Chinese Speaking Emotion Recognition and Emotion Synthesis based on BERT Model

People pay more and more attention to emotion recognition and synthesis in spoken Chinese. Aiming at emotional recognition of spoken Chinese, this paper uses BERT model to extract emotional features from text data, and fine-tunes the model to adapt to emotional goals. Through comparative experiments, the performance of BERT model is evaluated from three aspects: rated naturalness, stability of emotion recognition and macro average accuracy. This paper also compares the performance of BERT model with Word2Vec model and Lite BERT (ALBERT) model. Experiments show that the naturalness of Bert model is 80. In the aspect of stable emotion recognition, the performance of BERT model has been significantly improved, with less fluctuation in accuracy and the best robustness.