GROW makes TTS 2.9x faster with better intelligibility and voice similarity

Training text-to-speech models with reinforcement learning is tricky because flow-matching uses deterministic ODE sampling, forcing prior RL methods to convert to stochastic SDEs and track per-step likelihood ratios—adding noise and overhead. A new arXiv paper introduces GROW (Group-Relative Advantage-Weighted On-Policy RL), which operates directly on the standard flow-matching objective. For each prompt, GROW samples a group of utterances, standardizes intelligibility and speaker-similarity rewards within the group, and combines them to reweight regression. A Wasserstein-2 velocity penalty keeps the model close to a frozen pretrained reference, and a group-mean baseline converts reward weighting into advantage weighting—solving the self-imitation problem where positive exponential weighting ignores reward differences.

Instantiated on the DiTAR text-to-speech model and evaluated on LibriSpeech plus Seed-TTS in English and Chinese, GROW delivers clear gains: word error rate drops from 2.016 to 1.558 on average, speaker similarity rises from 0.676 to 0.715, and UTMOS speech quality stays stable. Even with 10-NFE training rollouts and 32-NFE evaluation, GROW achieves comparable performance while training 2.9x faster than DiTAR-GRPO. The authors promise to release complete code, a faithful DiTAR reproduction, and all model checkpoints, making this a practical recipe for high-quality, efficient TTS reinforcement learning.