GRPO reduces WER by 40% on synthetic-only ASR adaptation

Researchers demonstrate that Group Relative Policy Optimization (GRPO) significantly outperforms supervised fine-tuning (SFT) when adapting LLM-based automatic speech recognition models to regulated domains using only synthetic text-to-speech data. This approach addresses the acoustic mismatch between synthetic and real recordings while bypassing privacy constraints associated with collecting real speech.

• 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 performance, achieving a 45% reduction in WER.
• The improvement stems from behavioral changes rather than representation shifts, specifically better stopping calibration and speech-to-text alignment via attention anchoring.
• Early-layer representations remain intact while insertion errors are reduced.

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