RECOVER: Robust Entity Correction via agentic Orchestration of hypothesis Variants for Evidence-based Recovery

Researchers Abhishek Kumar and Aashraya Sachdeva have introduced RECOVER (Robust Entity Correction via agentic Orchestration of hypothesis Variants for Evidence-based Recovery), a novel framework designed to tackle a critical weakness in Automatic Speech Recognition (ASR). Standard ASR systems often struggle with rare, domain-specific entities like medical terms, financial jargon, or proper names, and can even omit them entirely. RECOVER addresses this by acting as a post-processing, tool-using agent that orchestrates multiple correction strategies. It doesn't rely on a single transcript; instead, it uses several ASR output hypotheses as evidence, retrieves relevant entities from knowledge sources, and applies a Large Language Model (LLM) to perform constrained corrections.

The system evaluates four distinct strategies for using these hypotheses: the standard 1-Best output, an Entity-Aware Select method, a ROVER Ensemble technique, and an LLM-Select approach. In comprehensive evaluations across five diverse datasets, RECOVER demonstrated significant improvements. The framework achieved relative reductions in entity-phrase word error rate (E-WER) ranging from 8% to 46%, and increased entity recall by up to 22 percentage points. The LLM-Select strategy emerged as the top performer, delivering the best entity correction results without degrading the overall word error rate (WER) of the general transcript. This makes it a practical upgrade for existing ASR pipelines in high-stakes fields.