MARSE uses continuous DAC codec representations for speech enhancement

MARSE performs speech enhancement by iteratively decoding masked clean speech frames using continuous latent representations from the DAC neural audio codec. Utilizing a Conformer-based architecture, the method allows for adjustable trade-offs between speech quality and computational overhead compared to discrete token approaches.