A diarization dataset needs more than a transcript with speaker names. Specify turn boundaries, overlapping speech, and the scope of speaker identifiers.
Define speaker and role labels
Decide whether identifiers are local to each recording or need a reviewed relationship across calls. Keep caller, agent, supervisor, and interpreter roles separate from acoustic speaker labels. Transfers can introduce several voices with the same business role, so include examples of the participant structure your system must handle.
Represent difficult timing cases
Write rules for short acknowledgments, laughter, interrupted words, simultaneous speech, and uncertain boundaries. Choose an annotation format that can express the required behavior. If overlap is simplified or excluded, record that limitation explicitly so later evaluation does not imply coverage that the references cannot support.
Inspect channel configuration
Request the original channel map and listen to each channel separately. Two-channel audio may contain a mixed conversation rather than isolated participants. Diarization labels and physical channel separation are different artifacts; specify which one the model needs and what leakage or missing-side behavior is acceptable.
Review a concrete pilot
Listen to difficult turns while inspecting the proposed labels and timing. Send Fiund examples, grouping rules, source-review requirements, and an acceptance plan. Suitability and permitted scope need assessment for each potential source, including any participant metadata and sensitive speech retained in the project.
Include in your inquiry
- Speaker-ID scope and role schema
- Overlap and boundary rules
- Original channel-map requirements
- Pilot listening and source-review plan