The part nobody warns you about when you decide to keep meeting recordings on your own machine is that Windows won't just let you record what you're hearing. Out of the box, it doesn't.
The mic is easy. Windows grabs that fine. But the far end of a Teams or Zoom or Meet call, the voices of everyone on the other side, those arrive through your speakers or headset. And there is no stock setting that says "also record that." Go look. I'll wait.
The standard answer is WASAPI loopback. It lets one program tap into your output device and pull the samples before they become sound, the same trick screen recorders lean on when they promise to also capture system audio. It's not exotic, but it's fiddly, and it's the part nobody writes about — because transcription and the LLM writing the summaries are the fun half. The capture is the boring half. It's also the part that decides whether your minutes actually cover both sides of the conversation.
Getting it right means handling a few things not in the blog posts:
- Loop back clean, not the audio that's already been through your chosen volume or any of the meeting flavor sounds.
- The feed comes straight off the device, so your job is matching it against what's coming out and what's coming in, without letting either get stomped.
- If the timing between system audio and your mic drifts, the transcript goes incoherent and the minutes are worthless. A conversation where people cut each other off needs both streams landing in the right order.
It's a real engineering problem, the kind you don't notice until it's wrong.
What I actually built is a small Windows app that records the mic and the system audio at the same time via that loopback, transcribes it locally with Whisper through faster-whisper, then passes the transcript to a small local model in LM Studio to write actual minutes — Summary, Key Decisions, Action Items with owners and dates when people said them, Open Questions. Export to Markdown, plain text, or PDF.
The reason that matters is the whole thing runs on my PC. No account, no subscription, nothing uploaded. The audio never leaves my machine. It genuinely doesn't. A GPU makes Whisper faster, but it all runs on CPU if that's all you've got.
One honest note, because it's true and it fits: recording meetings may be regulated where you work. Check your org's policy and get consent before you record people. The point isn't to record people in secret — it's to not have your meeting audio leave your machine at all.
Plenty of tools will transcribe your meetings for free, because you're the product and the audio is theirs to keep. I didn't want that. The ride where you're wondering which vendor's model is digesting your weekly staff all-hands is worse than just fixing the loopback once.
If the whole problem sounds like one you'd rather just have solved, I made it a small Windows app. $9 one time, no subscription, real code finally writing the minutes the meetings never did. https://symshah.gumroad.com/l/xgoypx
Top comments (0)