MAGE: A Coarse-to-Fine Speech Enhancer with Masked Audio Generator

Existing speech enhancement (SE) models face a trade-off: discriminative methods offer efficiency but generalize poorly, while powerful generative approaches are computationally expensive and difficult to train. To address this, we introduce MAGE, a lightweight mask-based generative audio enhancer model. Finetuned from the Qwen2.5-0.5B language model and built upon the BigCodec, MAGE has a small footprint of approximately 200 million parameters. MAGE employs a novel coarse-to-fine scarcity-aware masking strategy alongside an auxiliary corrector to improve learning efficiency and smooth the generation process, respectively. Extensive experiments on various enhancement tasks demonstrate that MAGE sets a new benchmark in enhancement performance while remaining significantly more efficient than previous methods.