Abstract: ‍ ‍Emotion recognition plays a crucial role in modern human-computer interaction， affective computing， and immersive virtual reality by enabling computers to automatically identify and classify human emotional states. With advancements in multimodal learning， emotion recognition has progressed from traditional single-modal approaches to more complex multimodal approaches. Multimodal emotion recognition involves processing data from various sources， such as text， speech， and visual information. These different modalities capture distinct emotional features， allowing models to better understand human emotional states. However， significant challenges remain in multimodal emotion recognition， including temporal misalignment and modality heterogeneity. To address these challenges， this paper presents a multimodal emotion recognition model that utilizes dual alignment and contrastive learning. The proposed model integrates text， speech， and visual modalities， achieving comprehensive alignment through a dual alignment module. Specifically， feature-level alignment employs a cross-attention mechanism to create dynamic temporal alignment， while sample-level alignment uses contrastive learning to align the modalities in feature space. Additionally， the model introduces supervised contrastive learning to leverage label information， allowing for the extraction of fine-grained emotional cues and enhancing the model’s robustness. Self-attention mechanisms are also applied to learn interactions among multimodal features， effectively improving overall model performance. Experimental results show that the proposed model performs exceptionally well on three public datasets， outperforming existing models on most metrics. In conclusion， the dual alignment and contrastive learning-based multimodal emotion recognition model successfully addresses key challenges in the field and achieves significant performance improvements.