Fact · Voice and AI
Pu and colleagues say that when an LLM corrects speech recognition output, it makes many unnecessary changes to the input transcription, which they call over-correction.
Checked against its source on October 17, 2026
In the source's own words
“LLMs make many unnecessary changes to the input transcription”
Multi-stage Large Language Model Correction for Speech Recognition, Jie Pu et al., arXiv, 2023, Introduction.
Open the original- What it measures
- The authors explain that the LLM steers sentences toward written language rather than the verbatim words that were spoken. Their pipeline responds by sending only transcripts that a confidence estimate flags as uncertain to the LLM.
- What it does not show
- A 2023 preprint about English benchmarks and one correction pipeline. It describes a failure mode that can appear with any LLM correction step, and it does not measure how often it happens in dictation.