Zimbabwe Language
Data Commons.
A governed, consent-based, AI-ready Shona and Ndebele speech-and-text corpus — the missing data foundation for Zimbabwean-language AI.
The problem
Zimbabwe has no governed, AI-ready corpus for its own languages. The AI systems deployed for Zimbabweans — e-government assistants, telemedicine triage, screen readers, chatbots — are trained overwhelmingly on English and foreign data, and speakers of local languages get measurably worse outcomes from digital services.
No AI team can build, evaluate or audit local-language systems while the underlying data does not exist. That absence is the problem this project solves.
What we'll build — pilot commitments
| Deliverable | Pilot target |
|---|---|
| Consented, transcribed Shona/Ndebele speech (≥400 distinct speakers) | 100+ hrs |
| Cleaned, provenance-tracked Shona/Ndebele text corpus | 1,000,000+ words |
| Health-triage vertical — scenario dialogue speech + phrase inventory | ≥15 hrs |
| Accessibility vertical — PWD & assistive-technology speech | ≥10 hrs |
| Held-out, double-transcribed evaluation benchmark, per language | ~5 hrs |
Everything ships AI-ready: a published schema and data dictionary, a per-file checksummed manifest, versioned releases, quality reports, a bias/representativeness register — and an evaluation layer that lets any deployed AI service's Shona/Ndebele accuracy be benchmarked and audited.
How the data is collected
Primary, consented, original data through four institutional channels — no web scraping, no single-institution dependence:
- 🏢
Community Information Centres
Supervised recording days at POTRAZ/USF centres, reaching rural and peri-urban speakers in every province.
- 🎓
State university language departments
The academic anchor — transcription standards, orthography authority, annotation QA, and trained fieldworkers.
- 📻
Community radio stations
Purpose-made consented studio sessions with presenters and community members, plus licensed archive segments where rights are clearly held.
- ♿
Disability organisations
Co-design and host the accessibility vertical — persons with disabilities represented as speakers and co-designers, not subjects.
Sampling uses documented quotas across province, district, urban/rural, gender, age and dialect region — collection sites in ≥6 of 10 provinces in the pilot, with ≥40% of speech hours from rural or peri-urban sites, and gaps measured and disclosed.
Consent, privacy & governance
The lawful basis for all primary collection is informed consent under Zimbabwe's Data Protection Act [Chapter 12:07] — plain-language consent in the participant's own language, with withdrawal honoured in every release.
Speakers are pseudonymised, identifiers redacted, and location generalised to district level; raw audio and consent registers sit under encryption and role-restricted access.
Hard boundary: the corpus is barred from surveillance, enforcement, profiling, credit denial and speaker-identification uses. Voice data is collected for language technology only. A named data steward and a governance group drawn from the collection partners approve every release, licence and takedown.
Access & sustainability
Open research tier
Anonymised core corpus under CC BY 4.0 for researchers, students and public-interest use.
Controlled-access tier
Sensitive subsets (e.g. health-triage) under logged access agreements.
Commercial tier
Paid licences for enterprise users fund ongoing stewardship and twice-yearly refresh releases.
Status & partners
Proposal submitted 3 July 2026 to the POTRAZ 2026 AI for Impact Challenge (Track 1: Data); institutional partners are being confirmed during the application window. IndabaX Zimbabwe has provided its signed letter of support.
We're looking for university language departments, community radio stations, disability organisations and the CIC network to join as partners — a short letter of support on your organisation's letterhead is all we ask to start.