Tingyi Lin, Prof. Wen-Ren Yang and Kuanwei Chen posted “Emotion as a Distribution: Joint Valence-Arousal Probability Learning for Speaker-Independent Multimodal Emotion Recognition” on arXiv (arXiv:2609.05755); it has been submitted to Speech Communication.
Instead of assigning one emotion label to an utterance, the model predicts a joint probability distribution over valence and arousal from text and audio. The work compares bidirectional Mamba state-space models with Transformer backbones on IEMOCAP. The code is public as EchoMind-Mamba-VA-SER.