Data governance
Who owns the voice of a community
A national-language corpus is produced by a community of speakers. The question is therefore not only who may use it, but who should benefit from it, and how each recorded person keeps control of their voice.
The licence and its tiers
The framework chosen, the Nwulite Obodo Open Data License version 1.0 (NOODL-1.0), breaks with the usual open licences, which grant the same rights to an underfunded African laboratory and to a technology multinational.
It is settled for the nine published datasets.
Free and open access, with attribution to the Maloba project.
No royaltyFree and open access, with attribution, for any use.
No royaltyBenefit sharing to be agreed with CDS before use; the proceeds return to the project and to the Lingala-speaking community.
Benefit sharingConsent
Consent to the recording and to the use of the voice is collected on site, before recording, in the ODK Collect collection app. It covers the recording, its transcription and its publication in a corpus under the NOODL-1.0 licence, with no name and no contact details. CDS keeps it with the recording, outside the public catalogue.
Right of withdrawal
To withdraw a recording, write to contact@congo-digital.com giving the place and the approximate date of the recording. CDS acknowledges receipt within a week and removes the segments concerned from the next version of the corpus, within a month. Versions already downloaded by third parties are covered by their licence undertaking.
Speaker protection
- No name and no contact details in the published datasets; annotator identities are not included.
- Identifying the voices or cloning them is forbidden: that is a condition of access to the speech corpus.
- Controlled access to the speech corpus: every request is reviewed by CDS.
- Raw collection files and consent forms are never published.
Model licences
A model keeps the licence of its base model, whatever the licence of the training data.
| Model | Licence | Commercial use |
|---|---|---|
| Whisper Small + QLoRA | Apache-2.0 | Yes |
| MMS 1B + adaptateur lingala | CC-BY-NC-4.0 | No |
| Qwen2-VL 2B + QLoRA | Apache-2.0 | Yes |
| NLLB-Maloba-h1 | CC-BY-NC-4.0 | No |
| Llama 3.1 8B + QLoRA | Llama 3.1 Community License | Under conditions |
Origin of the recordings
Broadcasts from the partner rural radio stations (Radio Rurale · ORPR, signed agreement) and community sessions (Health Voice-a-thon), covered by the partnership agreements and by individual consent. The detail, recording by recording, is not published.
Where the licence is
Nwulite Obodo Open Data License, version 1.0: licensingafricandatasets.com. It is settled for the nine published datasets.