Choose the speaker unit
Diarization describes who spoke when within the recording. Decide whether speaker labels are local to each call or need a reviewed relationship across calls. An agent role is not the same thing as an acoustic speaker identity, particularly when a call transfers between several agents.
Define segment boundaries and overlap
Write rules for short acknowledgments, interrupted words, laughter, silence, and simultaneous speech. State whether overlap can contain multiple active speakers or must be represented through a simplified format. Include difficult examples so reviewers understand how to label rapid turn exchanges rather than smoothing them away.
Separate role assignment from timing
Store the speaker segments and any caller, agent, interpreter, or supervisor roles as distinct information. Mark uncertainty when the recording does not support a confident role assignment. A transcript containing speaker names does not prove that turn boundaries are accurate, and diarization labels do not create isolated audio tracks.
Plan a reviewed pilot
Listen to a sample with the proposed labels visible and inspect transfers, overlap, and long silences. Define the annotation format and review criteria before discussing scale. Share these requirements with Fiund, alongside language coverage and source-review needs, so a potential collection can be assessed against your actual diarization task. Include a call where an agent transfers the customer to a colleague without a long pause. Review whether the reference introduces a new speaker, preserves the caller’s identifier, and updates roles independently. This concrete example can expose inconsistencies hidden in a simple two-speaker demonstration with long, clearly separated turns throughout the recording.
Review checklist
- Choose per-call speaker identifiers
- Specify overlap and boundary rules
- Keep role labels separate from segments
- Review difficult turns against the audio