Reproduce “F5R-TTS: Improving Flow-Matching based Text-to-Speech with Group Relative Policy Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaohui Sun, Ruitong Xiao, Jianye Mo, Bowen Wu, Qun Yu, Baoxun Wang
F5R-TTS integrates Group Relative Policy Optimization (GRPO) into flow-matching based non-autoregressive (NAR) text-to-speech models to enhance zero-shot voice cloning. The approach achieved a 29.5% relative reduction in Word Error Rate (WER) and a 4.6% relative increase in speaker similarity (SIM) on general test sets, demonstrating improved intelligibility and voice consistency.
Key insights
- Transforming the deterministic output of flow-matching models into probabilistic Gaussian distributions effectively enables their integration with reinforcement learning algorithms.
- Group Relative Policy Optimization (GRPO) can concurrently enhance speech intelligibility and speaker similarity in NAR TTS by employing a dual reward system based on WER and SIM.
- RL-driven fine-tuning provides a mechanism to improve the robustness of NAR TTS systems, particularly in noisy conditions and with complex textual inputs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.