What it is
On-policy RL for flow-matching TTS that reweights training by group-relative rewards. It improves WER and speaker similarity on DiTAR without hurting UTMOS. Fine-tunes autoregressive-diffusion TTS with on-policy rewards for intelligibility and speaker similarity, directly reweighting the flow-matching loss and anchoring to a frozen pretrained model.
Key results
- Average WER drops from 2.016 to 1.558 across 3 test sets
- Speaker similarity rises from 0.676 to 0.715 while keeping UTMOS
- Uses 10-NFE rollouts and trains 2.9× faster than 32-NFE DiTAR-GRPO
Shows that flow-matching TTS can be optimized directly with on-policy reward weighting, without SDE conversion or likelihood-ratio tracking.