The work

You listen to long stretches of text-to-speech audiobook narration in Dutch as a real listener would — but with a reviewer's discipline. When something slips, you mark the exact span in the review tool, classify the issue type (skipped word, inserted word, mispronunciation, number read wrong, abbreviation expanded incorrectly, unnatural phrasing or prosody, glitch or artifact), and add a short note in English explaining what a human narrator would have done instead. Alongside span-level tagging you give holistic judgements: where the narration held the illusion, where it collapsed, and whether you would keep listening.

The hardest part is not hearing the errors — it is hearing them consistently on hour three. Dutch throws up specific traps: the ij/ei contrast, final devoicing, stress in separable verbs (vóórkomen vs voorkómen), loanword pronunciation, dates and ordinals, currency, and abbreviations like bijv., o.a., d.w.z., n.a.v. that a model may spell out, misexpand, or read as a word. Belgian-Dutch or over-anglicised vowel realisations in a Netherlands-Dutch narration are also reportable.

What the screen looks for

Mercor's AI-led interview probes whether your Dutch judgement is real and articulable. Expect follow-ups asking you to name specific pronunciation failures you'd expect from TTS, to distinguish a genuine error from a stylistic choice you merely dislike, and to explain how you'd label something ambiguous rather than skip it. Vague enthusiasm for audiobooks scores poorly; concrete examples, named titles and narrators, and a clear rule for your own consistency score well. Prior annotation, transcription, proofreading, subtitling or editing work is a strong signal — say so plainly and describe the guideline you worked to.

Logistics

  • Remote and location-flexible, though candidates living in the Dutch-speaking Netherlands are preferred.
  • Roughly 20 hours per week, commonly worked as ~4 hours per day across five days.
  • Largely async, with written English used for reports, guideline questions and calibration threads.
  • Requires a quiet listening environment and decent headphones — artifact detection is unreliable on laptop speakers.
  • Pay observed at $15–20/hr; rates and hours are set per project and are not guaranteed.