In the process of human-computer emotional interaction, the accuracy of emotion recognition and real-time feedback directly affect the quality of interaction. The emotion recognition and intelligent feedback system designed based on multimodal deep learning integrates speech, expression and text features, and adopts the improved Transformer architecture to realise emotion state recognition, achieving an accuracy rate of 92.8%. Combined with the reinforcement learning algorithm to construct an intelligent feedback mechanism, experiments show that the system can effectively improve the user's emotion improvement rate, with good practicality and promotion value.