GRPO improves ASR on synthetic speech by 40% over SFT

Researchers demonstrate that Group Relative Policy Optimization (GRPO) significantly outperforms supervised fine-tuning (SFT) when adapting LLM-based automatic speech recognition (ASR) models to regulated domains using only synthetic text-to-speech data. This approach addresses the privacy constraints and acoustic mismatches that typically hinder model performance in sectors like banking.

• Synthetic-only adaptation with GRPO reduces word error rate (WER) by 40% relative to SFT, lowering it from 36.71% to 22.09%.
• Combining SFT with GRPO further improves results, achieving a total WER reduction of 45%.
• The performance gain stems from behavioral improvements rather than representation changes, specifically better stopping calibration and speech-to-text alignment.
• GRPO reduces insertion errors by anchoring attention to audio while leaving early-layer representations intact.

The authors conclude that when synthetic speech is the primary resource for adaptation, reinforcement learning methods like GRPO should be preferred over supervised fine-tuning.