Chinese Speech Emotion Recognition Based on Pre-Trained Model

This paper proposes a method based on large-scale speech pre-trained models for the task of Chinese speech emotion recognition. Similar to the transfer learning approach in image classification task, speech pre-trained model is used as a fixed feature extractor to extract features from speech data, followed by a fully connected network for classification. Experimental results show that, on the seven-class speech emotion recognition task of the Chinese M3ED dataset, the proposed method achieves classification accuracies of 0.506 and 0.560 using base and large sized pre-trained models respectively, surpassing the baseline accuracy of 0.454 for the ECAPA-TDNN model.