g2pW for Cantonese

A Cantonese Grapheme-to-Phoneme (G2P) conversion tool, adapted from the original g2pW repository for Chinese Mandarin.

Overview

This is a Cantonese adaptation of the g2pW repository, designed for polyphone disambiguation in Cantonese. The model converts Chinese characters (graphemes) to Jyutping (phonemes).

Model

The trained model is available at:
https://huggingface.co/Naozumi0512/g2pW-canto-20241206-bert-base

Usage

git clone https://github.com/Naozumi520/g2pW-Cantonese && cd g2pW-Cantonese
git switch 20241206-bert-base
git clone https://huggingface.co/hon9kon9ize/bert-base-cantonese
git clone https://huggingface.co/Naozumi0512/g2pW-canto-20241206-bert-base G2PWModel-v2-onnx
pip install -r requirements.txt

from g2pw import G2PWConverter

conv = G2PWConverter(model_dir='./G2PWModel-v2-onnx/', model_source='./bert-base-cantonese/', use_cuda=True)
# "校":
sentence = '調校溫度'
print(conv(sentence)) # [['tiu4', 'gaau3', 'wan1', 'dou6']]
sentence_1 = '校服'
print(conv(sentence_1)) # [['haau6', 'fuk6']]

Data Sources

This model was trained on data collected from:

• Rime Cantonese Input Schema (jyut6ping3.dict.yaml)
• 粵典 Words.hk
• CantoDict

Acknowledgements

This work builds upon the original g2pW model. I would like to express my sincere gratitude to:

• The original g2pW authors for their groundbreaking work
• The rime-cantonese Project contributors for their extensive Cantonese lexicon
• Words.hk team for their comprehensive Cantonese dictionary
• CantoDict team for their valuable Cantonese resources
• Our community hon9kon9ize for hon9kon9ize/bert-base-cantonese

This project is based on the g2pW model:

@inproceedings{chen22d_interspeech,
  title     = {g2pW: A Conditional Weighted Softmax BERT for Polyphone Disambiguation in Mandarin},
  author    = {Yi-Chang Chen and Yu-Chuan Steven and Yen-Cheng Chang and Yi-Ren Yeh},
  year      = {2022},
  booktitle = {Interspeech 2022},
  pages     = {1926--1930},
  doi       = {10.21437/Interspeech.2022-216},
  issn      = {2958-1796},
}

For the Cantonese adaptation specifically, please also acknowledge the data sources mentioned above.

License

https://github.com/GitYCC/g2pW/blob/master/LICENCE
https://github.com/rime/rime-cantonese/blob/main/LICENSE-CC-BY
https://words.hk/base/hoifong/