← All posts

AI meeting summaries, free and private

Nobody re-reads a full transcript. The thing you actually use after a meeting is the summary: what was decided, what you agreed to do, and who owns what. So it's a little strange that most tools treat the summary as an afterthought — a wall of transcript with a couple of bullet points bolted on top, generated by uploading your entire call to someone else's servers.

I wanted meeting summaries that are genuinely useful and stay on my machine. That's a big part of what Humla does, and it's free.

A transcript is not a summary

Automatic speech recognition has basically been solved — the top tools all transcribe English at 90-something percent accuracy. That's table stakes now, not a feature. The hard part is turning a messy, hour-long conversation into something a busy person can act on in thirty seconds.

A good meeting summary answers three things:

  • What was decided? The conclusions, not the debate that led to them.
  • What happens next? Concrete action items, ideally with an owner.
  • What was said that matters? The facts and numbers worth keeping, minus the small talk.

Most "AI summaries" get the third one half-right and skip the first two, because they only have the transcript to work from. They know what words were spoken; they don't know what any of it meant to you.

Why your own notes make the summary better

Here's the trick Humla uses: it summarizes from two sources at once — the transcript and the rough notes you typed during the meeting.

Your notes are a signal of intent. When you jot "decided: ship Friday" or "Sam owns release notes," you're telling the model what actually mattered out of the hour. Humla's summary leans on your notes for the what mattered and the transcript for the exact details — so you get a recap that reflects the meeting you were in, not just the audio a microphone caught.

It also has presets for different kinds of meetings — a 1:1 gets summarized differently from a client call or a lecture — so the output matches the shape of the conversation.

Free, and it doesn't have to leave your Mac

Two things make this different from the usual cloud summary:

  • It's free. No per-seat subscription, no monthly cap on how many summaries you can generate. See free, open-source AI meeting notes for the fuller picture.
  • It can run on-device. You can point summaries at a local model so the whole thing — transcription and the summary — happens on your Mac and nothing is uploaded. If you'd rather use a stronger cloud model, you bring your own OpenAI key; either way it's your account, not a vendor sitting in the middle. More on that in local & on-device.

For anyone summarizing client calls or anything under NDA, that distinction is the whole ballgame. A summary is a condensed, quotable version of a confidential conversation — arguably more sensitive than the raw audio. Keeping it on your own disk is the safe default.

How you actually get one

The workflow is boring in a good way:

  1. Take rough notes during the meeting, like you already do.
  2. Humla records your mic and the call audio in the background — no bot joins the call.
  3. When you're done, hit summarize. You get a structured recap that fuses your notes with the transcript.

That's it. No bot to invite, no upload step, no seat to pay for.

It works in your language too

Summaries aren't English-only. Humla handles around 100 languages and can route each to the engine you want, including a local Norwegian model — so you can get a clean summary in the language the meeting was actually held in. I'll write more about non-English meeting notes and summaries soon.

Try it

Download Humla for macOS — free, open source, no sign-up. Or see the on-device transcription running in your browser on the local page before you install anything.