Aims: Multimodal emotion recognition (MER) can outperform unimodal approaches by integrating complementary information from multiple sources. However, real-world applications often involve incomplete or missing modalities, limiting the reliability of existing MER models. This study aims to develop a framework that remains robust under missing modality conditions while preserving the benefits of multimodal integration.

Methods: To address this challenge, we propose a cross-modal latent regeneration and attention long short-term memory (CMLR-ALSTM) framework. The framework integrates pretrained variational autoencoder encoders with residual projection networks trained with L2 loss to achieve stable, effective latent-space alignment across modalities. The regenerated latent features, along with the original available ones, are then integrated through a cross-modal attention mechanism and passed to an long short-term memory (LSTM) network to capture temporal dependencies and enhance multimodal fusion under incomplete data conditions.

Results: We conducted two sets of experiments. First, we evaluated the proposed CMLR-ALSTM framework on three benchmark datasets under the complete-modality scenario, where all modalities were available, and compared its performance against state-of-the-art methods and baseline models. Second, to assess model robustness, we evaluated CMLR-ALSTM under missing and partially missing modality scenarios. Experimental results on the DEAP, MAHNOB-HCI, and SEED-IV datasets demonstrate that CMLR-ALSTM achieves up to 17.22% improvement under missing-modality conditions while maintaining competitive performance with state-of-the-art methods, highlighting its ability to preserve cross-modal relationships and maintain robust latent representations.

Conclusion: The experimental results confirm the effectiveness of the proposed CMLR-ALSTM framework for MER, particularly in realistic environments where data availability is inconsistent. By leveraging CMLR to reconstruct missing representations and modelling temporal dependencies through LSTM, the proposed approach provides a more robust and reliable MER framework for practical deployment. In addition, results across different combinations of modalities indicate that the proposed framework generalized well across heterogeneous multimodal settings.