Exploring Generative Error Correction for Dysarthric Speech Recognition

TL;DR

This paper introduces a two-stage speech recognition framework that combines advanced models with large language model-based error correction to improve transcription accuracy for dysarthric speech, addressing a significant challenge in accessibility.

Contribution

It presents a novel two-stage approach integrating generative error correction with speech recognition models, specifically tailored for dysarthric speech, and evaluates various configurations for optimal performance.

Findings

Effective improvement in transcription accuracy on structured speech

Challenges remain in single-word recognition accuracy

Insights into the roles of acoustic and linguistic modeling

Abstract

Despite the remarkable progress in end-to-end Automatic Speech Recognition (ASR) engines, accurately transcribing dysarthric speech remains a major challenge. In this work, we proposed a two-stage framework for the Speech Accessibility Project Challenge at INTERSPEECH 2025, which combines cutting-edge speech recognition models with LLM-based generative error correction (GER). We assess different configurations of model scales and training strategies, incorporating specific hypothesis selection to improve transcription accuracy. Experiments on the Speech Accessibility Project dataset demonstrate the strength of our approach on structured and spontaneous speech, while highlighting challenges in single-word recognition. Through comprehensive analysis, we provide insights into the complementary roles of acoustic and linguistic modeling in dysarthric speech recognition