Elamur Research⌕
☰
home / papers / 2026 / Graph-Based Phonetic Error Correction of Noisy ASR

Graph-Based Phonetic Error Correction of Noisy ASR

Paper · 2026 · Pratik Rakesh Singh, Mohammadi Zaki, Aneesh Mukkamla et al.

Summary

ASR output often looks accurate overall, yet the errors that remain tend to hit meaning-critical words such as names, negations and sentiment terms. These mistakes usually come from sounds being confused, not from random noise. Prompt-based LLM rewriting can hallucinate or over-correct. The authors (Sony Research India) propose G-SPIN, an inference-time pipeline with three stages:

  1. Detect suspect words. A masked language model scores each word by pseudo-log-likelihood and flags unlikely ones, with no reference transcript needed.
  2. Retrieve sound-alike candidates. A GNN, pretrained once over a phoneme-level graph built from the LLM vocabulary, proposes acoustically plausible replacements.
  3. Choose the correction. MLM scoring and an instruction-tuned LLM re-rank only within that small candidate set.

ASR systems and LLMs are not retrained. On English, Spanish, Hindi and Telugu, G-SPIN reaches the lowest WER and the best SeMaScore, beating the LLM-based baselines DoCIA, RLLM-CF and KG. Its BERTScore is equal to or above theirs. The approach is plug-and-play and backbone-agnostic, which makes it relevant for multilingual and lower-resource speech pipelines.

Key points

  • WER drops from the raw ASR baseline to G-SPIN:
    • en: 0.60 → 0.32
    • es: 0.47 → 0.34
    • te: 0.57 → 0.39
    • hi: 0.47 → 0.31
    • The strongest competitors (RLLM-CF, KG) land between 0.34 and 0.44.
  • Detection settings: α = 0.5, τ = 10⁻², up to 5 words flagged per sentence.
  • Graph settings: phoneme graph edges use a cosine threshold of 0.9, and the top-20 GNN candidates are kept.
  • GNN candidate recall: Hits@20 is at least 98% in all four languages (e.g., 99.02 for en). Hindi and Telugu score somewhat lower.
  • English error analysis (200 samples per category):
    • Sound-alike substitutions: WER 0.400 → 0.149
    • Named entities: 0.511 → 0.192
    • Insertions improve only moderately.
    • Deletions barely change (0.73 → 0.71), a stated limitation of text-only correction.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 1Research subjects and languages the work is about.Speech & spoken language · 480 works

Cite

Cite as Pratik Rakesh Singh, Mohammadi Zaki, Aneesh Mukkamla and Pankaj Wasnik. 2026. Graph-Based Phonetic Error Correction of Noisy ASR. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-industry.151

@inproceedings{singh2026graph,
  title = {{Graph-Based Phonetic Error Correction of Noisy ASR}},
  author = {Pratik Rakesh Singh, Mohammadi Zaki, Aneesh Mukkamla and Pankaj Wasnik},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-industry.151},
  url = {https://doi.org/10.18653/v1/2026.acl-industry.151}
}

References

19 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ghazi Bouselmi, D. Fohr, Irina Illina, and Jean Paul Ha-ton. 2006. Multilingual non-native speech recognition using phonetic confusion-based acoustic model modification and graphemic constraints. In Interspeech. find
  2. Yi-Chang Chen, Chun-Yen Cheng, Chien-An Chen, Ming-Chieh Sung, and Yi-Ren Yeh. 2021. Integrated semantic and phonetic post-correction for Chinese speech recognition. In Proceedings of the 33rd Conference on Computational Linguistics and Speech Processing (ROCLING 2021), pages 95–102, Taoyuan, Taiwan… find
  3. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. Preprint, arXiv:1911.02116. find
  4. Yangui Fang, Baixu Chen, Jing Peng, Xu Li, Yu Xi, Chengwei Zhang, and Guohui Zhong. 2025a. Fewer hallucinations, more verification: A three-stage llm-based framework for asr error correction. arXiv preprint arXiv:2505.24347. — [arXiv:2505.24347](https://arxiv.org/abs/2505.24347) find
  5. Yangui Fang, Baixu Chen, Jing Peng, Xu Li, Yu Xi, Chengwei Zhang, and Guohui Zhong. 2025b. Fewer hallucinations, more verification: A three-stage llm-based framework for asr error correction. Preprint, arXiv:2505.24347. — [arXiv:2505.24347](https://arxiv.org/abs/2505.24347) find
  6. William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 1025–1035, Red Hook, NY, USA. Curran Associates Inc.
  7. Chao-Wei Huang and Yun-Nung Chen. 2020. Learning asr-robust contextualized embeddings for spoken language understanding. Preprint, arXiv:1909.10861. — [arXiv:1909.10861](https://arxiv.org/abs/1909.10861) find
  8. Yichong Leng, Xu Tan, Wenjie Liu, Kaitao Song, Rui Wang, Xiang-Yang Li, Tao Qin, Edward Lin, and Tie-Yan Liu. 2023. Softcorrect: Error correction with soft detection for automatic speech recognition. Preprint, arXiv:2212.01039. — [arXiv:2212.01039](https://arxiv.org/abs/2212.01039) find
  9. Xinglin Lyu, Wei Tang, Yuang Li, Xiaofeng Zhao, Ming Zhu, Junhui Li, Yunfei Lu, Min Zhang, Daimeng Wei, Hao Yang, and Min Zhang. 2025. Docia: An online document-level context incorporation agent for speech translation. Preprint, arXiv:2504.05122. find
  10. Rao Ma, Mengjie Qian, Mark Gales, and Kate Knill. 2025. Asr error correction using large language models. Preprint, arXiv:2409.09554. — [arXiv:2409.09554](https://arxiv.org/abs/2409.09554) find
  11. Rena Nemoto, Ioana Vasilescu, and Martine Adda-Decker. 2008. Speech errors on frequently observed homophones in French: Perceptual evaluation vs automatic classification. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), Marrakech, Morocco. European… find
  12. Titouan Parcollet, Yuan Tseng, Shucong Zhang, and Rogier van Dalen. 2025. Loquacious set: 25,000 hours of transcribed and diverse english speech recognition data for research and commercial use. Preprint, arXiv:2505.21578.
  13. Wei Ruan, Yuriy Nechaev, Lin Chen, Chang Su, and Istvan Kiss. 2020. Towards an asr error robust spoken language understanding system. In Proceedings of Interspeech 2020, pages 901–905. find
  14. Zitha Sasindran, Harsha Yelchuri, and T. V. Prabhakar. 2024. Semascore : a new evaluation metric for automatic speech recognition tasks. ArXiv, abs/2401.07506. — [arXiv:2401.07506](https://arxiv.org/abs/2401.07506) find
  15. Piotr Szymański, Piotr Żelasko, Mikolaj Morzy, Adrian Szymczak, Marzena Żyła Hoppe, Joanna Banaszczak, Lukasz Augustyniak, Jan Mizgajski, and Yishay If the correction operator C satisfies: Carmiel. 2020. Wer we are and wer we think we are. Preprint, arXiv:2010.03432. find
  16. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penche… find
  17. Seamless Communication Team. 2023. Seamless: Multilingual expressive and streaming speech translation.
  18. Yi-Cheng Wang, Hsin-Wei Wang, Bi-Cheng Yan, Chi-Han Lin, and Berlin Chen. 2024. Dancer: Entity description augmented named entity corrector for automatic speech recognition. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LRE… find
  19. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. Preprint, arXiv:1904.09675. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/singh-2026-graph-based-phonetic-error-correction/