Preprint
Towards Robust Dysarthric Speech Recognition: LLM-Agent Post-ASR Correction Beyond WER
Xiuwen Zheng et al., arXiv (ICASSP 2026), 2026
Read the originalRead on October 17, 2026
- What it is
- A 2026 study of LLM correction for speech recognition on dysarthric speech, the speech of people with motor speech disorders. It splits its test set into utterances the recognizer got right and utterances it got wrong, and reports how a zero-shot LLM and a fine-tuned LLM change the error rate of each group.
- How much weight it carries
- A preprint of an ICASSP 2026 paper. The test set is dysarthric speech, not typical dictation, and the LLMs are 7 to 8 billion parameter models, so the numbers show the pattern (correction helps the erroneous group and can hurt the rest) rather than dictation performance.
Facts taken from this source
- On a dysarthric speech test set, the recognizer alone had an overall word error rate of 13.63 percent: 0 percent on sentences it transcribed without error, and 21.98 percent on sentences that had errors.
“The ASR baseline achieves an overall WER of 13.63% (0% on the NoErr group and 21.98% on the Err group).” (Results, Table 1)
- After supervised fine-tuning, the word error rate on the erroneous sentences fell from 22 to 19 percent, and the authors report this across all the models they tested.
“across all models, SFT consistently reduces WER on the Err group (from 22% to 19%)” (Results, discussion of Table 1)