About

What is published here

Datasets, models and papers for Cantonese speech, contributed by registered users and published under their own names. Every release carries a version, a licence, a checksum for each file and a ready-made citation. Its owner can replace files or correct details. The release keeps its address, and each file change is dated on the release page and on the Timeline. Earlier files are no longer offered for download, but their names and checksums stay listed on the release page; note the version and checksum of what you download.

Focus

The platform is organised around one goal: a Cantonese speech–language model that performs well in both directions — speech to text and text to speech — with Jyutping as the shared representation between sound and writing.

How to cite

Every release page shows an APA 7 reference and a BibTeX entry generated from its metadata: the authors as entered by the contributor, the year of the current version, the title, the version, CantoAI as publisher and the release's address. The version identifies what was used, so the APA reference carries no retrieval date; the BibTeX note records when the files last changed and the date of access. The BibTeX uses the @misc entry type, which every BibTeX and BibLaTeX style accepts, and states the type and version in the howpublished field, because most styles leave out a separate version field. The examples below show the three kinds.

Dataset

APA 7: Chan, T.-M., & Wong, M.-L. (2026). Example Cantonese read-speech corpus (Version 1.0) [Data set]. CantoAI. https://cantoai.org/releases/example-corpus/

@misc{chan2026example,
  author       = {Chan, Tai-Man and Wong, Mei-Ling},
  title        = {{Example Cantonese read-speech corpus}},
  year         = {2026},
  howpublished = {Dataset, version 1.0. CantoAI. \url{https://cantoai.org/releases/example-corpus/}},
  note         = {Accessed 2026-09-07}
}

Model

APA 7: Wong, M.-L. (2026). Example Cantonese speech-to-Jyutping model (Version 0.3) [Computer software]. CantoAI. https://cantoai.org/releases/example-model/

@misc{wong2026example,
  author       = {Wong, Mei-Ling},
  title        = {{Example Cantonese speech-to-Jyutping model}},
  year         = {2026},
  howpublished = {Model, version 0.3. CantoAI. \url{https://cantoai.org/releases/example-model/}},
  note         = {Accessed 2026-09-07}
}

Paper

APA 7: Chan, T.-M. (2026). Tone-aware evaluation of Cantonese speech recognition. CantoAI. https://cantoai.org/releases/example-paper/

@misc{chan2026example,
  author       = {Chan, Tai-Man},
  title        = {{Tone-aware evaluation of Cantonese speech recognition}},
  year         = {2026},
  howpublished = {Paper. CantoAI. \url{https://cantoai.org/releases/example-paper/}},
  note         = {Accessed 2026-09-07}
}

A paper that has also been published in a journal or conference should be cited by that venue; the platform copy then serves as the accessible version. The platform does not mint DOIs; contributors who need one can archive a copy with a DOI service such as Zenodo and add the DOI to the release description.

Contributing

Anyone can publish data, corrections, models or papers: create an account, then use Upload. Your release is public at once and recorded on the Timeline with your username and the date. Native speakers who would like to help check transcripts with the site's annotation tool, which is open to invited annotators, can write to the contact address above.

Human verification

The site has an annotation tool for human verification of resources published here. It is open to invited annotators: once added to a project, they find it under Annotate after signing in. Verification is optional. Contributors who have had a dataset checked are welcome to publish the details with the release — how the sample was drawn, the questions as the annotator saw them, and the resulting figures with their confidence intervals and sample sizes — so that readers can see for themselves which part of the data was checked and how.