What the work actually is
You receive batches of recorded English conversations — speakers with Nigerian, Kenyan, Ghanaian, Tanzanian, South African or Afrikaans-influenced accents — and produce verbatim, timestamped transcripts against a style guide. The audio is used to build and evaluate AI voice agents, so the transcript is ground truth: a dropped filler word, a guessed name, or an inconsistent timestamp offset propagates straight into model evaluation. Alongside transcription you flag what the audio itself cannot support — crosstalk, clipping, background noise, segments where speaker attribution is genuinely ambiguous — rather than quietly resolving them. Some contributors may also be asked to help produce conversations that others will later transcribe.
What the screen looks for
Mercor's application is AI-led and asks you to declare which accent group you're applying for; the Afrikaans group requires actual Afrikaans fluency, not familiarity. Beyond that, screeners probe for evidence of real transcription work — tools you've used, conventions you've worked under (verbatim vs. clean read, speaker labels, inaudible tags), typical turnaround, and how you handled QA passes. Expect follow-ups that test whether you can distinguish accent features from errors: a speaker's realisation of a vowel or a code-switched Yoruba or Afrikaans word is not a mistake to be normalised away.
- Confident comprehension of the specific accents you claim, not English generally
- Consistency across a whole batch, not one polished sample
- Judgment about when to tag unclear speech instead of guessing
Logistics
Remote and asynchronous, hourly, with pay observed in the $15–20/hr range — bands vary by project and are never guaranteed. Expect around 20 hours total across one to two weeks, with the possibility of more if the batch extends. Reliable internet, decent headphones and immediate availability matter here; this is a short-deadline engagement and slow starters are simply replaced.