More than
a dictionary.
Taishanese (台山話) connects families and communities across generations. This project starts with the words and voices of its speakers.
A path toward live translation
- Collect and review. Build a dictionary of useful words and phrases, with aligned Taishanese text, English meanings, regional context, and recordings.
- Prepare a speech corpus. Collect multiple speakers, add accurate utterance transcripts, normalize audio, and measure recording quality. Keep evaluation speakers separate from training speakers.
- Train and evaluate. Assess speech recognition, text translation, and speech generation with Taishanese speakers. Publish measured accuracy and limitations.
- Connect live speech. Add streaming recognition, translation, and spoken output once each component is dependable.
The current app supports the first step and exports reviewed contributions for corpus preparation. It is a reviewed phrase lookup. It does not translate sentences, recognize speech, or generate speech yet, and it does not present Mandarin or Cantonese machine translation as Taishanese.
Privacy & contribution terms
We store your sign-in provider ID, account name, and verified email to manage contributions. Your email is not published. Contributions, regional descriptions, recordings, and review decisions are stored using Cloudflare D1 and R2.
By submitting, you confirm you have the right to contribute the text and that an attached recording is your own voice. You give this project permission to host, publish, copy, adapt, and reuse your contribution to develop dictionaries, language datasets, and translation and speech models. This permission is non-exclusive; you retain your rights.
Approved text and approved recordings become publicly accessible. Pending and rejected recordings are accessible only to their contributor and project administrators. An audio recording can identify its speaker even without a name. Do not include private conversations or other people’s voices.
To request account or contribution removal, contact the project operator. Removal from the project cannot recall copies already made by others or undo completed model training.
No third-party dataset license is granted by these terms. The operator must settle redistribution and licensing terms with the community before publishing a training corpus.