agentic-ger: terminology recovery in long-form speech using global context

Yanqiao Zhu; Wupeng Wang; Zhifu Gao; Xiangang Li; Xie Chen

cs.CLSep 24, 2026

agentic-ger: terminology recovery in long-form speech using global context

Authors: Yanqiao Zhu, Wupeng Wang, Zhifu Gao, Xiangang Li, Xie Chen

Organizations: X-LANCE Lab, Shanghai Jiao Tong University · Shanghai Innovation Institute · Alibaba Token Foundry

Abstract

Recent advances in speech language models have improved automatic speech recognition (ASR) for long-form audio. However, accurately and consistently transcribing domain-specific terminology remains challenging. Motivated by the world knowledge and contextual capability of large language models (LLMs), we propose Agentic-GER, an LLM-based agent for terminology correction in long-form speech. The agent uses global context from the full transcript to identify suspicious terms and resolve ambiguous hypotheses. It selectively re-transcribes the source speech to check candidate corrections, and uses accepted edits to guide subsequent decisions. Experiments with four LLMs and two ASR systems on GigaSpeechBench show consistent terminology improvements in both Chinese and English, with and without thinking. On Chinese speech, Agentic-GER achieves up to a 36.8% relative reduction in biased character error rate (B-CER) over the Whisper baseline.

Figures & tables

Figure 1: Overview of Agentic-GER. Global context guides error discovery and correction. The agent re-transcribes selected segments to evaluate edits, updates the transcript, and records decisions for subsequent rounds.

Figure 2: Resolving an ambiguous ASR hypothesis using global context. The agent flags “death certificates” in a discussion of medical device evolution. Re-transcription yields “deathscapes.” An earlier segment mentions a “stethoscope.” The agent corrects the term to “stethoscopes,” matching the reference. Excerpts are shortened for display.

	Chinese		English
	B-CER $\downarrow$		B-WER $\downarrow$
	FunASR	Whisper	FunASR	Whisper
Baseline	10.91	35.07	12.94	14.67
+Qwen3.8-27B	10.19 (6.7)	28.88 (17.7)	12.35 (4.6)	14.08 (4.0)
+Gemma4-31B	9.88 (9.5)	25.09 (28.5)	12.05 (6.9)	13.71 (6.5)
+Qwen3.8-Flash	9.64 (11.7)	26.61 (24.1)	11.91 (8.0)	13.72 (6.5)

Table 1: Main terminology results without thinking. Error rates (%,

\downarrow

); parentheses show relative error reductions (%,

\uparrow

) from Baseline. Column minima are in bold.

Language	Error type	Proportion	Reduction $\uparrow$
Chinese	Substitution	85.0	21.3
Chinese	Deletion	14.7	3.0
English	Substitution	56.0	6.4
English	Deletion	43.1	0.4

Table 2: Terminology error analysis for Qwen3.8-27B on Whisper without thinking. Proportion (%) is each error type’s percentage of all baseline terminology errors, including insertions. Reduction (%,

\uparrow

) is the relative decrease in its error count.

	Thinking	B-CER $\downarrow$	CER $\downarrow$
Baseline	—	10.91	3.12
yq +Local GER	✗	10.80	3.37
yq +Local GER	✓	10.63	3.42
yq +Agentic-GER	✗	10.19	3.11
yq +Agentic-GER	✓	9.03	3.01

Table 3: Local GER versus Agentic-GER. Error rates (%,

\downarrow

); best values in bold. “Thinking” indicates whether intermediate token generation is enabled ( ✓ ) or disabled ( ✗ ).

	Chinese		English
	B-CER $\downarrow$		B-WER $\downarrow$
	FunASR	Whisper	FunASR	Whisper
+Qwen3.8-27B	9.03	25.89	11.96	13.87
$\Delta$ Thinking (pp) $\uparrow$	$+10.6$	$+8.5$	$+3.0$	$+1.5$
+Gemma4-31B	9.32	23.20	11.79	13.60
$\Delta$ Thinking (pp) $\uparrow$	$+5.1$	$+5.4$	$+2.0$	$+0.8$

Table 4: Terminology error rates (%,

\downarrow

) with thinking.

\Delta

Thinking reports the change in relative error reduction compared with Table 1 (percentage points,

\uparrow

). Positive values indicate improvement; negative values (bold) indicate degradation.

Explore similar work

AgenticASR: Refining Speech Recognition in Real-World Scenarios via an Agentic Approach
Jul 30, 2026Zixuan Jiang, Binghao Qiang, Jiaying Chi +3Automatic Speech Recognition Transcript
X-LANCE lab, Shanghai Jiao Tong University · Shanghai Innovation Institute · College of Artificial Intelligence, Xi’an Jiaotong University
Voice Memory for Agentic Speech Recognition
Jul 29, 2026Chao-Han Huck Yang, Zih-Ching Chen, Piotr Zelasko +3Error Correction
NVIDIA
LOGIC: Efficient and Robust Contextual Biasing for Speech LLMs via Logit-Space Integration
Jan 21, 2026Peidong Wang, Jian Xue, Jinyu LiSpeech Language Models Logit Lens
Microsoft, USA

Jul 30, 2026, cs.AIJ/K move · Enter open · S save

AgenticASR: Refining Speech Recognition in Real-World Scenarios via an Agentic Approach

Zixuan Jiang, Binghao Qiang, Jiaying Chi, Yanqiao Zhu+2

X-LANCE lab, Shanghai Jiao Tong University · Shanghai Innovation Institute · College of Artificial Intelligence, Xi’an Jiaotong University

Automatic Speech Recognition Transcript Agentic Real-World

Open paper AskPDF arXiv

Automatic speech recognition (ASR) has achieved substantial gains in transcription accuracy, yet verbatim transcription does not necessarily produce readily usable text. It retains fillers, repetitions, false starts, and self-corrections that increase reading effort, obscure the speaker's final intent, and propagate unresolved or abandoned content to downstream tasks. Existing spoken-to-written methods process completed audio or transcripts but cannot revise emitted text when later speech changes how preceding content should be interpreted. We therefore formulate Agentic Speech Recognition (AgenticSR), an audio-to-clean-text task that removes disfluencies, resolves self-corrections, and normalizes written form while preserving the speaker's final intent. AgenticASR implements this task through an ASR--Refiner architecture that repeatedly transforms a bounded active context and replaces its corresponding output span as audio arrives. This enables continual emission and revision over streams of arbitrary duration. We also introduce AASR-Bench, a bilingual benchmark with fine-grained atomic rubrics. Across multiple ASR front ends, AgenticASR attains the highest AASR-Bench scores among evaluated systems. A human--AI agreement study shows that rubric-based judgments align with independent expert assessments. Ablations characterize Refiner capacity, context length, and the quality--latency trade-off between online and offline inference. Together, these results establish AgenticASR as a practical framework for intent-preserving clean transcription during ongoing speech. Code, AASR-Bench, and a demo will be released at https://github.com/AnXMuy/AgenticASR.

Abstract

Explore similar work

AgenticASR: Refining Speech Recognition in Real-World Scenarios via an Agentic Approach

Voice Memory for Agentic Speech Recognition

LOGIC: Efficient and Robust Contextual Biasing for Speech LLMs via Logit-Space Integration