Phonetic-to-Character Generation for Low-Resource Chinese Dialects: A Benchmark and Retrieval-Augmented Correction Framework
Keywords:
low-resource dialects, phonetic-to-character, retrieval augmentationAbstract
This study addresses the challenge of phonetic-to-character generation for low-resource Chinese dialects, where standard input methods fail due to data scarcity and orthographic ambiguity. We introduce a benchmark covering six under-documented dialect families and propose a retrieval-augmented correction framework that combines a lightweight acoustic model with a dialect-specific character retrieval index. The benchmark includes over 120,000 aligned phonetic sequences and character transcriptions, annotated for tone sandhi and regional variants. Our framework first generates candidate characters via a phoneme-to-grapheme neural transducer, then refines them using a retrieval module that queries a curated dialect lexicon and a context-aware language model. Experimental results show that the proposed method improves character accuracy by 18.4 percentage points over a strong baseline, and reduces the character error rate from 32.7% to 14.3% on unseen dialect samples. Ablation studies confirm that retrieval augmentation contributes the largest gain, while tonal normalization and data augmentation each add smaller but significant improvements. We also evaluate cross-dialect transfer, finding that a model trained on three high-resource dialects generalizes to low-resource ones with only 5% of the original training data. The benchmark and framework are released to support reproducible research in dialectal input method engineering. Our findings suggest that retrieval-augmented generation is a viable path for low-resource phonetic input, and that careful benchmark design is essential for measuring progress in this domain.References
1. F. Alam, S. A. Chowdhury, S. Boughorbel, and M. Hasanain, "LLMs for low resource languages in multilingual, multimodal and dialectal settings," in Proc. 18th Conf. Eur. Chapter Assoc. Comput. Linguistics: Tutorial Abstracts, Mar. 2024, pp. 27–33.
2. A. C. Peng, K. T. Huang, T. H. Lo, H. S. Lee, H. M. Wang, and B. Chen, "Efficient dialect-aware modeling and conditioning for low-resource Taiwanese Hakka speech processing," arXiv preprint arXiv:2602.22522, 2026.
3. W. Zhang, T. Liao, and Y. Feng, "Multi-Agent LLM Collaborative Reasoning and Task Planning for Complex Task Solving," in Proc. IEEE CISAT, 2026, doi: 10.1109/CISAT71030.2026.11682610.
4. J. Wu, "A Methodological Framework for Training Tool-Aware AI Entertainers for Live Streaming," Artificial Intelligence and Digital Technology, vol. 3, no. 3, pp. 1–12, 2026.
5. K. A. Ta, N. Van Dinh, and K. Van Nguyen, "ViDia2Std: A parallel corpus and methods for low-resource Vietnamese dialect-to-standard translation," in Proc. AAAI Conf. Artif. Intell., vol. 40, no. 36, Mar. 2026, pp. 29995–30004.
6. S. Balamurugan, R. Uthra, T. Janarthanan, and K. S. Madasamy, "Low-resource dialect identification in global English variants using few-shot meta-learning and acoustic-phonetic features," in Proc. 2026 Innovations Mach., Eng., Digit. Conf. (IMED), Mar. 2026, pp. 1–6.
7. H. Azis, H. Darwis, and T. Hasanuddin, "Development of a low-resource automatic speech recognition system for the Makassar dialect," in Proc. 2026 20th Int. Conf. Ubiquitous Inf. Manage. Commun. (IMCOM), Jan. 2026, pp. 1–8.
8. E. Khan, F. Saidani, O. Van Esbroeck, R. Khoury, and L. Kosseim, "Low-resource dialect adaptation of large language models: A French dialect case-study," arXiv preprint arXiv:2510.22747, 2025.
9. S. Qin, L. Wang, S. Li, J. Dang, and L. Pan, "Improving low-resource Tibetan end-to-end ASR by multilingual and multilevel unit modeling," EURASIP J. Audio, Speech, Music Process., vol. 2022, no. 1, p. 2, 2022.
