How can I implement reinforcement learning for more LLM-based TTS models?

[ZongWei-HUST]

Great work! If I want to implement reinforcement learning experiments for more LLM-based TTS models, is there a workflow I can reference for code construction?

[yuekaizhang]

@ZongWei-HUST The cosyvoice_rl recipe is based on the VeRL framework. To support other LLM-based TTS systems, you can directly use the reward server in the recipe. For other components like rollout and log_prob computation, you can either adapt your model to be compatible with the VeRL framework or write your own code.
I'm currently developing custom code to support TTS RL training without using the VeRL repository, but it may take a month to complete.