Conceptual
Browse Instruct Curate

Adapting Speech-Input LLMs to Disordered Speech with Reinforcement Learning

A tuning strategy that turns a large language model into a disordered-speech recognizer by swapping low-frequency text tokens for audio tokens, fine-tuning on transcribed speech, then applying reinforcement learning with syntactic- and semantic-accuracy rewards. It shows that RL with custom rewards adapts a speech LLM to a new speaking style substantially better than supervised fine-tuning alone.