Interactive ASR

Interactive ASR teaser figure

Comparison between daily human communication, the traditional ASR paradigm, and the proposed Agentic ASR paradigm.

Abstract

Automatic speech recognition (ASR) is a core component of human-computer interaction and an increasingly important front-end for LLM-based assistants and agents. However, most current ASR systems still follow a single-pass paradigm, which is poorly aligned with human communication, where misunderstandings are resolved through iterative clarification and refinement. This mismatch makes it difficult to correct meaning-critical errors once they occur. Meanwhile, token-level metrics such as WER or CER cannot adequately reflect such a problem. To address these limitations, we formulate Interactive ASR as a multi-turn refinement task and propose Agentic ASR, a closed-loop framework that combines a single-pass ASR front-end with semantic correction, intent routing, and reasoning-based editing. We further introduce the Sentence-level Semantic Error Rate (S2ER), an LLM-based semantic evaluation metric, together with an Interactive Simulation System for scalable and reproducible benchmarking. Experiments on multilingual, named-entity-intensive, and code-switching benchmarks show that iterative interaction consistently reduces semantic errors, with much larger gains in S2ER than in conventional token-level metrics. Human-AI alignment and ablation studies further validate the reliability of the semantic judge and the robustness of the proposed framework.

Method

Agentic ASR framework

Agentic ASR

Interactive Simulation System

Interactive Simulation System

Result

Category Benchmark Loop 0 Loop 1 Loop 3 Loop 10
Multilingual GigaSpeech Test 21.47% 12.35% 7.00% 3.49%
Multilingual WenetSpeech Test_Net 19.46% 8.69% 4.15% 1.80%
Name Entity AISHELL-NER Dev† 17.38% 8.45% 4.45% 1.97%
Name Entity AISHELL-NER Test† 19.91% 9.55% 5.47% 2.02%
Code Switching CS-Dialogue Test† 19.73% 10.83% 6.58% 4.16%
Code Switching ASRU2019 Test 28.57% 10.32% 3.98% 1.36%
Different Base ASR Model Ablation

Different Base ASR Model Ablation

Examples

ID: TEST_NET_Y0000000000_-KTKHdZ2fb8_S00043 2 Rounds
Ground Truth: 吴克儿待会儿去趟监控室
User Audio
Model Output
Loop 0: 吴克,带我去趟办公室。
User Correction
不是吴克,是吴克儿。要纠正的是待会儿去监控室,不是带我去办公室。
User Audio
Model Output
Loop 1: 吴克儿,待会儿去监控室。
User Correction
是去趟监控室,不是去监控室,有个趟字。
User Audio
Model Output
Loop 2: 吴克儿待会儿去趟监控室

Citation

@misc{wang2026interactiveasrhumanlikeinteraction,
  title={Interactive ASR: Towards Human-Like Interaction and Semantic Coherence Evaluation for Agentic Speech Recognition},
  author={Peng Wang and Yanqiao Zhu and Zixuan Jiang and Qinyuan Chen and Xingjian Zhao and Xipeng Qiu and Wupeng Wang and Zhifu Gao and Xiangang Li and Kai Yu and Xie Chen},
  year={2026},
  eprint={2604.09121},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2604.09121},
}

@misc{jiang2026humanlikeinteractivespeechrecognition,
  title={Towards Human-Like Interactive Speech Recognition With Agentic Correction and Semantic Evaluation},
  author={Zixuan Jiang and Yanqiao Zhu and Peng Wang and Qinyuan Chen and Xinjian Zhao and Xipeng Qiu and Wupeng Wang and Zhifu Gao and Xiangang Li and Kai Yu and Xie Chen},
  year={2026},
  eprint={2605.29430},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2605.29430},
}