Masked diffusion models (MDMs) have emerged as a competitive approach for zero-shot text-to-speech, offering an alternative to autoregressive (AR) models. Unlike AR models, which generate speech tokens sequentially, MDMs define a corruption process that maps clean speech-token sequences into noisy versions, then learn to reverse that process. The new arXiv preprint, Revise-TTS, proposes a refinement framework built on this foundation.
The key contribution is a non-masking refinement strategy. Rather than relying on the conventional masking-based corruption, Revise-TTS introduces a refinement step that aims to improve generation quality. The abstract does not provide full details of the corruption process or experimental results, but the methodological shift is clear: moving from masking to a non-masking approach within the MDM framework.
Because this is a single preprint, the claims are not yet peer-reviewed, and no comparisons to other systems are available in the source. Still, the work points to an active area of research: making diffusion-based TTS more efficient and higher quality without the sequential constraints of autoregressive models.