed into a feature fusion module, which contains a set of convolutional layers for extracting local features and the self-attention mechanism for capturing global interactions of multimodal information. Finally, the fused features are input into linear layers to obtain the prediction results. To verify the effectiveness of the proposed method, experiments are performed on RAVDESS and a newly collected dataset named PKU-ER. The experimental results show that the proposed model achieves state-of-the-art performance in audio-only, video-only, and audio-visual fusion experiments.