Yuhang Jia

Explicit acoustic guidance in H-SAGE leads to significant performance improvements in multi-talker ASR, especially in challenging overlapping speech conditions.

Training for speech editing with reinforcement learning not only enhances editing quality but also unexpectedly boosts zero-shot TTS performance.

LLM post-training isn't just about objectives; it's about strategically intervening on model behavior through support expansion, policy reshaping, and behavioral consolidation.