Speech emotion recognition based on SE-ResNet

Emotion recognition plays an important role in the field of human-computer interaction. It can help robots understand human needs more accurately. However, the impact of noise signal and model architecture on accuracy has not been fully explored. In addition, individual datasets are often used for algorithm testing, making it challenging to ensure algorithm generalization. To address these issues, we explore the impact of noise and algorithms on emotion recognition tasks based on two datasets. First, we use SE-ResNet as an emotion recognition network, which guarantees the effectiveness of the algorithm through the attention mechanism and residual structure. Experiments show that SE-ResNet performs better than other classical convolutional neural networks, and it validates the effectiveness of the attention mechanism. Second, we verify that noise can cause the algorithm to lose accuracy by setting up experiments with or without noise. Besides that, we analyze noise’s effect for each emotion class by the confusion matrix. The results show that noise has the most significant impact on the recognition accuracy of natural emotion.