Purpose
Improve speaker attribution without pretending diarization is certainty.
When to use it
- Transcript labels are Speaker 1, Speaker 2, or inconsistent names.
- Action items depend on who said what.
- A meeting summary would be risky if owners are misattributed.
Inputs
- Normalized transcript segments
- Candidate attendee list
- Known names, roles, and optional voice or phrase cues
Outputs
- Speaker label mapping
- Attribution confidence
- Review-required utterances
Steps
- Group transcript segments by original speaker label.
- Look for self-identification, direct address, role references, and calendar roster clues.
- Assign only high-confidence mappings automatically.
- Keep ambiguous labels unresolved and attach review evidence.
- Apply mappings to downstream output with confidence annotations.
Quality checks
- No label is mapped from attendance alone.
- Every mapped speaker has at least one evidence note.
- Low-confidence labels remain unresolved instead of being guessed.
Failure modes
- A participant speaks on behalf of another person.
- Two people share a role or first name.
- The recorder missed a joining or leaving participant.
Privacy notes
- Speaker attribution can affect accountability.
- Require human review before using low-confidence attribution in follow-ups.
- Keep evidence notes private when they include sensitive wording.
Example prompt
Map speaker labels to likely attendees. Use only transcript evidence, invite roster, and known role clues. Mark unresolved labels and explain confidence.
Example structured output
{"mappings":{"Speaker 1":{"person":"Dana","confidence":"high","evidence":["Introduces self as Dana"]},"Speaker 2":{"person":null,"confidence":"low","evidence":["No unique cue"]}},"reviewRequired":["Speaker 2 action ownership"]}