Trust
Provenance & consent
Every recording in the BantuNomics dataset comes from a named, paid, consented contributor on the amina platform (amina.ai). Consent is not a checkbox bolted onto a scrape — it is the gate the entire collection pipeline is built around. This page documents how.
1. How contributors are recruited
Registration is invitation-gated. A contributor signs up for exactly one country + language pair, validated against the canonical Bantu language registry, and provides demographics (gender, age band, education, dialect region) used only as anonymized linguistic metadata. New contributors are pending approval and see only the qualifying task until an operator vets their audio quality — they are not auto-authorized for paid work.
2. The consent instrument (v2.0)
Before any account is usable, the contributor must read and accept the v2.0 BantuNomics Data Contributor Consent. It is all-or-nothing — every term must be accepted to participate — and it explicitly grants the uses BantuNomics licenses to partners:
- internal research and AI/speech/language tool development;
- commercial licensing of de-identified derivative datasets to third parties;
- voice cloning, synthesis, and text-to-speech training, by BantuNomics and licensees;
- public / open-source release of derived datasets;
- use of de-identified excerpts in educational and promotional materials.
The instrument grants a perpetual, worldwide, royalty-free, sublicensable, transferable license over the contributor's recordings and associated data for those purposes (§3), states that contributors are paid for quality-passing work (§4, via Western Union, at per-task rates shown in-app), confirms contributors are ≥18 and participating voluntarily (§6), and describes storage and rights (§5–§6). The full versioned text is available in the partner data room.
3. The per-consent audit trail
Consent is recorded as a durable, queryable record — not merely a boolean. Each signed consent stores:
So for any speaker we can show exactly what document they agreed to, which version, when, from what IP/agent, and the materialized scope grants that downstream export filters read.
4. Consent is per-language
Licensing scope is per-language, so consent is too. A contributor who consents to commercial use of their Zulu recordings has not thereby consented for Xhosa; a contributor recording a second language re-signs the v2.0 instrument for that language before any task work. This keeps the rights chain clean at the granularity we license.
5. The export gate (live publishable allowlist)
Only contributors on the consented public-speaker allowlist
(abs_public_speakers) are ever exportable. The allowlist is maintained
live: it is derived from per-language consent + acoustic QC, and it is
tombstone-only — a contributor who withdraws or is deleted is marked
status=removed (never hard-deleted), and this site reads active speakers
only, so withdrawal takes effect immediately across coverage, the player, and the API.
This is why the audio counts here are consent-clean: the 14
languages with audio reflect only active consented speakers
(10,498 clips across 53 speakers). Identities
appear only as anonymized public IDs (e.g. BEM-A04) with coarse demographics — never
raw names, emails, user IDs, or file paths. See the live roster on
Speakers.
6. The licensing chain
The rights flow end-to-end:
contributor (paid; v2.0 §3 commercial grant) →
BantuNomics / 3 Mega LLC (de-identification + packaging) →
licensee (Founding Partner). Each licensed delivery emits a
manifest — the public_ids, consent versions, gate criteria, and date —
persisted as the auditable rights-chain record. The consent scope and the per-language key travel
with the data. An independent legal review of the v2.0 instrument and this rights chain is provided to
partners under diligence. For the full pipeline with live numbers, see
How it works, end to end.