What the work actually involves

You receive batches of Spanish-language audio — recordings that may be spontaneous speech, regionally accented, overlapping, noisy, or code-switched between Spanish and English — and produce transcripts that conform to a written style guide. That guide dictates things most people never think about: whether to transcribe pa' as spoken or normalize to para, how to tag inaudible segments, how to mark speaker changes, whether filled pauses and false starts are kept. A second stream of work is review: correcting someone else's transcript and documenting why. Annotation sits alongside — labelling speaker attributes, audio conditions, language switches, and other metadata the customer's model needs.

The consistent failure mode on this kind of project is not poor Spanish. It is drifting from the spec — silently fixing a speaker's grammar, guessing at a muffled word instead of flagging it, or applying one convention in the morning and a different one after lunch. Transcription for model training is not transcription for readability.

What the screen looks for

  • Dialect range. Which varieties you genuinely handle — rioplatense, caribeño, andaluz, Mexican regional, Andean — and which you would decline rather than guess at.
  • Verbatim discipline. Whether you can articulate the difference between clean-verbatim and true-verbatim and say which one a given instruction implies.
  • Ambiguity handling. What you do with an unclear three seconds: flag it, timestamp it, escalate it, or invent something plausible.
  • Throughput realism. Honest minutes-of-audio-per-hour figures for difficult audio, not flattering ones.

Logistics

Remote and contractor-based. Work arrives as assignments with deadlines rather than fixed shifts, so hours are largely self-scheduled, though batches can be time-sensitive and volume fluctuates with customer demand. You will need a quiet environment, reliable internet, and good headphones; some transcription tools are browser-based with foot-pedal or keyboard-shortcut support. Observed pay for this listing runs $20–36/hr, typically banded by dialect coverage, review-tier work, and demonstrated quality scores — not guaranteed, and rates on AI data projects move with the project.