An Initial Study on Reconstructing Communicative Intent from ASR Pre-Decision Information in Dysarthric Speech: Toward a Personal Speech Interpreter

This preprint presents an initial study toward a Personal Speech Interpreter (PSI) for people with dysarthria. It investigates whether information related to a speaker’s intended utterance can remain in ASR pre-decision information even when it is lost from the final 1-best output. Using dysarthric Japanese speech and Reazon Japanese Zipformer, the study examines token posterior information and discusses its potential use for downstream intent reconstruction by a large language model (LLM), with particular attention to the safe handling of critical information such as numbers, times, negation, and agreement or disagreement.