Mask, Sample, Revise: A Revisable CTMC Inference Stack for Guided Discrete Flow Matching Text-to-Speech

2026-06-11

AI Summary

This work addresses the challenge of generating stable and controllable discrete speech with unaligned non-autoregressive text-to-speech (TTS) models under low-step inference. To this end, the authors propose a reversible inference framework based on continuous-time Markov chains (CTMCs), integrating predictor-agnostic guidance, acoustic prompt-aligned coupling, and a schedule-constrained SC-ReMask re-masking mechanism. This approach enables efficient and controllable discrete flow-matching speech synthesis within a single tau-leaping sampling step. As the first study to introduce revisable re-masking and prompt-matching coupling in discrete flow-matching TTS, the proposed method significantly improves intelligibility and stability under low NFE (number of function evaluations) conditions without requiring fine-tuning, substantially outperforming existing unguided or solely guided sampling strategies.

Abstract

Recent alignment-free non-autoregressive (NAR) text-to-speech (TTS) models formulate synthesis as a conditional infilling task, bypassing explicit duration predictors and external aligners. When speech is represented with neural codec tokens, the infilling problem becomes discrete, making Discrete Flow Matching (DFM), a Continuous-Time Markov Chain (CTMC) framework for discrete generation, a natural fit. However, inference-time control for stable low-step conditional infilling remains underexplored. We propose Mask, Sample, Revise, an inference-time CTMC stack for alignment-free DFM-TTS. The stack combines predictor-free guidance to strengthen text conditioning, prompt-matched conditional coupling to align the probability path with the acoustic prompt, and SC-ReMask, a schedule-constrained remasking mechanism that introduces token-to-mask transitions so early de-masking decisions can be revised. These components require no post-hoc fine-tuning and operate in a single tau-leaping sampler. Controlled ablations show that this stack improves intelligibility and robustness in the low-NFE prompted setting, outperforming unguided and guidance-only samplers with substantially more steps.

Problem

• Discrete Flow Matching
• Text-to-Speech
• Conditional Infilling
• CTMC
• Non-autoregressive

Innovation

• Discrete Flow Matching
• Continuous-Time Markov Chain
• non-autoregressive TTS
• conditional infilling
• SC-ReMask