AI 研究
Agentic-GER Revisits Audio Segments, Cutting Chinese Terminology Error Rates by Up to 36.8%
The method has an agent read the full transcript, then re-transcribe suspicious segments to correct inconsistent proper nouns across segments. The official implementation is public, but the largest gains are limited to a specific model configuration; deployment still requires weighing the rate of incorrect edits against inference costs.
Researchers from Shanghai Jiao Tong University, Alibaba, and other institutions made Agentic-GER public on September 24. The method breaks long-form speech correction into finding suspicious items, checking the audio, and making minimal edits. It builds on existing speech recognition results and uses the full transcript to help identify proper nouns. The paper is currently a preprint and is marked as submitted to ICASSP 2027. [Paper details](https://arxiv.org/abs/2609.29428)
The process first summarizes the topic and terminology, then selects suspicious segments using the full transcript and working memory, and sends them to Qwen3-ASR-1.7B for re-transcription. A language model combines the new recognition result with the surrounding context to decide whether to keep or edit the text; if the evidence is insufficient, it keeps the original. Both edit and keep decisions are recorded in memory for the next round. Segment boundaries remain unchanged. Each round checks at most four candidates in Chinese, and each recording is limited to 32 rounds and 48 accepted edits. [Implementation guide](https://github.com/QwenAudio/FunResearch/blob/main/Agentic-GER/README.md)
The evaluation covers 524 Chinese and 387 English recordings across 12 domains in GigaSpeechBench. The authors report that the largest relative improvement in Chinese terminology came from pairing a Whisper-Large-v3 draft with Qwen3.8-Max with thinking disabled: B-CER fell from 35.07% to 22.16%, a relative reduction of 36.8%. This metric counts only annotated terms, so it should not be read as an equivalent reduction in errors across the full transcript. When the initial draft used FunASR, the same corrector achieved a relative reduction of 17.9%. [Evaluation results](https://arxiv.org/html/2609.29428v1)
The gains also depend on the type of error: substitution errors often leave suspicious text behind, while omitted terms are harder to find by scanning the full transcript. More thinking is not always better: Max had worse terminology results in three of the four comparisons. Overall recognition error rates may also rise slightly with some configurations. Evaluations of the corrector should track both terminology and full-transcript quality. [Limitations analysis](https://arxiv.org/html/2609.29428v1)
The official repository provides prompts, evaluation adapters, and Chinese and English configurations, but its default model is Qwen3.8-27B, which differs from the configuration that produced the largest gains. The quick start processes only one recording per language and does not include raw experiment results for each recording. Reproducing the results requires a separate model service, and teams should verify that each run completed successfully so that an original transcript carried forward after a failure is not mistaken for a corrected result. [Reproduction notes](https://github.com/QwenAudio/FunResearch/blob/main/Agentic-GER/README.md)
The workflow suggests that this method’s value for Chinese technical content comes from using terms repeated across segments to make up for the limited semantic context in short clips. However, it depends on a complete transcript, and this study has not demonstrated real-time benefits for streaming services. Engineering teams can first add it after their existing transcription pipeline, measure the rate of incorrect edits, latency, and call costs on recordings from their own domains, and then assess its suitability for long audio such as meetings and lectures.