MEGA Hub

GROW: Group-Relative Advantage-Weighted On-Policy Reinforcement Learning of Autoregressive-Diffusion Text-to-Speech model

Authors

Do you know Guanrou Yang?You can claim authorship or link another user.Do you know Tian Tan?You can claim authorship or link another user.Do you know Qian Chen?You can claim authorship or link another user.Do you know Ziyang Ma?You can claim authorship or link another user.Do you know Yakun Song?You can claim authorship or link another user.Do you know Zhikang Niu?You can claim authorship or link another user.Do you know Qi Chen?You can claim authorship or link another user.Do you know Wenming Tu?You can claim authorship or link another user.Do you know Haitao Li?You can claim authorship or link another user.Do you know Shan Yang?You can claim authorship or link another user.Do you know Xie Chen?You can claim authorship or link another user.

Abstract

Reinforcement learning for flow-matching text-to-speech is complicated by deterministic ODE sampling: trajectory-level policy-gradient methods typically convert the ODE into an SDE and track per-step likelihood ratios, introducing stochastic perturbations and substantial overhead. We propose GROW, a group-relative advantage-weighted on-policy RL method that acts directly on the standard flow-matching objective. For each prompt, GROW samples a group of on-policy utterances, separately standardizes intelligibility and speaker-similarity rewards within the group, and combines them to reweight flow-matching regression. A Wasserstein-2 velocity penalty anchors the updated model to a frozen pretrained reference. A group-mean reward baseline is introduced to convert reward weighting into advantage weighting. For strong pretrained TTS models with concentrated rewards, positive exponential weighting is dominated by reward-agnostic self-imitation, whereas a zero-mean signed advantage preserves effective within-group credit assignment. Instantiated on DiTAR and evaluated on LibriSpeech and Seed-TTS EN/ZH, GROW reduces average WER from 2.016 to 1.558 and raises speaker similarity from 0.676 to 0.715 while keeping UTMOS. With 10-NFE training rollouts and 32-NFE evaluation, GROW retains comparable performance while training 2.9x faster than 32-NFE DiTAR-GRPO. We will open-source complete GROW codes, faithful DiTAR reproduction, and all model checkpoints.

Community

00