New · Linux first
Speak instead of typing — in any app.
Press Ctrl+Shift+Space and talk. The text appears word by word, about half a second after you say it — in every field in Undan and, on Linux, in your browser, terminal or word processor too. The speech models run on your computer. Nothing you say leaves it.
No network, no account, no speech-service subscription. Undan keeps no recordings — only the text remains.
How it works
Download the model once
Under Settings → Dictation, Undan downloads the speech model (682 MB, once) and checks it against its checksum. After that, dictation needs no network.
Press the key, talk
In any text field — a note, an email, search, chat. A bar appears at the bottom: the waves show the microphone hears you, below it what is coming in.
Ctrl+Shift+Space
Press again — done
Or
Esc, or simply stop talking: after the silence you have set, it ends by itself.Ctrl+Ztakes back what you just dictated.
What dictation does
In any app
When another program is in front, Undan types the words where its cursor is — in Slack, the browser, the terminal. What appears there is what the model hears as you speak; for the most accurate text, dictate into a note and copy it from there. Linux for now.
On your computer
The speech models — Nemotron, Parakeet, Whisper — run here, with no network and no account. The microphone is only open while the bar is showing. Undan keeps no recordings.
Fast first, then accurate
One model writes along word by word. A second one re-reads at every pause and puts the more accurate version, punctuation included, in its place — but only if your computer is fast enough. Undan measures that itself.
You keep writing
Type, click, fix a word while you talk: what you said so far stays where it is, what comes next goes to the new spot. Dictation never overwrites what you changed yourself.
German, English, Swedish
Switch the language in the bar, even mid-dictation — from the next sentence on. “New paragraph” and “new line”, said on their own, structure the text; the models add punctuation themselves.
An instruction to the AI
With Ctrl+Alt+Space you say what should happen to the text: “add paragraphs”, “shorter”, “write to Irma asking whether she can come tomorrow, and add a task for Friday”. The AI proposes,
Returnapplies. Instructions need the network — dictation itself does not.
Where it works today
Dictation is built and in daily use on Linux. What is still open is listed here.
Works
- In Undan: every text field, in every window of the application.
- In other programs on Linux (KDE Plasma, Wayland): Undan types into the program in front.
- German, English, Swedish, with re-reading for a more accurate version.
- Instructions to the AI with their own key.
Not yet
- Other Linux desktops (GNOME, X11): not yet tested.
- Windows and macOS: dictation into other programs comes later there.
- Phone: a first trial runs; not yet tested on a device.
Questions
Does dictation need the internet?
No. After downloading the model once, everything runs on your computer. Only an instruction to the AI goes over the network — to Claude, or to your own model if you have set one up.
What does my computer need?
An ordinary laptop is enough: the fast model and re-reading with Parakeet run on the processor. The models take about 1 GB of memory, 1.8 GB with re-reading. Whisper is the most accurate and fast enough with a graphics card; without one, the fast model writes on its own.
Does it work in Word, the browser, the terminal?
On Linux, yes: Undan types into the program in front. It only appends there and does not re-read, so nothing you have clicked in the meantime gets overwritten. For the most accurate text, dictate into a note in Undan — a second model re-reads there — and copy it from there.
Is Undan always listening?
No. The microphone is only open while the bar is showing. Dictation ends when you press the key, Esc, or after a silence you set. Recordings are not kept.
And on Windows and macOS?
Undan itself is designed for them but not yet tested there. Dictating into other programs needs its own way to the keyboard on each system — it comes as soon as it can be tested there.
Which models, which licences?
Nemotron 3.5 (NVIDIA, OpenMDW-1.1) writes along; Parakeet v3 (NVIDIA, CC-BY-4.0) and Whisper large-v3-turbo (OpenAI, MIT) re-read. Everything runs on sherpa-onnx, ONNX Runtime and whisper.cpp — all open source.
Join the waitlist
We'll write to you when there's something to try — and otherwise not. No newsletter rhythm, no ads for anything else.