Define roles operationally
Choose the roles relevant to your workflow and explain how a reviewer can recognize them from supplied evidence. Caller and agent may be enough for a simple two-person interaction, while transfers or interpreted calls need additional roles. Include unknown when the recording or source metadata cannot support a confident assignment.
Keep roles separate from identifiers
A speaker identifier groups segments attributed to one voice within a defined scope. A role describes that participant’s function. The same role can be occupied by multiple speakers in one call, and source-provided agent identifiers may have a different scope from annotation labels created during preparation.
Record metadata origin
State whether each field came from the source system, a reviewer, or a derived process. Avoid treating business location, phone region, or an annotator’s impression as verified personal attributes. Collect only the metadata needed for the task and include uncertainty and missingness conventions in the dictionary.
Check the difficult cases
Inspect conferences, handoffs, voice prompts, and participants speaking on behalf of someone else. Confirm that role labels remain aligned with diarization segments after edits. Share the role schema, evidence requirements, and handling limits with Fiund so a prospective source can be reviewed against the actual project rather than a generic speaker-metadata list. For a role-label check, select an interpreted conversation where one participant speaks on behalf of another. Ask reviewers to distinguish the acoustic speaker from the represented person and the business role. Document which distinctions are supported by the source and which remain uncertain, rather than forcing every voice into a caller-versus-agent pair.
Review checklist
- Define roles with observable evidence
- Separate speaker IDs from role labels
- Record metadata source and uncertainty
- Review transfers, conferences, and prompts