When Synthetic Speech Is All You Have: Better Call GRPO
Group Relative Policy Optimization (GRPO) improves LLM-based Automatic Speech Recognition when using synthetic speech as the primary adaptation resource, outperforming supervised fine-tuning (SFT). GRPO optimizes a sequence-level objective over sampled hypotheses, making it robust to localized acoustic artifacts in synthetic audio. This method achieves significantly better performance on synthetic data alone compared to SFT.