Abstract: Multimodal Emotion Recognition (MER) aims to predict human emotions by leveraging multiple modalities, including language, vision, acoustic. Existing methods typically perform interaction or fusion directly upon raw features without explicit modeling and filtering of noisy information This paper proposes a Quality-aware Denoising Fusion framework with Modality Balancing (QDFMB). Specifically, the framework first decomposes each modality's representation into informative and noisy components via semantic disentanglement, enabling quality-aware selective absorption and recombination for denoising-based cross-modal fusion. A routing mechanism is introduced to adaptively adjust modality weights during fusion, which contributes to the effective noise suppression and efficient utilization of each modality. The comprehensive experimental results on the CMU-MOSI, CMU-MOSEI, and CH-SIMS datasets demonstrate that QDFMB achieves performance comparable to or better than state-of-the-art methods.