Elamur Research⌕
☰
home / papers / 2026 / G&P2P: A Multi-Source Approach to Grapheme-to-Phoneme Conversion

G&P2P: A Multi-Source Approach to Grapheme-to-Phoneme Conversion

Paper · 2026 · Chun-Yi Jerry Peng

Summary

Grapheme-to-phoneme (G2P) models usually learn from a single pronunciation lexicon, even though most languages have several dictionaries that differ in coverage, notation and reliability. This paper presents G&P2P, a framework that adds a "side pronunciation" from a second English dictionary to each training pair. The author draws on three expert-built resources (CELEX, PronLex, NETTalk) and two crowd-sourced Wiktionary lexicons (WikiPron_UK, WikiPron_US). Inputs are merged in three ways: extra columns, control symbols, or role-marking subscripts. Four architectures are trained with Yoyodyne: attentive LSTM, pointer-generator LSTM, transformer and pointer-generator transformer.

Pairing two expert dictionaries cuts word error rate (WER) sharply. Adding crowd-sourced data helps little and sometimes hurts, especially for transformers. The fusion method makes little difference. Pointer-generator LSTMs generally beat transformers, but they are much larger and slower to train. The practical lesson for multi-source G2P is that choosing clean sources matters more than how the sources are combined.

Key points

  • Lexicon sizes: CELEX 72,995 entries; NETTalk 20,006; PronLex 96,941; WikiPron_US 78,633; WikiPron_UK 79,520. A custom extractor normalized the English "r" in the Wiktionary data.
  • Largest gains (PG LSTM, median WER over 5 seeds):
    • CELEX alone scored 26.90, falling to 19.81 with PronLex as the side pronunciation.
    • PronLex alone scored 31.60, falling to 19.79 with CELEX, an 11.81-point (about 37% relative) reduction.
    • The abstract reports a different headline figure: 11.26 points, 22% relative.
  • Crowd-sourced side data: adding WikiPron improved WER by only about 0.1 points and degraded the pointer-generator transformer results.
  • Fusion strategy: the three strategies differed by roughly 1 WER point or less. A Friedman test found no significant effect (χ²(2) = 2.40, p = .30).
  • LSTM vs. transformer:
    • A Wilcoxon test favored the PG LSTM.
    • Model size: the PG LSTM has about 53.6M parameters, against 3.2M for the PG transformer.
    • Training time: about 3h 42m for the PG LSTM, against about 1h 2m for the PG transformer.

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

 Institutions

🎓City University of New YorkNew York · United States17 works · 2,564 citations🎓The Graduate Center, CUNYNew York · United States13 works · 207 citations

Also on the byline: Borough of Manhattan Community College.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksPsychology · 1,649 worksExperimental and Cognitive Psychology · 930 works

Cite

Cite as Chun-Yi Jerry Peng. 2026. G&P2P: A Multi-Source Approach to Grapheme-to-Phoneme Conversion. Proceedings of the Third Workshop on Computation and Written Language (CAWL 2026) @ LREC 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/4f2x3fda6jj7

@inproceedings{peng2026g,
  title = {{G\&P2P: A Multi-Source Approach to Grapheme-to-Phoneme Conversion}},
  author = {Chun-Yi Jerry Peng},
  booktitle = {Proceedings of the Third Workshop on Computation and Written Language (CAWL 2026) @ LREC 2026},
  year = {2026},
  doi = {10.63317/4f2x3fda6jj7},
  url = {https://doi.org/10.63317/4f2x3fda6jj7}
}

