What the work involves

Voice coaching for AI evaluation sits between language teaching and audio annotation. On a typical shift you listen to synthesised or human-recorded English speech and judge it against a rubric: segmental accuracy (which phonemes are wrong and how), stress and rhythm, intonation contours, pacing, breath and disfluency placement, and whether the delivery matches the intended register — a customer-service line, an audiobook passage, an emergency announcement. Where the rubric asks for it, you write the correction rather than just scoring it: an IPA transcription, a note that the model is placing primary stress on the wrong syllable of a compound noun, or a clean reference take recorded from your own microphone.

Expect a mix of task types across a contract. Some batches are pure rating with a forced-choice comparison between two audio clips. Some are coaching notes written to be read by a non-linguist engineer. Some are elicitation — reading prompt sets aloud with controlled variation, or coaching a speaker through retakes until the target prosody lands. Consistency between annotators matters more than elegance; calibration rounds and gold-standard checks are normal, and disagreement is usually resolved by pointing at the rubric rather than at your ear.

What the screen looks for

micro1 runs an AI-led interview before any project match. It probes whether your phonetics vocabulary is real or borrowed — expect follow-ups that push past "it sounded unnatural" toward what exactly was unnatural and how you would fix it. It also tests accent literacy: whether you can describe a variety of English without ranking it, since much of this work involves rating models on speakers who are not General American or RP. Coaching credentials (CELTA/DELTA, speech-language pathology, voice or dialect coaching, phonetics coursework) help, but demonstrated ability to articulate a diagnosis under questioning helps more.

Logistics

  • Remote and largely asynchronous, with occasional live calibration or coaching sessions scheduled across time zones.
  • Project-based hours; commitments listed on micro1 commonly land in the 10–20 hour per week range, though this varies by batch.
  • You supply your own quiet recording space and a headset or condenser microphone; audio submissions are rejected for room noise and clipping.
  • Pay is undisclosed on this listing and negotiated per engagement — treat any figure you see quoted elsewhere as unverified.