What the work involves

You audit other people's output rather than producing your own. In the Multilingual Transcription Audit stream, you listen to a recording of German speech, read the transcription an annotator submitted, and decide whether it faithfully captures what was said. Every deviation gets tagged with the applicable error codes — all of them, not just the first one you notice — followed by a pass or fail verdict and a short written rationale in English, with critical errors stated up front.

In the Forced Alignment Audit stream, you work at the word level on audio that an automated aligner has already segmented. For each segment you check that the start and end timestamps land on the real acoustic boundaries of the word and that the reference text matches the spoken form. Where the machine is wrong you fix it: nudge a boundary, split a run-on segment, merge two that cover a single word, insert a segment the aligner missed, or delete one containing no speech at all.

What the platform screens for

  • Language provenance. Native or near-native German as spoken in Germany — born and raised in such a region, or fully fluent with five or more years of residence there. This gate is not negotiable.
  • English as a working language. The rulebook, the error codes, and your rationales are all in English. Screens probe whether you can read a dense conventions document and write a reviewer-legible justification under time pressure.
  • Rubric discipline over instinct. Interviewers look for evidence you have applied a written standard hundreds of times consistently — transcription, subtitling, localization, linguistic annotation, or QA backgrounds read strongly here.
  • Acoustic judgment. The ability to separate genuine acoustic ambiguity (mumbling, overlap, coarticulation, elision in connected speech) from a clear annotator mistake, and to say which is which in writing.

Phonetics or linguistics training and familiarity with waveform or audio-editing tools help but are not required.

Logistics

Fully remote and asynchronous. Tasks are metered at 5.00 audit hours apiece, so this is deliberate listening work, not throughput work — expect to replay passages and consult the conventions rather than clear a queue. A short calibration phase precedes production work, and your agreement with the gold standard there determines whether you continue. Hours are self-scheduled; sustained availability matters more than long single sittings, because auditing quality degrades fast when you push past a few concentrated hours.