Why meeting transcripts should not guess speaker names

A transcript should say “Speaker 1” when all it knows is that one voice differs from another. Diarization separates voices. It does not identify people.

Scripta refuses to turn a neutral speaker label into an attendee's name automatically, even when the meeting already has a participant list. A plausible attribution is still a guess, and a guessed name can assign an objection, decision, or promise to the wrong person.

Separating voices is not identifying people

Diarization answers a structural question: where does one speaker stop and another begin? The result can group turns under labels such as Speaker 1 and Speaker 2.

Identity is a different question. The audio service does not know that Speaker 1 is Mira. It does not know whether everyone on the attendee list spoke, whether two people shared one room, or whether an unlisted guest joined the conversation.

The attendee list makes a name possible. It does not make the attribution true.

A wrong name changes the meeting record

Neutral labels may create a few minutes of review work. A wrong name creates false history.

Consider these transcript lines:

Speaker 1: I can own the rollout checklist.

Speaker 2: I still think the consent language needs review.

If the software guesses incorrectly, the finished minutes can give one person work they never accepted and attach another person's objection to them. Once the names appear in polished prose, later readers rarely see them as tentative.

This is especially dangerous around action items. An executable task needs an owner and a date, but ownership should be established by the conversation and human review—not filled from the nearest available name.

Let a person resolve identity

Scripta returns diarized speakers with neutral labels. Someone who recognizes the voices can name them. The mapping belongs to that recording; Speaker 1 in Monday's meeting is not assumed to be Speaker 1 on Tuesday.

That last constraint prevents a convenient-looking shortcut from becoming a quiet source of error. Speaker order can change between recordings. A persistent mapping would attach a person to whichever voice happened to appear first next time.

The useful workflow is straightforward:

1. listen to a short, distinctive turn;
2. assign a name only when you recognize the voice;
3. leave the neutral label when you do not;
4. promote the reviewed transcript to the meeting.

Verified words and verified identity are different

When Scripta proposes a decision or action item, its supporting quote is checked against the meeting transcript. That check answers: “Do these words exist in the source?”

It does not independently answer: “Did this person say them?” If the transcript still says Speaker 2, the quote remains grounded in the text while the identity remains unresolved.

Those two claims should not be collapsed. Evidence can be strong about what was said and incomplete about who said it. A useful review interface preserves both facts.

This is the broader standard for AI meeting notes that show their work: generated structure should expose its source, and uncertainty should remain visible until a person resolves it.

Completeness is not the same as accuracy

Software often tries to remove every blank because finished-looking output feels better. In a meeting record, that instinct can be harmful.

“Speaker 1” asks for review. “Mira” states a fact. If the system only possesses evidence for the first statement, adding the second does not improve the transcript. It invents certainty.

The same rule applies when documenting disagreement. A record can preserve an objection without reopening a decision, but only if it does not put that objection in the wrong person's mouth.

When voices return as Speaker 1 and Speaker 2, the labels tell you exactly what the system knows and no more. A person can supply the names; the software should not manufacture them.