10. Y. Shu et al., "DialectS2S: End-to-end speech dialogue modeling for low-resource Chinese dialects," arXiv preprint arXiv:2608.08067, 2026.
11. Y. Zou et al., "Benchmarking Large Language Models for Chinese and Japanese IMEs: Phonetic-to-Character Generation and Textual Error Correction," in Proc. LREC, vol. 11, no. 16, 2026.
12. L. Pan, S. Li, L. Wang, and J. Dang, "Effective training end-to-end ASR systems for low-resource Lhasa dialect of Tibetan language," in *Proc. 2019 Asia-Pacific Signal Inf. Process. Assoc. Annu. Summit Conf. (APSIPA ASC)*, Nov. 2019, pp. 1152–1156.
13. M. Tzudir, P. Sarmah, and S. R. Prasanna, "Analysis and modeling of dialect information in Ao, a low resource language," J. Acoust. Soc. Amer., vol. 149, no. 5, pp. 2976–2987, 2021.
14. A. Singh et al., "Model adaptation for ASR in low-resource Indian languages," arXiv preprint arXiv:2307.07948, 2023.
15. N. Aepli, "There is plenty of room at the bottom: Challenges & opportunities in low-resource non-standardized language varieties," Doctoral dissertation, Univ. Zurich, Switzerland, 2024.
16. R. Liu, W. Guo, T. Liao, and X. Fang, "Uncertainty-Aware Adaptive Debate for Robust and Efficient Large Language Model Reasoning," Research Square Preprint, Aug. 2026, doi: 10.21203/rs.3.rs-10581777/v1.
17. S. Yuan, "Data flow mechanisms and model applications in intelligent business operation platforms," Financ. Econ. Insights, vol. 2, no. 1, pp. 144–151, Dec. 2025, doi: 10.70088/m66tbm53.
18. S. K. Sahoo, B. K. Mishra, S. Parida, S. R. Dash, J. N. Besra, and E. V. Tello, "Automatic dialect detection for low resource Santali language," in Proc. 2021 19th OITS Int. Conf. Inf. Technol. (OCIT), Dec. 2021, pp. 234–238.
19. V. Ondrejová and M. Šuppa, "Can LLMs handle low-resource dialects? A case study on translation and common sense reasoning in Šariš," in Proc. 11th Workshop NLP Similar Languages, Varieties, Dialects (VarDial 2024), Jun. 2024, pp. 130–139.
20. J. Wu, "Personality Is a Training Target: Data Curation and SFT for an Entertaining, Agentic AI Streamer," AI and Data Science Journal, vol. 3, no. 2, pp. 29–36, 2026.
21. Z. J. Wong, W. X. Koh, and K. Y. T. Lim, "Exploring the use of natural language processing in the translation of dialects using limited data: A case study of the Hokkien dialect," in EDULEARN25 Proc., 2025, pp. 2509–2515.
22. M. Tzudir, S. Baghel, P. Sarmah, and S. R. Prasanna, "Under-resourced dialect identification in Ao using source information," J. Acoust. Soc. Amer., vol. 152, no. 3, pp. 1755–1766, 2022.
23. R. Amooie, W. De Vries, Y. Hao, J. Dijkstra, M. Coler, and M. Wieling, "Enhancing standard and dialectal Frisian ASR: Multilingual fine-tuning and language identification for improved low-resource performance," in Proc. ICASSP 2025—2025 IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), Apr. 2025, pp. 1–5.
24. A. A. Gedi, Y. M. Ahmed, S. A. Mohamed, Y. A. Yusuf, and A. U. Ebdiyow, "Somali dialect identification in low-resource settings using machine learning and deep learning," Discover Data, vol. 4, no. 1, p. 10, 2026.
25. A. Suwanbandit et al., "Thai-dialect: Low resource Thai dialectal speech to text corpora," in Proc. 2023 IEEE Autom. Speech Recognit. Understanding Workshop (ASRU), Dec. 2023, pp. 1–8.
