Reproduce “Speech Recognition With LLMs Adapted to Disordered Speech Using Reinforcement Learning” 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
Chirag Nagpal, Subhashini Venugopalan, Jimmy Tobin, Marilyn Ladewig, Katherine Heller, Katrin Tomanek
The authors adapt a Gemma 2B recognizer for English disordered speech using RL rewards for word accuracy and meaning preservation. Continued supervised fine-tuning on disordered speech alone performs worse than the mixed-data starting model. On Euphonia test, RL reports WER 42.6 and MP 55.7, versus 57.1 and 42.8 for continued fine-tuning. Experiments are limited to Gemma 2B and English.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.