Who Said What: The Feature That Turns a Transcript Into a Record

Open a meeting transcript with no speaker labels and you have a wall of text: a hundred lines of dialogue and no idea who said any of it. You can see that a decision was made, but not who made it. You can see a commitment, but not who owns it. The recording captured every word and lost the one thing that makes words actionable: attribution.

Why "who said what" is the whole game:

• Decisions need an owner. "We'll ship Friday" is a rumor until you know who said it and who agreed.
• Commitments need a name. An action item with no assignee is a task nobody will do.
• Accountability needs a record. "I thought you said you would handle it" is one of the most expensive sentences in software. A speaker-labeled transcript ends that argument before it starts.

Attribution is also the hard part. Anyone can run audio through a speech model and get text back. Telling voices apart, what is called speaker diarization, is where most tools quietly fail. Real meetings have people talking over each other, similar-sounding voices, cross-talk, and a single shared system-audio stream where everyone is mixed together. Generic cloud tools tend to smear all of that into one undifferentiated transcript, or guess wrong about who is speaking.

Doing it well takes more than a transcription API. It means models tuned for overlapping speech, careful handling of the captured audio, and labeling every line with the actual participant behind it, not just "Speaker 1" and "Speaker 2." Get that right and the transcript stops being a recording and becomes a record: searchable, attributable, and ready to turn into assigned work.

Celeritas ties every line in the transcript to the person who said it, on your own device. When the call ends, you do not get a blob of text. You get who decided what, who owns which task, and exactly when it was said. That is the difference between a transcript you skim and a transcript you can act on.