arXiv cs.CL· Shashi Kumar, Yanis Labrak, Hasindri Watawana, Sergio Burdisso, Esa\'u Villatoro-Tello, Kadri Hacio\u{g}lu, Petr Motlicek, Andreas Stolcke·· 12 小时前AI 评分31
合成语音是唯一资源时，GRPO 比 SFT 更有效
When Synthetic Speech Is All You Have: Better Call GRPO
AI 导读
在银行等受监管领域，用合成语音适配 LLM 语音识别模型时，GRPO 比 SFT 效果更好：纯合成语音下 GRPO 将 WER 相对降低 40%（36.71%→22.09%），SFT 后再接 GRPO 可进一步降低 45%。增益来自行为而非表征，GRPO 通过改善停止校准减少插入错误，并让注意力更好地锚定音频以提升语音-文本对齐，早期层表征保持不变。
正文 · 原文
Abstract:LLM-based ASR adapted to regulated domains such as banking is bottlenecked by privacy: real speech is costly and legally constrained to collect, making synthetic text-to-speech (TTS) an attractive substitute. Yet synthetic speech stays acoustically mismatched with real recordings, and work on this gap has stayed within supervised fine-tuning (SFT). We instead turn to reinforcement learning, and show that Group Relative Policy Optimization (GRPO) extracts far more from the same synthetic speech than SFT. Synthetic-only adaptation of the model with GRPO, a critic-free method rewarding low-WER hypotheses, reduces WER by 40% relative to SFT (36.71%$\to$22.09%), and an SFT-then-GRPO combination pushes this further to 45%. We trace the gain to behavior rather than representation: GRPO reduces insertion errors by improving stopping calibration and speech-to-text alignment by better anchoring attention to audio, leaving early-layer representations intact. When synthetic speech is the main resource, reinforcement learning should be preferred over supervised fine-tuning.
| Comments: | 7 pages; accepted to IEEE SLT 2026 |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2607.08409 [cs.CL] |
| (or arXiv:2607.08409v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2607.08409 arXiv-issued DOI via DataCite |
Submission history
From: Shashi Kumar [view email]
[v1]
Thu, 9 Jul 2026 12:34:56 UTC (448 KB)
[v2]
Tue, 20 Oct 2026 16:29:26 UTC (439 KB)
来源：arXiv cs.CL · arxiv.org