Everyone sells local AI like you need a serious GPU under your desk. Every blog post is an RTX this and a CUDA core that. But my work laptop is a four-year-old Dell with integrated graphics and a fan that sounds like a blender when you open Teams. And it's what records and transcribes my meetings, fully on its own, every single day.
I'm an IT admin. I live in meetings. Design calls, vendor calls, incident bridges, the weekly where my boss re-explains the thing we decided last week. I used to type pseudo-minutes into OneNote while half-listening, then I'd go back and fix them from memory and get half of it wrong. People started noticing the action items I misattributed.
I'm not going to lie to you: local transcription on a CPU is not fast. A one-hour meeting takes my laptop maybe fifteen minutes to chew through, running in the background while I get on with actual work. Whisper via faster-whisper handles the heavy lifting; I load it with whatever small model LM Studio has sitting around. The whole thing - record, transcribe, summarize, format - reads like:
python record_and_transcribe.py --mic --system --out minutes/
Then a local LLM takes the transcript and turns it into a structure I actually forward to people instead of leaving it in my head: Summary, Key Decisions, Action Items with owners and dates where anyone actually said them, Open Questions. Exported to Markdown or PDF and dropped into a shared folder. No one re-asks who owns what, because it's in writing, from the meeting, not from my memory.
Why bother with local at all? Because I watched my employer evaluate a tidy cloud minutes app once, and the sales pitch basically required us to hand them every word of every internal call. I said no. My company's IP stays on my disk. There's no account, no upload, no subscription - the recording dies on the machine it was made on. That's the entire point.
My laptop doesn't care that there's no GPU. It just takes a little longer, and I don't notice because the minutes aren't blocking anyone. Meetings are linear anyway - an hour of audio is an hour of content. If it took a GPU and an hour of holding it to the system, it would still be cheaper than me trying to summarize from a chat bubble afterward.
One honest note: recording meetings may be regulated where you work. Get consent from the people you're recording, or at least check the policy. I flag muting myself into my own personal notes on internal calls, but that's up to you and your compliance folk.
The thing I built is a small Windows app that does record-mic-plus-system, transcribe locally, drop structured minutes, export. Costs $9 once, no subscription, no cloud. https://symshah.gumroad.com/l/xgoypx if you want to give up on messy notes.
I keep coming back to it because the minutes come out the way meetings actually went: decisions, owners, dates, questions. Not a wall of words. For a laptop with no gaming GPU, that's not bad.
Top comments (0)