Twi Speech and Transcription Contributor

Transcribe short Twi audio clips with speaker turns and code-switch tagging. Asante and Akuapem Twi both welcome. Remote, flexible hours, paid per accepted clip.

Bilingual annotation is the work of deciding what text or speech actually means, then recording that decision in a form a model can learn from. For Twi that covers labelling intent and sentiment, checking whether a translation preserves meaning rather than words, transcribing speech accurately, and judging which of two model outputs a native speaker would actually accept.

Asante and Akuapem Twi, plus the code-switching with English that is normal in real speech. Corpshore treats language varieties as distinct rather than flattening them, because a model trained on a flattened dataset fails precisely on the speakers who were flattened.

Corpshore runs annotation, evaluation and data collection for AI teams across 18+ countries in 35+ languages. Work is assigned by qualification and availability, not by location and every task shows its rate before you reserve it.

What you would be doing

• Label and review Twi text or audio against a project rubric
• Judge whether Twi model outputs are natural, accurate and appropriate in register
• Flag ambiguous cases rather than guessing, so the rubric can be corrected
• Keep quality steady across long batches, which is harder than it sounds

What we look for

• Native or near-native Twi
• Strong written English, since rubrics and reviewer feedback are in English
• A reliable internet connection and a device you can work on comfortably
• Ability to follow a written rubric precisely, including its edge cases
• Willingness to have your work reviewed, and to act on that review

Helpful, not required

• Formal training in Twi linguistics, translation or teaching
• Familiarity with a specific regional variety and its differences
• Previous annotation, transcription, moderation or research experience
• Comfort with spreadsheets or annotation tooling
• A subject you know unusually well, whatever it is

How applying works

1. Apply once. One application is matched against every role you qualify for, so you do not apply to these one at a time.
2. A person reads it. No automated screening and no automated rejection. We reply either way.
3. A short recorded interview. Once approved, you answer a few questions to camera. Your audio is transcribed and scored against a rubric, and a member of staff makes the decision. Nothing is decided by software.
4. Work opens. Tasks appear in your queue with the rate shown before you reserve anything.