RECOVER framework uses multiple ASR hypotheses and LLM critics to correct missed entities
Researchers introduce RECOVER, an agentic post-ASR entity-correction framework that recovers rare or domain-specific terms missing from transcripts, reporting large entity-error reductions across five datasets.
In plain words
Speech recognition systems often miss rare or specialized terms, and if a term never appears in the transcript, later text-only fixes have nothing to work with. RECOVER instead treats the recognizer's multiple output hypotheses as evidence, retrieves candidate entities, and has a large language model propose constrained corrections that pass deterministic guardrails and a critic step. The authors report this recovers entities the first pass missed.
What was announced
The framework, described in an arXiv paper by its authors, works after automatic speech recognition has run. It draws on several recognizer hypotheses as evidence, retrieves relevant entities, and has a large language model propose corrections under constraints. Deterministic guardrails filter the edits, and an LLM critic judges each surviving one; the critic can also trigger a targeted feedback round to recover an entity the first pass missed.
RECOVER offers several hypothesis-use strategies: 1-Best, Entity-Aware Select, ROVER Ensemble, LLM-Select and Adaptive Routing. Evaluated across five datasets against a Whisper-large-v3 baseline, the authors report up to a 49% relative reduction in entity-phrase word error rate and up to 14 percentage points higher entity recall, while preserving overall word error rate, with LLM-Select the strongest strategy.
What it does not show
- The evidence is the paper's abstract only; the full methodology, datasets and results have not been independently verified.
- Reported gains are author claims from their own evaluation; no third-party replication is described.
For security and IT teams
- Scope
- The framework targets entity correction in domains such as finance, medicine and air-traffic control, where recognition errors on rare terms are costly.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.