Somewhere in a privacy policy you clicked past, most meeting-notetaker apps reserve the right to train on your own calls. I read that clause about Otter's suite after the 2026 Illinois lawsuit landed, and I got quiet for a second. A third of my calendar is vendor calls, HR syncs, and salary talk. The idea that any of it was fair game for someone's model is what finally pushed me to run meeting notes entirely local.
What private meeting notes actually mean
Private meeting notes don't mean the tool has a pretty lock icon. It means the audio never leaves your machine. Not a transcription service on some vendor's cluster, not a summarization API, not a training corpus. When I say nothing leaves the PC, I mean literally nothing leaves the PC. If you unplug the router mid-meeting, the whole pipeline still finishes.
Where cloud notes quietly leak
The leaks are rarely a breach. Usually it's boring and contractual. A bot joins as a participant and everything said goes to someone else's server. The vendor's own LLM summarizes it. The raw recording gets retained for "quality" and staff can see it. Maybe that's fine for a general standup. On a call about a client's unpaid invoice or a peer's performance review, it's not.
That's the gap I was trying to close. I didn't want a cheaper Otter. I wanted the same structured output with the recording and the minutes never existing anywhere but my desk.
The setup that keeps meeting minutes local
I'm on Windows, so the app I settled on records mic and system audio — Teams, Zoom, Meet, or a room full of people — then transcribes on-device with Whisper, then feeds it to a small local model in LM Studio that writes the minutes. Summary, key decisions, action items with owners and dates where they were stated, open questions. It exports to Markdown, text, or PDF.
Requirements are modest: Windows 10/11, Python 3.10+, and the free LM Studio with any small model loaded. It runs on plain CPU. A GPU just makes the transcription step faster — mine's a work laptop with no gaming card and the bottleneck is still my attention, not the box.
# no cloud call here - faster-whisper and LM Studio both run locally
pip install faster-whisper
I keep the raw transcript too, because a transcript is a record and minutes are a promise. When someone swears a decision went differently, I have the exact words from my own disk, not a link to a shared doc I don't control.
Why I stopped caring about the "best" notetaker
Every cloud notetaker compares on accuracy percentage and UX polish. I stopped comparing once the constraint was privacy. If the audio can't leave the building, most of the market is off the table by definition, and the choice shrinks to how well a local tool handles my meeting structure. Structured action items beat a pretty transcript every week.
I've been running this for a few months now. The minutes land as files, I keep what matters, and searchable history stays under my own folder. It costs $9 once, no subscription, no account, no upload queue. If you want the same without handing your calls to a vendor, it's at https://symshah.gumroad.com/l/xgoypx
FAQ
Will a local setup keep up with a long call?
Yes. Transcription runs as the meeting happens (or right after), and a small LLM writes the minutes. On CPU a marathon three-hour session waits longer for the transcript pass, but it always finishes — and it finishes on your machine.
Do I need an Nvidia GPU for this?
No. CPU works fine with faster-whisper's small models and any half-decent LM Studio model. A GPU speeds up the transcription, that's the whole difference.
What happens to the recordings afterwards?
What you choose. They stay in a local folder, get moved to your backup routine, or get deleted. There is no vendor retention policy because there is no vendor.
Top comments (0)