I work in IT for a mid-size company. Last year our legal team asked me to audit how we handle meeting recordings and transcriptions. The answer was embarrassing: half the department was using cloud-based transcription services that sent internal meetings over the internet for processing.
Not just any internal meetings either. Vendor negotiations, personnel discussions, product roadmaps — all of it getting piped to some remote server with an API key we found on a shared Google Doc from 2022.
So I built something that does what those services do without sending anything anywhere. It cost me $9 once. Not because I'm cheap (though I am), but because the alternative was writing a compliance report explaining why we shouldn't be doing this in the first place.
How it works
The app captures both your microphone and system audio through WASAPI loopback on Windows. That means Teams calls, Zoom meetings, Google Meet — everything that comes out of your speakers gets recorded alongside what you say. Or if people are sitting in the room with you, just point the mic at them.
From there it's two local tools:
- faster-whisper for transcription (runs entirely offline)
- LM Studio running a small LLM locally to read the transcript and produce structured minutes
The output is exactly what you'd expect from any paid meeting tool: a summary of what happened, key decisions with context, action items (owners and due dates if stated), and open questions nobody could answer in the moment.
Export to Markdown, text, or PDF. Done.
What makes this different from the cloud stuff
Three things:
- Nothing leaves your machine. At no point does any audio or transcript go anywhere. Not even a hash. The whole pipeline runs locally.
- No account required. No signup, no email, nothing tied to your identity. You download it and run it.
- One payment. $9 one-time on Gumroad. No monthly subscription. No per-seat pricing. No enterprise tier that costs more than my actual salary.
The technical stack is straightforward: Python 3.10+, Windows 10/11, LM Studio with any small model loaded (free), and a local installation of faster-whisper. It runs on CPU — a GPU makes transcription faster but isn't required. I've tested it on machines that are three or four years old.
The setup is boring, which is the point
This isn't one of those projects where you spend weekends configuring Docker containers and debugging environment variables. You install LM Studio (free), load a model (any small one works — I use Qwen 2.5 7B quantized), and then run this thing.
The app handles the orchestration between recording, transcription, and summarization. You don't need to understand how Whisper tokenizes audio or why LM Studio uses GGUF format. It just works.
A quick note on compliance
Recording meetings is regulated differently depending on where you live. Some places require all-party consent. Others only need one person to agree. Check whatever applies in your jurisdiction before you start recording things.
If you're tired of explaining cloud transcription services to your security team (or if your company just doesn't let you use them), this might save you some headaches:
Top comments (0)