Preprint
Multi-stage Large Language Model Correction for Speech Recognition
Jie Pu et al., arXiv, 2023
Read the originalRead on October 17, 2026
- What it is
- A 2023 paper from Zoom Video Communications that corrects speech recognition output with a large language model in two stages. It names over-correction, where the model changes words that were already right, as the main obstacle to using LLMs for this job.
- How much weight it carries
- A preprint, not peer-reviewed in the version read. The experiments use English benchmarks such as LibriSpeech and TED-LIUM, and the paper says that earlier LLM methods gave very limited gains on LibriSpeech when the word error rate was below 2 percent.
Facts taken from this source
- 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.
“LLMs make many unnecessary changes to the input transcription” (Introduction)