My meeting app has a built-in recorder. Click the button, it records, it transcribes. Sounds great until you actually need the output.
For a start, half the time it's the wrong side of the call. Screen-share captures the presenter's voice in one channel and leaves everyone else muddy, or it only records what one person said, or it quietly depends on the host being on the right plan to even give you the file. And every one of those recordings is sitting in someone's cloud before you've ended the call.
Then there's the output. You get a transcript. A wall of text. Sixty people saying a lot of words, none of them agreed to anything, and you're the one who's supposed to turn it into minutes at 9pm.
I built something instead. It records the mic and system audio on my own PC - which means it catches a Teams call, a Zoom, or a person sitting across a table, all the same way. Uses WASAPI loopback so it grabs what the app is playing, not a screen recording with a voice on top. Then Whisper transcribes it locally, and a small LLM in LM Studio reads the transcript and writes the minutes the way I would if I'd paid attention all meeting: a Summary, the Key Decisions, the Action Items with owners and dates where anyone actually said one, and the Open Questions nobody answered.
Export to Markdown, text, or PDF. Done. No account, no subscription, no upload.
That last part isn't a slogan, it's the whole design. The audio never leaves the machine. Whisper runs here. The model that writes the minutes runs here. There's no pipeline that even could send it anywhere, which is kind of the point when your employer's setup treats every recording as a thing to be governed.
One honest note: recording meetings may be regulated where you work, and whether your colleagues know they're being recorded is a genuine question, not a checkbox. Get consent, read your policy, then use a tool that doesn't make the situation worse by also spraying the audio to a third party.
You don't need a monster machine. Windows 10 or 11, Python 3.10+, and LM Studio running any small model - it chews through a 45-minute meeting on CPU fine; a GPU just makes the transcription snappier.
I put it up as a one-time thing for $9: https://symshah.gumroad.com/l/xgoypx. No surprise monthly charge like the apps it replaced. Install it, point it at a meeting, and see if you actually stop dreading the minutes.
People who attend meetings get to doze off. One of you still has to write it down. Might as well be a local model instead of you.
Top comments (0)