EmoMamba: Real-time emotional state recognition and adaptive feedback generation in auditory learning environments

Emotion recognition and feedback play a critical role in enhancing user engagement and emotional learning experiences. However, existing models often struggle to accurately recognize complex emotions and provide adaptive feedback, particularly in real-time scenarios. This paper proposes EmoMamba, a deep learning model that integrates Bi-Mamba for bidirectional temporal feature extraction, GAT for emotional association reasoning, and DQN for dynamic feedback optimization. The model aims to address challenges in emotion recognition and feedback accuracy, focusing on real-time emotional state adaptation. Experiments on the RAVDESS and TESS datasets demonstrate that EmoMamba outperforms traditional methods and state-of-the-art models in terms of accuracy (94.2% on RAVDESS, 93.8% on TESS) and Macro-F1 (93.1% on RAVDESS, 92.4% on TESS). Additionally, it excels in feedback effectiveness, achieving high user satisfaction and emotion improvement degree (EID of 0.87 on RAVDESS and 0.85 on TESS). The proposed model provides a robust solution for dynamic emotional interaction, offering significant potential for applications in personalized learning and emotion-aware systems.