Utilizing Self-Supervised Learning Features and Adapter Fine-Tuning for Enhancing Speech Emotion Recognition

This paper employs an innovative approach that combines self-supervised learning feature fusion and adapter fine-tuning for a speech emotion recognition task. The results show a 5.57% improvement in accuracy compared to the baseline (Hubert upstream model) in the speech emotion recognition task, reaching the highest level of 75.1%. This progress illustrates the significant improvement of adapter fine-tuning on the migration performance of self-supervised learning features in different tasks. In summary, this study introduces an innovative solution to address the discrepancy between self-supervised learning features and downstream tasks, offering valuable insights for enhancing performance in speech emotion recognition. The integration of self-supervised learning with adapter fine-tuning not only holds promise for achieving comparable success across diverse speech-related tasks but also contributes fresh perspectives to the advancement of speech processing.