CosyEdit2: Speech-Editing-Oriented Reinforcement Learning Unlocks Better Zero-Shot TTS

Researchers introduced CosyEdit2, a two-stage post-training method for speech editing.

It first uses supervised editing training, then applies editing-oriented GRPO on target-speech-free data.

The approach improves speech editing performance and also boosts zero-shot TTS quality.

The results suggest speech editing optimization and zero-shot TTS are more closely linked than previously thought.