What the work actually involves

Contributors record unscripted Japanese group conversations. Those recordings pass through an automatic transcription and diarization pass, then a human annotator corrects them. You are the last reviewer before the data becomes training or evaluation material. A typical session review means listening to multi-channel audio for bleed between channels, clipping, low gain and background noise; then reading the transcript against the audio to confirm it is genuinely unnormalized — every えーと, あの, stutter, false start, repair and backchannel present as spoken, with no silent grammar tidying. You check turn-level and word-level timestamps against what you hear, confirm speaker attribution through overlapping dialogue, and open the underlying JSON to confirm metadata tags, demographic markers and timestamp logic are coherent. You then assign pass/fail or agree/disagree and write comments specific enough that the annotator can act on them.

Safety and privacy auditing is part of every review: sessions containing PII, toxic or sensitive content have to be caught here, not downstream.

What the screen looks for

  • Native Japanese ear, not fluent Japanese. Screens probe disfluency handling, dialect and register, and how you transcribe things Japanese orthography gives you choices about — numerals, particles dropped in speech, 縮約形 like 〜ちゃう or 〜てる.
  • Verbatim discipline. The most common failure among excellent language teachers and translators is instinctive correction. Expect to be asked what you would do with an ungrammatical utterance.
  • Audio judgement. Whether you can describe the difference between clipping, low gain and room noise, and what channel bleed sounds like on a two-speaker recording.
  • Comfort in JSON. You do not need to code, but you need to read a diarization payload and notice that a segment's end timestamp precedes the next segment's start when the speakers clearly overlapped.
  • Reviewer temperament. Calibrated pass/fail decisions with written reasoning, not blanket approval or blanket rejection.

Logistics

Fully remote contractor assignment through Turing, roughly 15 weeks with possible extension based on performance and project need. The stated commitment is a minimum of 4 hours per day and 40 hours per week, including 4 hours of overlap with PST — a real constraint if you are based in Japan, where that window falls in the late evening or very early morning. No medical or paid leave. Pay is not disclosed in this listing; treat any rate you see quoted elsewhere as observed rather than promised. Shortlisted candidates are reviewed internally before an onboarding conversation.