Voice Notes

What it does

I press record on my phone, and minutes later a finished note appears: who said what, the key points, the actions to take. The recording lands on my Mac Mini at home, which picks it up, sends it off to a couple of AI services to transcribe and summarise, separates the speakers, works out who was in the room, and files the note so it links to everything else I have already written. All I do is press record.

Why I built it

I am quite bullish that every conversation will one day be recorded, so AI can draw on it later. The paid tools that do this now - Granola, Otter and the rest - rent you the software, keep your recordings on someone else's servers, and mostly stop at a transcript. There is no self-improving memory layer. I wanted it running on hardware I own, dropping the note straight into my own system where it is searchable and linked. My first version did all that too, but leaned on cloud services that cost about £20 a month. This one costs me around £3 a month in compute - far less.

The bit I am proudest of

I wanted it fully automated, end to end: press record, done. Separating the voices in a transcript is easy. The hard part is getting the AI to work out who is actually on the call without me telling it anything. So I have given it access to my calendar and the full history of my Obsidian second brain. It infers the likely speakers, cross-checks them against the recording, and labels the note with real names instead of Speaker 1 and Speaker 2. It does feel like magic. The finished note lands on my phone, and if anything is wrong I just reply to the message to fix it - but it is about 99% accurate as it stands.

Where it stands

Live on the Mac Mini at home, processing real recordings. Built in Python, running as a background service that checks for new audio every minute. AssemblyAI handles transcription; Claude does the summarising, speaker-matching and filing. Nothing leaves my own machines apart from one pay-per-use transcription call.

Back to Projects