Fact · Voice and AI
19percent word error rate
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.
Checked against its source on October 17, 2026
In the source's own words
“across all models, SFT consistently reduces WER on the Err group (from 22% to 19%)”
Towards Robust Dysarthric Speech Recognition: LLM-Agent Post-ASR Correction Beyond WER, Xiuwen Zheng et al., arXiv (ICASSP 2026), 2026, Results, discussion of Table 1.
Open the original- What it measures
- SFT is supervised fine-tuning: the language model is trained on labeled pairs of recognizer output and the correct text, and it is then used to correct new output.
- What it does not show
- The gain is on the sentences that already had errors. The same paper reports that the group without errors stayed stable after fine-tuning, with a small loss. The method needs labeled pairs that most dictation apps do not have.