The work

You are not producing transcripts — you are grading them. Two collections run in parallel. In Multilingual Transcription Audit, you listen to a recording of Korean speech, read the transcription an annotator submitted, and decide whether it faithfully captures what was said. You tag every applicable error code (not just the first one you hit), issue a pass or fail, and write a short rationale in English naming the critical error first so a reviewer can see the reason for failure without replaying the file. In Forced Alignment Audit, you review audio already segmented word by word by an automated system, checking that each start and end timestamp lands where the word actually begins and ends and that the reference text matches the spoken form. Where the machine is wrong you fix it: move a boundary, split a segment, merge two, insert a missing one, or delete an empty one.

Much of the judgment sits in the gap between genuine acoustic ambiguity and clear annotator error. Fast connected speech, particle elision, regional colouring, crosstalk and clipped audio all produce cases where a reasonable transcription is defensible. The rubric decides those cases, not your instinct — and the platform screens hard for people who can apply a written standard the same way on judgment 400 as on judgment 4.

What the screen looks for

  • Native or near-native Korean as spoken in South Korea — born and raised there, or fully fluent with five-plus years of residence. Stated as a hard requirement.
  • English reading and writing at working strength. The rulebook, the error codes and your rationale are all in English. A rationale that a reviewer has to reread is a failed rationale.
  • Evidence of rubric discipline from transcription, subtitling, localization, linguistic annotation or QA work — expect follow-ups asking how you handled a case where the written standard disagreed with your ear.
  • Ear-level detail: where word boundaries fall in continuous speech, what a 60ms early onset sounds like, when a segment contains breath rather than a word.

Logistics

Fully remote and asynchronous. Tasks are budgeted at 5.00 audit hours each, so this is deep, sustained listening rather than piecework — plan sessions in blocks and expect a short unpaid-to-low-volume calibration phase before production access opens. Observed pay on this listing is $25/hr; rates on Mercor are set per project and are not guaranteed. Quiet workspace and reliable closed-back headphones are effectively mandatory; waveform-tool familiarity (Audacity, Praat, ELAN) helps on the alignment collection but is not required.