Capturing meetings and keeping the transcripts

The setup people call a second brain is four steps: record the audio, transcribe it, summarise it, and save the text somewhere you can search. Every step can run on your own machine, and for this particular job that matters more than usual.

Why local is the right default here. A meeting recording is mostly other people's words. Sending it to a hosted API puts your colleagues' unguarded sentences in a third party's infrastructure under whatever retention policy applies that day — a decision you are quietly making for everyone in the room. Transcribing on your own machine means the audio never leaves it. One thing this guide cannot do for you: check the consent and recording rules that apply where you work before you hit record.

The transcription step is genuinely solved. The Whisper family runs on almost anything — Macs, PCs, even phones — through free apps like MacWhisper or the whisper.cpp tools, and models like Parakeet are dramatically faster if you have an NVIDIA card. Speed is measured as a multiple of real time, so a fast model turns an hour-long meeting into a transcript in a couple of minutes. The audio tab of the catalogue shows accuracy, speed and languages side by side.

Know what the accuracy number means. The measured scores are English-only, and every model degrades in the same places: crosstalk, distant microphones, heavy accents, noise. The condition boards show exactly how much, per model — worth a look before you trust a transcript of a chaotic eight-person call.

The summarising step is the other half of this site. Feed the transcript to a local model for the summary and action items — a mid-sized open model handles this well, and the pipeline stays entirely on your machine. If you already run Ollama, the whole chain is scriptable: audio in, dated markdown note out.

Store it as plain text. Markdown files in a folder beat any app: searchable with anything, portable everywhere, readable in thirty years. The second brain is the folder, not the software.

Where next: The audio catalogue · Run your first local model · What your machine can run