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

Zixuan Jiang, Yanqiao Zhu, Peng Wang, Qinyuan Chen, Xinjian Zhao, Xipeng Qiu, Wupeng Wang, Zhifu Gao, Xiangang Li, Kai Yu, Xie Chen · 2026

Read Towards Human-Like Interactive Speech Recognition with Agentic Correction and Semantic Evaluation by Zixuan Jiang, Yanqiao Zhu, Peng Wang, Qinyuan Chen, Xinjian Zhao, Xipeng Qiu, Wupeng Wang, Zhifu Gao, Xiangang Li, Kai Yu, Xie Chen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-29430.txt.

Description

Zixuan Jiang frames Interactive ASR as a multi-turn task. Most speech recognition systems follow a single-pass pattern, poorly aligned with human talk, where mix-ups are resolved through iterative follow-ups. Their Agentic ASR system combines a single-pass front-end with meaning fixes, intent routing, and reasoned editing. Tests on many-language and code-switching benchmarks show back-and-forth interaction always reduces meaning errors. Gains are much larger in the meaning metric than in token-level metrics.

Catalog record

CREATORZixuan Jiang, Yanqiao Zhu, Peng Wang, Qinyuan Chen, Xinjian Zhao, Xipeng Qiu, Wupeng Wang, Zhifu Gao, Xiangang Li, Kai Yu, Xie Chen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39286
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.29430

Browse the full library · About the Tower of Babel