Mask, Sample, Revise: A Revisable CTMC Inference Stack for Guided Discrete Flow Matching Text-to-Speech
Researchers have developed a new inference stack for text-to-speech models that utilizes discrete flow matching. This approach formulates speech synthesis as a conditional infilling task, bypassing the need for explicit duration predictors and external aligners. The proposed "Mask, Sample, Revise" stack enhances text conditioning, aligns acoustic prompts, and allows for revision of early de-masking decisions, leading to improved intelligibility and robustness, especially in low-step settings. AI
IMPACT This research could lead to more natural and robust text-to-speech systems by improving conditional infilling and allowing for revision of synthesis steps.