DualDistill: Self and cross-modal knowledge distillation for multimodal emotion recognition in conversation

Emotion Recognition in Conversation (ERC) has become increasingly critical in diverse applications, including health care and virtual assistants. This paper proposes DualDistill, a multimodal ERC model. The proposed model employs a self-distillation strategy based on the Exponential Moving Average (EMA) to incorporate soft-label signals and enhance text representations, providing rich emotional cues. In addition, cross-modal Knowledge Distillation (KD) is applied to transfer contextual emotional cues from text to non-verbal modalities (audio and visual), alleviating modality imbalance and improving multimodal fusion. The DualDistill method achieves state-of-the-art performance on IEMOCAP and MELD benchmarks, demonstrating robustness and strong generalizability.