Define the role of sound
Audio can provide instructions, conversation, or sounds associated with tool use and object interactions. Decide which of these your model will receive and which labels depend on them. A visual-only benchmark should not quietly rely on information available only through speech. Ask reviewers to distinguish events visible in the image from events supported by sound alone.
Review recording conditions
Inspect examples from quiet and noisy activities, including movement that may affect microphones. Check whether relevant sounds are audible and whether automatic processing changes their character. State whether separate channels or microphone positions matter. A track existing in the file does not establish that it provides useful supervision or that it was recorded consistently across sessions.
Address timing and derivation
If audio and video are delivered separately, request their time relationship and verify alignment using identifiable events. Document trimming, channel changes, and transcription links. An edited transcript may omit hesitations or background speech, so define whether text is intended as a readable summary or a time-aligned representation of what was actually recorded.
Specify capture and handling boundaries
First-person microphones can capture people and conversations outside the intended activity. Describe exclusions and the review process before collection. If audio must be removed, confirm that all requested outputs follow that decision, including derived clips and previews. The proposed use and permission scope should be reviewed for the specific source rather than inferred from video availability.
Review checklist
- State whether models receive audio
- Identify sound-dependent labels
- Verify alignment and transcript meaning
- Define exclusions across derived outputs