The Transcript Does Not Know Who Promised

You recorded the meeting. You have the transcript. It is all there, word for word, everything that was said.

And still there is nothing you can do with it.

Not because the text is poor. Because it is missing one thing that looks small and is not: it does not know who said what.

The difference between a note and a commitment

"We decided the quote goes out before Friday."

That is a note. It is true, it is accurately transcribed, and it drives nothing. Nobody wakes up on Thursday feeling it was their job.

"Anna told Markus the quote goes out before Friday, and Markus said fine."

That is something else. It can be followed up. It can be raised again. There is someone who can say "that was not mine to take" and someone else who can say "yes, it was".

A commitment without a sender is not a commitment. It is a phrasing. And that, I think, is the real reason so many meeting notes never lead anywhere. They are not too short. They are not too long. They are just missing who.

Three layers, three questions

It gets easier to see if you separate what each layer actually answers.

The transcript answers: what was said? It gives you a verbatim text of one occasion.

The structure answers: what belongs together? It connects occasions to one another, so that ten meetings become a thread instead of ten files.

The origin answers: who said it, and to whom? It turns utterances into something a person can carry.

Most people are on the first layer. Some have started on the second. Almost nobody talks about the third, and it is the only one that turns the material into something you can run a project on.

Teams is solved. The room is not.

Here I want to be honest about where the difficulty sits, because there is a risk this sounds easier than it is.

In Teams, Zoom and Meet the problem is effectively solved. Every participant has their own audio stream and their own login. The transcript arrives labelled, without anyone having to work anything out. That is good, and it is not where the challenge lies.

The challenge is the ordinary meeting room. Five people around a table. One microphone in the middle. People interrupt each other, talk over each other, sit at different distances. Nobody recorded a voice profile in advance and nobody intends to.

There is only the audio. And working out from a single track that there were four different people, and where one stopped and the next took over, is a very different thing from reading a label that already exists.

That part is unsolved enough to be worth working on. It is also where the conversations that matter most actually happen.

The uncomfortable part

Someone is now thinking: this sounds like surveillance.

That is a fair reaction and it deserves an answer rather than a wave of the hand.

The point is not to be able to single anyone out. Nobody needs a system that keeps track of who said something foolish in an internal meeting. The point is to capture the direction of a conversation. Who asked and who answered. Who took something on, and in front of whom. That is what allows a conversation to lead somewhere, and it is information that is lost every single time today.

And there is a comparison worth making. The alternative to a system knowing who said what is not that nobody knows. The alternative is that one person writes it down afterwards and decides it alone, from memory. That person did not catch everything, misremembered a couple of things, and unconsciously wrote their own role a little more clearly than everyone else's. That is not neutral. It is just invisible.

A verbatim record with stated origin is in fact fairer to everyone in the room than an account written by whoever had the least time to spare.

Two things that sound alike and are not

There is a distinction almost nobody makes here, and it matters more than it sounds.

Speaker separation is knowing that there were four different voices, and which one said what. Speaker A, speaker B, speaker C. Nobody knows who they are. That is signal processing.

Speaker identification is binding speaker B to a named person by matching the voice against a stored voice profile. That is something else entirely, and it is something else in the legal sense too: a voice profile used to single out a specific person is handled as biometric data, the strictest category of personal data there is.

Merging the two is easy, because in everyday language both are called "knowing who said what". But one is counting voices in a room. The other is measuring a body.

The order that follows from that seems fairly obvious once you have seen the difference:

Always separate. Speaker A, B, C, without storing any voice profile.

Never identify automatically.

Let the person bind the name. Whoever was in the meeting marks "speaker B" as herself afterwards, if she wants to. Then it is a person making the connection, not a system that measured her.

That last step is also the only reasonable one for a different reason, which has nothing to do with law: whoever binds a name to a commitment ought to be a human being who can be held to account for getting it right.

The layer that makes the difference

I have written a great deal about capture. About the transcript being the raw material and the summary being the least interesting part. About how anything left loose drives nothing, and how value only appears once the parts connect over time.

I stand by all of it. But it stops at a well-ordered archive.

What turns the archive into something that actually makes things happen is that it knows who said what to whom. Without it you have an excellent history. With it you have something that can run a project.

And as so often, the model is not the hard part. Knowing what you are missing is.