Selective Gating for LLM-Based Chinese ASR Error Correction
DOI: https://doi.org/10.62517/jbdc.202601409
Author(s)
Yazhou Zhao
Affiliation(s)
Ningxia University, Yinchuan, Ningxia, China
Abstract
This study examines over-correction in Chinese automatic speech recognition (ASR) post-processing with a general-purpose large language model (LLM). Whisper-small generates ASR hypotheses from Common Voice zh-CN, and Qwen2.5-3B-Instruct performs text correction. Three selective strategies-edit-ratio gating, ASR-confidence gating, and a conservative joint gate-are developed on a 100-utterance development set and evaluated with frozen settings on an independent 1,000-utterance test set. Direct LLM correction increases the macro-averaged character error rate (Macro CER) from 34.75% to 44.75%. The joint gate reduces the number of worsened utterances from 469 to 5, but yields a Macro CER of 34.79%, slightly above the Whisper baseline. These results show that, under the present setting, a small general-purpose instruction-tuned LLM is not a reliable unconditional ASR post-processor. Conservative gating substantially limits error propagation, but does not produce a net accuracy gain, highlighting the need for more reliable error detection and named-entity protection.
Keywords
Automatic Speech Recognition; Large Language Model; Error Correction; Whisper; Selective Gating; Character Error Rate
References
[1] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust Speech Recognition via Large-Scale Weak Supervision,” in Proceedings of the 40th International Conference on Machine Learning, PMLR, vol. 202, pp. 28492–28518, 2023.
[2] R. Ardila, M. Branson, K. Davis, et al., “Common Voice: A Massively-Multilingual Speech Corpus,” in Proceedings of the Twelfth Language Resources and Evaluation Conference, Marseille, France: European Language Resources Association, pp. 4218–4222, 2020.
[3] A. Yang, B. Yang, B. Zhang, et al., “Qwen2.5 Technical Report,” arXiv:2412.15115, 2024.
[4] Z. Fang, R. Zhang, Z. He, H. Wu, and Y. Cao, “Non-Autoregressive Chinese ASR Error Correction with Phonological Training,” in Proceedings of NAACL 2022: Human Language Technologies, Seattle, United States: Association for Computational Linguistics, pp. 5907–5917, 2022, doi: 10.18653/v1/2022.naacl-main.432.
[5] R. Ma, M. Qian, M. Gales, and K. Knill, “ASR Error Correction using Large Language Models,” arXiv:2409.09554, 2024.
[6] T. Udagawa, M. Suzuki, M. Muraoka, and G. Kurata, “Robust ASR Error Correction with Conservative Data Filtering,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track, Miami, Florida, US: Association for Computational Linguistics, pp. 256–266, 2024.
[7] V. J. Wei, W. Wang, D. Jiang, Y. Song, and L. Wang, “ASR-EC Benchmark: Evaluating Large Language Models on Chinese ASR Error Correction,” in Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, Suzhou, China: Association for Computational Linguistics, pp. 1567–1575, 2025, doi: 10.18653/v1/2025.emnlp-industry.110.
[8] Mozilla Data Collective, “Common Voice Scripted Speech 26.0 - Chinese (China),” dataset, 2026.