B-GRPO: Unsupervised Speech Emotion Recognition based on Batched-Group Relative Policy Optimization

Researchers have developed a new reinforcement learning technique called B-GRPO for unsupervised speech emotion recognition. It treats the selection of training samples as a long-term decision process, using self-reward and teacher-reward functions to guide learning. This approach avoids the need for costly, biased human annotations. Experiments show the method improves performance by 19.8% over a baseline system that does not use reinforcement learning.