What the work is
The role
We are hiring native and near-native Italian speakers to audit AI training data in their own language, not to produce it.
You will work across two Amazon Sonic audit collections:
- Multilingual Transcription Audit. You listen to a recording of human or agent speech, read the transcription an annotator produced, and judge whether it accurately captures what the speaker actually said. You apply a fixed set of error codes, assign a pass or fail verdict, and write a short rationale explaining the decision.
- Forced Alignment Audit. You review audio that has already been segmented at the word level by an automated system. For each segment you verify that the start and end timestamps match where the word actually begins and ends, and that the reference transcription matches the spoken form. Where the machine got it wrong, you correct the boundary, split a segment, merge two, add a missing one, or remove one that contains no word.
Both collections run at 5.00 audit hours per task. Every judgment carries a written reason, so this is careful listening work rather than volume work.
Who gets hired
Native or near-native command of Italian as spoken in Italy. You were either born and raised in a region where this language and variety is dominant, or you are fully fluent with five or more years of residence in such a region. This is a hard requirement.
Strong English reading and writing. All conventions, error codes, and audit rationale are written in English, so you need to read a detailed rulebook and write clear feedback in it.
Disciplined rubric application. The work rewards people who apply a written standard consistently across hundreds of judgments rather than relying on instinct. Prior transcription, subtitling, localization, linguistic annotation, or QA experience is a strong signal.
Careful listening. You can distinguish similar sounds, identify where one word ends and the next begins in continuous speech, and tell genuine acoustic ambiguity apart from a clear annotator error.
Helpful but not required: prior AI data annotation or model evaluation work, a phonetics or linguistics background, and experience with audio editing or waveform tools.
What the work looks like
You listen, judge, and submit. Where you find an error you tag every applicable error code, not just the first one, and you explain in plain language why the task passed or failed. Critical errors are called out first so the reason for a failure is obvious to the reader.
You will complete a short calibration phase before production work begins.
The platform posted this on 18 Sep 2026.