References

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

  1. Lucas F. E. Ashby, Travis M. Bartley, Simon Clematide, Luca Del Signore, Cameron Gibson, Kyle Gorman, Yeonju Lee-Sikka, Peter Makarov, Aidan Malanoski, Sean Miller, Omar Ortiz, Reuben Raff, Arundhati Sengupta, Bora Seo, Yulia Spektor, and Winnie Yan. 2021. Results of the second SIGMORPHON shared tas… find
  2. PG transformer 23.09 20.37 56.52 22.78 22.34 59.56 38.39 45.40 55.93 42.25 *Table 6: WERs from pointer-generator LSTM and pointer-generator transformer models.* Trainable params Total params Total size (MB) Training Time PG LSTM 53.6 M 53.6 M 214.512 3h 42m PG Trans. 3.2 M 3.2 M 12.732 1h 1m 53s ing… find
  3. Nikhil Prabhu and Katharina Kann. 2020. Making a point: Pointer-generator transformers for disjoint vocabularies. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processi… find
  4. Kanishka Rao, Fuchun Peng, Haşim Sak, and Françoise Beaufays. 2015. Grapheme-to-phoneme conversion using long short-term memory recurrent neural networks. In 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4225–4229, South Brisbane, Australia. find
  5. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations (ICLR 2015), Conference Track Proceedings, San Diego, USA.
  6. Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073–1083, Vancouver, Canada. Association for Comput… find
  7. Jacob Eisenstein. 2019. Introduction to Natural Language Processing. MIT Press. find
  8. Kyle Gorman, Lucas F.E. Ashby, Aaron Goyzueta, Arya McCarthy, Shijie Wu, and Daniel You. 2020. find
  9. The SIGMORPHON 2020 shared task on multilingual grapheme-to-phoneme conversion. In Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 40–50, Online. Association for Computational Linguistics. find
  10. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, Long Beach, USA. Curran Associates, Inc.
  11. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural Computation, 9:1735–1780. find
  12. Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2015. Pointer networks. In Advances in Neural Information Processing Systems, volume 28, pages 2692–2700, Montréal, Canada. Curran Associates, Inc. find
  13. Piotr Kłosowski. 2022. A rule-based grapheme-to-phoneme conversion system. Applied Sciences, 12(5). find
  14. Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Effective approaches to attention-based neural machine translation. In Proceed-Adam Wiemerslage, Kyle Gorman, and Travis M. Bartley. 2025. Yoyodyne: Small-vocabulary neu-Dataset CELEX → NETTalk CELEX → PronLex CELEX → WikiPron_UK PronLex → CE…
  15. PG LSTM 20.95 19.79 52.06 19.81 20.36 55.71 26.77 44.60 24.03 30.01 *Table 8: WERs from attentive and pointer-generator LSTM models.* ral sequence-to-sequence models. Computer software. find
  16. Sevinj Yolchuyeva, Géza Németh, and Bálint Gyires-Tóth. 2019a. Grapheme-to-phoneme conversion with convolutional neural networks. Applied Sciences, 9(6). find
  17. Sevinj Yolchuyeva, Géza Németh, and Bálint Gyires-Tóth. 2019b. find
  18. Transformer Based Grapheme-to-Phoneme Conversion. In Interspeech 2019, pages 2095–2099, Graz, Austria. 9. find
  19. Language Resource References Baayen, R. H. and Piepenbrock, R. and Gulikers, L. 1995. CELEX2. Linguistic Data Consortium. LDC Catalog No.: LDC96L14, ISBN: 1-58563-085-3. find
  20. Kingsbury, Paul and Strassel, Stephanie and McLemore, Cynthia and MacIntyre, Robert. 1994. CALLHOME American English Lexicon (PRON-LEX). Linguistic Data Consortium. LDC Catalog No.: LDC97L20, ISBN: 1-58563-110-8. find
  21. Lee, Jackson L. and Ashby, Lucas F.E. and Garza, M. Elizabeth and Lee-Sikka, Yeonju and Miller, Sean and Wong, Alan and McCarthy, Arya D. and Gorman, Kyle. 2020. Massively Multilingual Pronunciation Modeling with WikiPron. European Language Resources Association. find
  22. Sejnowski, Terry and Rosenberg, Charles. 1988. Connectionist Bench (Nettalk Corpus). DOI: https://doi.org/10.24432/C5VP6T. — [doi:10.24432/c5vp6t](https://doi.org/10.24432/c5vp6t) ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/peng-2026-g-p2p-multi-source-approach/