No cloud. No account. No telemetry.
Press a hotkey, speak for as long as you need — your text appears where your cursor is.
Aparté exposes a WebSocket bridge on your local network. Open a page on your phone, type or speak — the text lands directly on your desktop. No internet. No app to install. Nothing leaves the room.
Useful when your hands are busy, when you're reading on your phone and want to push a note to your desktop, or simply because your phone's microphone is better than your laptop's.
Not a demo project. Extracted from months of real daily dictation — with all the edge cases that come with it.
Audio is chunked in 30-second windows and processed while you speak. Stopping a long dictation is nearly instant — no waiting after the fact.
Your names, project names, technical jargon — defined once, corrected deterministically. No model fine-tuning, no re-training.
Whisper invents text on silence. Aparté detects and drops it. What you said — and nothing you didn't say.
Failed chunks are spooled and replayed. Aborted dictations are salvaged. Partial text survives a crash.
A model cleans up your transcript — but may only delete words, never add them. One invented word and raw text ships instead.
Select text, press Super+Shift+R, speak an instruction. The rewrite lands in your clipboard, ready to paste.
I wanted to write with my voice on Linux. Nothing matched Wispr Flow. The closest open-source tool broke on day one — a bug that was straightforward to diagnose.
Fixing that bug was the trigger: with today's AI assistants, one person can build the complete tool they actually want. Aparté is the result — extracted from months of daily use, refined by every edge case that came up.
It's MIT. It runs entirely on your machine. It will stay that way.
Wayland-first. Tested daily on Arch + Hyprland. Bring your own whisper.cpp server with the model of your choice.
yay -S aparte
aparte initpipx install aparte
aparte initaparte init writes a fully commented config file — no docs to read first.