CantoAI · Cantonese speech
Sign in
Dataset version 2.1 Wang, Yan · uploaded by tom · published 2026-09-10 files updated 2026-09-15

iCantonese pronunciation corpus

17.3 hours of Cantonese speech from one presenter, cut into 4,911 clips and labelled syllable by syllable with Jyutping (dictionary and acoustic readings) by an automatic pipeline. The published export holds 164,693 syllables; none has yet been checked by a listener.

Description

This is an unverified build. Every file in it is the output of a single pipeline run on the author's own machine, and no reading, timestamp or verdict has been checked by a listener.

The archive holds the corpus tables, the same corpus as a SQLite database, the 4,911 window clips and the 569 normalised full-length tracks in 16 kHz mono FLAC, the stage intermediates, the figures, the pipeline scripts and PIPELINE.md. 567 videos, 171,867 syllables over 17.3 hours of speech; the published export covers tiers A and B, which is 164,693 syllables in 4,439 utterances. Built with qwen3-asr-flash, ToJyutping 3.2.0, Qwen3-ForcedAligner-0.6B and hon9kon9ize/wav2vec2bert-jyutping.

The tables, the database, the figures and the scripts may be used and redistributed with attribution. The recordings are the work of the icantonese YouTube channel.

Files

FileTypeSizeSHA-256
icantonese_run_2026-09-15.zip archive · zip 3.2 GiB 155590609f063a7710e882233bc11220440b76e0f003ec37976f5b0652da5fa5
Download from the command line, and check the file

A large download can be resumed: run the same command again after an interruption.

curl -C - -o 'icantonese_run_2026-09-15.zip' 'https://cantoai.org/files/8559a7b7bdcc1e214ca707970f63ce20/icantonese_run_2026-09-15.zip'

Compare the SHA-256 above with:

sha256sum FILE                     # Linux
shasum -a 256 FILE                 # macOS
certutil -hashfile FILE SHA256     # Windows

Licence

See description

History

  • Dataset updated: iCantonese pronunciation corpus (2.0 → 2.1).
  • Dataset added: iCantonese pronunciation corpus (2.0).

Discussion

No discussion about this release yet.

Start a thread about this release

Citation

Generated from this release. The version identifies what you used; the BibTeX note also records when the files last changed and the date you accessed them.

Wang, Y. (2026). iCantonese pronunciation corpus (Version 2.1) [Data set]. CantoAI. https://cantoai.org/releases/icantonese-pronunciation-corpus/

@misc{wang2026icantonese,
  author       = {Wang, Yan},
  title        = {{iCantonese pronunciation corpus}},
  year         = {2026},
  howpublished = {Dataset, version 2.1. CantoAI. \url{https://cantoai.org/releases/icantonese-pronunciation-corpus/}},
  note         = {Files updated 2026-09-15; accessed 2026-10-23}
}

Citation guide →

Report a problem with this release · How reports are handled