Define the conversational phenomena
Separate simultaneous speech from fast back-and-forth turns, brief acknowledgments, and a speaker restarting after interruption. Specify which behaviors matter for the task. A dataset can have many short turns without substantial overlap, so ask for actual examples rather than accepting conversational as a sufficient description.
Choose representation rules
Decide whether labels support two active speakers at the same time and how overlapping words align to timestamps. State whether uncertain words are retained with confidence markers or excluded from a particular score. Keep the audio available for review because a flattened transcript can conceal what the model actually heard.
Inspect channel effects
Native separated channels may make some overlapping speech easier to inspect, while mixed audio can contain competing voices that are hard to resolve. Record channel configuration and leakage so evaluation slices remain interpretable. Diarization annotations alone do not provide the same information as original isolated microphone or channel recordings.
Measure and review the slice
Report overlap coverage by duration, call count, and speaker combinations when those measures are available. Listen to representative hard cases before adopting the labels. Send Fiund your overlap definition, annotation format, and channel requirements so a proposed source can be assessed against the actual conversation behaviors you need. A useful comparison includes both a short acknowledgment spoken over another participant and a longer interruption that changes the conversation. Review how each is represented and scored. The distinction can matter for turn timing and attribution, so the annotation policy should explain the expected handling rather than leaving both cases to individual reviewer preference.
Review checklist
- Distinguish overlap from rapid turns
- Define simultaneous-speaker labeling
- Record channel setup and leakage
- Review hard cases against original audio