Speech RL

Apply reinforcement learning to fine-tune speech generation models for objective quality metrics, naturalness, and expressiveness beyond supervised training. Speech RL applies policy gradient methods to speech generation models, using reward models trained on human judgments of speech quality or objective metrics (MOS prediction, intelligibility scores). The speech generator is treated as a policy that produces audio, which is scored by the reward model. This allows optimization for aspects of speech quality that are hard to capture with supervised loss: natural prosody, expressiveness, listener preference.