Agentic ASR evaluation
Towards Human-Like Interactive Speech Recognition With Agentic Correction and Semantic Evaluation
Zixuan Jiang, Yanqiao Zhu, Peng Wang, Qinyuan Chen, Xinjian Zhao, Xipeng Qiu, et al. · 2026-05-28
- Modality
- Audio & speech
- Data
- Hybrid (real + synthetic/simulated)
- Task · ground truth
- Gen./Rev. · Text+Sem.
- Scope
- component evaluation
What it tests
- CECross-modal evidenceDirectly evaluated
- LELong-range evidencePartially evaluated
- SCState changesDirectly evaluated
- CFConflicting memoriesDirectly evaluated
- UQUnanswerable queriesNot evaluated
- MAMemory-grounded actionNot evaluated
- SPSpeaker trackingNot evaluated
- ACAcoustic-cue retentionPartially evaluated
- RPRepair persistenceDirectly evaluated
Survey notes
Each episode iteratively corrects one speech transcription until semantic equivalence or a round limit. It tests interactive ASR repair, not persistence and later use of corrected agent memory.