This page shows the shape of future Voice Bench SA tasks before any
live contributor collection. No upload happens on this page, and these
examples are not production data, paid work, or a claim of live users.
The goal is to make each task tiny, private-detail-free, reviewable,
and useful for South African language evaluation.
If someone wants to help, the contributor page asks only for source,
18+, preferred explanation language, language or accent, role, and no
task before consent before any live task can begin.
Make a 5-15 second voice note in the requested language or accent.
Example: say a simple greeting naturally, without names, addresses,
ID numbers, bank details, or private stories.
Task version
voice-simple-greeting-v0
Language tags
xh, af, en-ZA, mixed
2. Type it
self_transcript_exact_words_v0
Type exactly what was said. The task is not to improve the sentence
or make it sound formal. South African wording, pronunciation, and
ordinary local phrasing should stay visible.
Output
Exact transcript plus optional correction note
Guard
Private-content hold blocks export
3. Check it
Independent review
A different person reviews whether the audio, transcript, and
correction match. The final reviewer cannot be the speaker or the
correction worker.
Decision
accept, needs_more_review, or reject
Benchmark rule
Self-reviewed rows stay out of export
Readiness gates
Checked before live collection
Consent, 18+ confirmation, points, withdrawal, and STOP or DELETE handling are explained first.
WhatsApp warm-up and provider health must pass before outbound task routing.
Human/native local-language review is still required before non-founder isiXhosa outreach.
Accepted export rows need expected_language, task_version, provenance, consent, correction, and review.
Backed by npm run contributor:task-library:check, npm run contributor:first-task:check, and npm run data:quality:check.