Preprint
LLM-based Generative Error Correction for Rare Words with Synthetic Data and Phonetic Context
Natsuo Yamashita et al., arXiv (accepted by INTERSPEECH 2025), 2025
Read the originalRead on October 17, 2026
- What it is
- A 2025 paper from Japanese researchers on using a large language model to correct speech recognition output after the recognizer has run, which they call generative error correction. It documents over-correction, where a text-only model rewrites words that were already right, and tests a fix that adds the recognizer's alternative hypotheses and phonetic context.
- How much weight it carries
- Peer-reviewed venue (INTERSPEECH 2025) but cited here from the arXiv abstract, which is open and quotable. Results are reported on English and Japanese datasets for rare-word correction, not on everyday dictation, so the over-correction finding is a mechanism to check rather than a measured rate.
Facts taken from this source
- The 2025 INTERSPEECH authors add the speech recognizer's N-best hypotheses (its ranked alternative transcripts) and phonetic context to the language model, to reduce over-correction. They report lower word and character error rates on English and Japanese datasets.
“we integrate ASR's N-best hypotheses along with phonetic context to mitigate over-correction.” (Abstract)
- In the 2025 INTERSPEECH paper, existing language-model error correction for speech recognition relies mainly on the text, ignores phonetic cues, and that leads to over-correction, meaning the model changes words that were already right.
“existing LLM-based GER approaches primarily rely on textual information, neglecting phonetic cues, which leads to over-correction.” (Abstract)