Reproduce “Customizing Speech Recognition Model with Large Language Model Feedback” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shaoshi Ling, Guoli Ye

Researchers at Microsoft Core AI developed a reinforcement learning framework that leverages large language model feedback for unsupervised domain adaptation of automatic speech recognition models. This approach resulted in an average 15% relative reduction in Word Error Rate and a 50% relative reduction in Entity Word Error Rate compared to the baseline ASR.

Key insights

  • LLMs can serve as effective, implicit human preference models, providing high-quality, contextual feedback to guide ASR model adaptation without requiring manual annotation or a separate reward model.
  • Integrating LLM intelligence directly into the ASR model's training loop through RL offers a more fundamental and robust adaptation compared to external post-processing approaches.
  • The quality of the LLM-derived reward signal is more impactful for ASR domain adaptation than the specific choice of RL optimization algorithm among those tested (RAFT, DPO, GRPO).

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.