Speech data collection

How to brief an Arabic speech collection pilot

A buyer’s checklist for dialect screening, single-channel recordings, consent, audio quality and acceptance rules.

By the Annotvia AI team

Define one usable recording

“Collect Egyptian Arabic audio” leaves room for several incompatible deliveries. Is one item a prompted sentence, an uninterrupted speaker turn or a conversation with your model endpoint? Specify who speaks, what starts and ends an item, and whether the delivered file should contain only the participant’s voice.

Illustrative brief: 10 Egyptian Arabic speakers × 5 prompted recordings = 50 planned clips before QA. One speaker per file. Each clip is accepted or rejected independently. A retake replaces the rejected clip; it does not count as a second accepted item.

Recruit for the audience you need

Ask separately about native dialect, where the participant grew up and current country of residence. A Cairo speaker living abroad might be exactly right for a dialect test and ineligible for an in-country device study. Agree on what is self-reported, what must be checked and which data the buyer actually needs.

Set the capture rules before recruitment

Specify file format, sample rate, channel count, recording environment and device requirements. If participants speak to a model, test whether its voice leaks back through the microphone. If a study is supposed to reflect everyday phone use, do not reject every trace of room noise unless that rule was in the brief.

Review one small batch together

Have the buyer and QA lead listen to the same clips. Label the reason for each rejection: missing words, clipped speech, another voice, wrong prompt, or technical failure. Compare judgments on borderline examples. If the brief changes, record which batch uses the new rule.

Agree on handoff and payment units

Decide whether the quote is per accepted clip, accepted audio minute or completed participant. Fifty clips is not fifty people, and a minute spent reading instructions is not a minute of delivered speech. Agree how consent is recorded, who can access recordings, how long they are retained and who authorizes deletion.

What comes back from the pilot

Ask for accepted files or hosted-study completion records, a manifest with non-identifying participant IDs, rejection reasons and actual time taken. Use those observations to plan the next batch rather than assuming all invited speakers will complete every prompt.