Fact · Voice and AI
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.
Checked against its source on October 17, 2026
In the source's own words
“existing LLM-based GER approaches primarily rely on textual information, neglecting phonetic cues, which leads to over-correction.”
LLM-based Generative Error Correction for Rare Words with Synthetic Data and Phonetic Context, Natsuo Yamashita et al., arXiv (accepted by INTERSPEECH 2025), 2025, Abstract.
Open the original- What it measures
- GER is generative error correction: a large language model rewrites the output of a speech recognizer. The authors state the problem as a limitation of earlier text-only approaches, then propose adding phonetic context and the recognizer's alternative hypotheses.
- What it does not show
- The abstract names over-correction as a problem but does not give a rate of changed words. The experiments target rare and domain-specific words, not everyday dictation.