Re: PSA: Automatic Minutes Generation

[Eric Rescorla <[e-mail]>]

On Mon, Oct 13, 2025 at 3:26 AM Dirk Kutscher <[e-mail]> wrote:

> Hi Ekr,
>
> > In an attempt to partly address this situation, I've put together an
> > automatic minutes generator at (https://ietfminutes.org/) The site
> > takes as input the Meetecho transcript and produces AI-generated
> > minutes. I've gone back and made minutes through IETF 112 (November
> > 2021) and I plan to produce minutes for IETF 124 (Montreal) once
> > transcripts are available, shortly after the meeting.
>
> Nice – thanks for putting this together!
>
> I had a look at some of the generated minutes. The gist of the meetings
> and the next steps are generally well captured. It currently does not
> reproduce the discussions (questions/comments/answers) explicitly (as many
> human note takers would). Do you think it's possible/useful to increase the
> level of detail a bit?
>

Probably? Perhaps a little more technical background would be helpful
here.

The entire meeting is recorded in both audio and (if available) video
and then posted on YouTube. The IETF proceedings [0] links to two
recordings:

- A "Video Recording" which just goes to YouTube.
- A "Session Recording" which goes to meetecho.ietf.org

The "Session Recording" [1] embeds the YouTube video but also includes
an auto-generated transcript [2] feature (see the attached
screenshot), which is what I use to generate the minutes.

Internally, that transcript is a set of text fragments with attached
timestamps that aren't tagged with who is speaking, though you can
sometimes infer it from the transcribed text. For instance, here is a
fragment from the Dispatch session in IETF 123:

{
  startTime: "00:14:00",
  text: " Okay, so I see Eckers in the queue, go ahead Hi, Erica
  Skorla, thanks to the presentation. I think I'm a little still
  unclear on what this thing does You said dating, but it wasn't quite
  provocative enough for me Could you, did you? like, give me a
  concrete example of what I would use? this So I can answer so um we
  have uh cloud storage solutions imagine you have an account on
  Google Drive, let's say someone else has an account on Dropbox and
  you want to share data with the person in Dropbox Currently, doesn't
  work for them because they're world gardens, but with our solution,
  if they could adopt open cloud mesh and you could share a folder or
  a file an application even with the user of the other system. So
  that's what you can do with next cloud you can share to own cloud or
  open cloud or to cernbox or something like that so in this instance
  Dropbox and Google Cloud would have to adopt Open Cloud Mesh. If
  they want to use it, of course Ah, Okay. Yeah. So then I think the
  predicate question is how many people who operate systems like this
  are interested? And I think the best way to discover that is have a
  buff. Yeah we already have a lot of uh installations and user one
  thing that's important to mention is that like the European Open
  Science Cloud, has opted to use use OCM in their tender. So if you
  want to be able to join you have to support the protocol and some
  German german yeah is the german and a lot of German institutions
  Bundeslenders has adopted it and indeed",
}

As far as I know, it's not posted anywhere on the IETF Web site other
than as part of the session recording feature (I used browser dev
tools to determine the URL pattern), though obviously it would be
possible to clean it up and post it. I think what you're asking for,
however, is a little more, namely what's sometimes called narrative
minutes, in which the minutes capture the essence of what someone said
(e.g., boiling down your email into "Dirk Kutscher: would it be
possible to produce narrative minutes").

It's probably possible to generate this from the provided transcripts,
and in fact if you ask Gemini to do that from this fragment, it
produces something plausible and even corrects my name and and
correctly interpretes "buff" as BOF:

   Eric Rescorla then joined the queue, saying he was "a little still
   unclear on what this thing does" and asked for a concrete example.

   The presenter answered by explaining that the protocol, Open Cloud
   Mesh (OCM), solves the "walled gardens" problem for cloud
   storage. He gave an example: "imagine you have an account on Google
   Drive... someone else has an account on Dropbox and you want to
   share data... with our solution... you could share a folder or a
   file."

   Rescorla confirmed his understanding that this would require
   providers like Google and Dropbox to adopt OCM. He then said, "I
   think the predicate question is how many people who operate systems
   like this are interested? And I think the best way to discover that
   is have a BoF."

   In response, the presenter said that they "already have a lot of
   installations and users." He pointed out that the European Open
   Science Cloud has mandated OCM in its tenders and that it has also
   been adopted by a number of German state-level institutions.

If others want this, I could probably add this kind of more verbose
version as well. I'm reluctant to just replace the shorter version
with that, because part of the purpose here is to make it easy for
chairs to provide the minutes required by RFC 2418 and so I want to
provide something which is an easy drop in for those, but it's not
much more difficult to generate things in multiple forms.

This would be easier, I suspect, if we had better labeling for who was
speaking. Meetecho already knows who is in the queue and which
microphone input is coming in from, so in principle it should be able
to tag things appropriately, e.g., have the transcript look more like
this:

    {
      startTime: ...,
      input: "Chair microphone",
      text: "Okay, so I see Eckers in the queue, go ahead"
    },

    {
      startTime: ...,
      input: "Remote, Eric Rescorla",
      text: "Hi, Erica
        Skorla, thanks to the presentation. I think I'm a little still
        unclear on what this thing does You said dating, but it wasn't
        quite provocative enough for me Could you, did you? like, give me
        a concrete example of what I would use? this"
    }

-Ekr


[0] https://datatracker.ietf.org/meeting/123/proceedings
[1] See, for instance:
https://meetecho-player.ietf.org/playout/?session=IETF123-DISPATCH-20250721-0730
[2] Note that YouTube also has its own transcript, but AFAICT
that's not what's on the IETF site.
[3] For instance:
https://meetecho-player.ietf.org/playout/transcripts/IETF123-DISPATCH-20250721-0730
[4] https://datatracker.ietf.org/doc/draft-rescorla-auto-minutes/