Fact · Voice and AI
Amazon researchers report that combining LLM prompting with fine-tuning brought speech recognition error rates below the N-best oracle level, a benchmark for choosing among the recognizer's candidate transcripts.
Checked against its source on October 17, 2026
In the source's own words
“By combining prompting techniques with fine-tuning we achieve error rates below the N-best oracle level”
Generative Speech Recognition Error Correction with Large Language Models and Task-Activating Prompting, Chao-Han Huck Yang et al., IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) 2023, 2023, Abstract.
Open the original- What it measures
- The tests used out-of-domain ATIS and WSJ tasks, where a large language model corrects the output of a recognizer.
- What it does not show
- Two datasets and 2023 models. The abstract reports the direction of the result, and the exact error rates are in the full paper, which I did not use here.