Research questionHow can speech enhancement use continuous audio representations while balancing quality against decoding cost?Masked generative speech enhancement reconstructs clean speech from noisy inputs, while the representation and decoding process affect both enhancement quality and inference cost.