Towards Human-Like Interactive Speech Recognition With Agentic Correction and Semantic Evaluation

Agentic ASR wraps a standard single-pass ASR front-end in a closed-loop multi-turn refinement framework that adds semantic correction, intent routing, and reasoning-based editing to fix meaning-critical errors that WER and CER metrics cannot capture. The system treats speech recognition as an iterative clarification task rather than a one-shot decode, aligning more closely with how humans resolve misunderstandings in conversation.