Kinetic-Optimal Scheduling with Moment Correction for Metric-Induced Discrete Flow Matching in Zero-Shot Text-to-Speech

Researchers introduce GibbsTTS, a new zero-shot text-to-speech system using metric-induced discrete flow matching with kinetic-optimal scheduling and moment correction. The method achieves superior naturalness and speaker similarity compared to existing masked generative models and state-of-the-art TTS systems without requiring hyperparameter tuning.

GibbsTTS represents a significant advancement in text-to-speech synthesis by solving two fundamental computational problems that have limited practical deployment of discrete flow matching models. The researchers derived a kinetic-optimal scheduler that eliminates the need for manual hyperparameter search, replacing heuristic approaches with a theoretically grounded solution that traverses probability paths at constant Fisher-Rao speed. This training-free numerical schedule reduces complexity and improves reproducibility across implementations.

The technical innovation addresses path-tracking errors inherent in finite-step solvers for continuous-time Markov chains by introducing moment correction that adjusts jump probabilities while preserving the destination distribution. This dual contribution—theoretical optimization combined with practical error correction—represents meaningful progress in discrete generative modeling.

The experimental validation demonstrates tangible improvements in both objective and subjective metrics. GibbsTTS achieves the best naturalness scores against masked discrete generative baselines under unified architectural and dataset conditions, with human evaluators showing clear preference. Performance on speaker similarity metrics proves competitive with state-of-the-art systems, ranking first on three of four test datasets. These results validate that the theoretical improvements translate into practical quality gains.

For the broader AI and speech synthesis community, this work demonstrates how principled mathematical approaches can improve generative models without increasing computational burden. The method's training-free nature and strong empirical results position it as a viable alternative for applications requiring high-quality zero-shot speech synthesis, particularly in scenarios where computational efficiency and reproducibility are priorities.

• →GibbsTTS eliminates hyperparameter search through kinetic-optimal scheduling derived from Fisher-Rao geometry
• →Moment correction technique reduces finite-step path-tracking errors while maintaining CTMC jump distributions
• →Achieves superior objective naturalness compared to masked discrete generative baselines in controlled experiments
• →Demonstrates state-of-the-art speaker similarity performance, ranking first on three of four test datasets
• →Training-free numerical approach reduces complexity and improves reproducibility for discrete flow matching models