> For the complete documentation index, see [llms.txt](https://ai4commsci.gitbook.io/formosanbank/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://ai4commsci.gitbook.io/formosanbank/the-bank-architecture/developers/repository-contracts/audio-publication-policy.md).

# Audio publication policy

{% hint style="warning" %}
**Synced copy — do not edit here.** The canonical file is [`AUDIO-PERMISSIONS.md` in the FormosanBank repository](https://github.com/FormosanBank/FormosanBank/blob/main/AUDIO-PERMISSIONS.md); it versions together with `audio_permissions.json` and the parity checks that enforce it. Edit it there, then run `python sync_upstream_docs.py` in this repository.
{% endhint %}

## Audio Publication Policy

FormosanBank uses the public repository as the publication boundary:

* If a corpus's XML is published under `Corpora/` in FormosanBank, its associated audio is public under the same license recorded in the XML.
* Publication permission is recorded on the corresponding Basecamp corpus card.
* Audio from a private development repository stays private and is not licensed for reuse.
* When a development corpus is published into FormosanBank, its audio access and license must be updated at the same time.

`audio_permissions.json` records this state for every FormosanBank Hugging Face audio repository. `audio_sources.json` lists the 21 canonical public datasets used by the download scripts. The 16 per-language ILRDF repositories are public compatibility mirrors, not additional canonical downloads.

### License versions

Use the latest version of the license named in the XML unless the publication permission comes from material already released under a specific Creative Commons version. In that case, retain the source's version.

### Published audio

| Published XML                           | Audio license   | Hugging Face repositories                                    |
| --------------------------------------- | --------------- | ------------------------------------------------------------ |
| ILRDF dictionaries                      | CC BY-NC 4.0    | Aggregate dataset plus 16 language mirrors                   |
| NTU Formosan Corpus grammar             | CC BY-NC 4.0    | Grammar dataset                                              |
| NTU Formosan Corpus stories             | CC BY-NC 4.0    | Stories dataset                                              |
| NTU Paiwan ASR                          | CC BY 4.0       | Two published audio datasets                                 |
| Paiwan Stories                          | CC BY-NC 4.0    | One dataset; the three WAV files also remain in FormosanBank |
| Apay Tang / PARADISEC Taroko recordings | CC BY-NC 4.0    | One dataset                                                  |
| John Whitehorn / Cambridge recordings   | CC BY-NC 4.0    | One dataset                                                  |
| Wilang Yutas recordings                 | CC BY-NC 4.0    | One dataset                                                  |
| Yedda Palemeq recordings                | CC BY-NC 4.0    | One dataset                                                  |
| ePark published topics                  | CC BY-NC-SA 4.0 | Eleven topic datasets                                        |

The canonical download contains 525,770 audio files. Files listed in `audio_extras.json` belong to the same published corpora and use the same licenses.

### Private development audio

These repositories remain private and are not licensed for reuse:

* `FormosanBank/NTU_Paiwan_Raw`
* `FormosanBank/ePark3_Raw`

Raw or working audio from any other private Formosan-X development repository follows the same rule.

### Publishing a development corpus

When a corpus moves from a private development repository into FormosanBank:

1. Confirm publication approval on its Basecamp card.
2. Add the final XML and its `copyright` license to `Corpora/<corpus>/`.
3. Change the associated Hugging Face audio from private to public.
4. Set the dataset-card license to match the XML, using the version rule above.
5. Add the repository to `audio_permissions.json` and, if canonical, `audio_sources.json`.
6. Run `python QC/validation/validate_hf_audio.py` anonymously.

CI checks that published repositories are public, development repositories are private, and every canonical dataset matches its XML audio contract.