aiqbench analysis
This research addresses the training-inference mismatch in LLM-based speech translation by introducing joint recognition and translation fine-tuning via Group Relative Policy Optimization (GRPO). The study demonstrates that this reinforcement learning approach significantly improves BLEU scores and reduces word error rates compared to standard supervised fine-tuning methods.
Transcribe, Translate, and Optimize: Joint Reward Learning for Speech Translation
arXiv:2609.26536v1 Announce Type: new Abstract: In LLM-based speech translation, transcription-based chain-of-thought (CoT) suffers from a mismatch between reference transcripts used in supervised fine-tuning (SFT) and model-generated transcripts at inference. To address this, we propose joint recognition and translation fine-tuning via group relative policy optimization (GRPO). We score both transcripts and translations, with translation conditioned on model-generated transcripts, and compare three token advantage strategies. Using Qwen2.5-Omni-3B across four languages, we evaluate CoT against direct speech translation (Direct ST) under SFT and GRPO, training on CoVoST 2 and testing on CoVoST 2 and FLEURS. CoT GRPO outperforms Direct ST GRPO by 1.77 and 0.83 average BLEU points on CoVoST 2 and FLEURS. Compared to CoT SFT, GRPO boosts BLEU by 0.82 and 0.67 points and reduces word error rate (WER) by 8.8% and 7.2% relatively. These results highlight reinforcement fine-tuning as an effective method to mitigate the training-inference mismatch, jointly improving recognition and translation.
Abstract:In LLM-based speech translation, transcription-based chain-of-thought (CoT) suffers from a mismatch between reference transcripts used in supervised fine-tuning (SFT) and model-generated transcripts at inference. To address this, we propose joint recognition and translation fine-tuning via group relative policy optimization (GRPO). We score both transcripts and translations, with translation conditioned on model-generated transcripts, and compare three token advantage strategies. Using Qwen2.5-Omni-3B across four languages, we evaluate CoT against direct speech translation (Direct ST) under SFT and GRPO, training on CoVoST 2 and testing on CoVoST 2 and FLEURS. CoT GRPO outperforms Direct ST GRPO by 1.77 and 0.83 average BLEU points on CoVoST 2 and FLEURS. Compared to CoT SFT, GRPO boosts BLEU by 0.82 and 0.67 points and reduces word error rate (WER) by 8.8% and 7.2% relatively. These results highlight reinforcement fine-tuning as an effective method to mitigate the training-inference mismatch, jointly improving recognition and translation.
Computation and Language (cs.CL); Audio and Speech Processing (eess.AS)
arXiv:2609.26536 [cs.CL]
(or arXiv:2609.26536v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.26536 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 247 words
Related coverage
Putting sign language AI into users’ hands
Google DeepMind · Aug 12, 2026
Microsoft’s AI chief says superintelligence is near, but won’t take your job
The Verge · Jun 8, 2026
DuraS2ST: Chain-of-Thought and Reinforcement Learning for Duration-Aligned Speech-to-Speech Translation
arXiv · Sep 30, 2026
Reflections on unlearning and inoculation
LessWrong · Sep 20, 2026