Private · Pay once · Windows and Linux
Dusk is private dictation that types wherever your cursor is, and runs on your computer. Also use voice commands to control your computer and navigate. No cloud, no account, no subscription. $39 once, yours forever.
For Windows 10 and 11, and Linux on GNOME (Wayland or X11). macOS is on the way.
Dictation
Nobody talks in finished sentences. You pause, say “um”, change your mind halfway through. Dusk keeps up, and the words on screen read like you typed them carefully.
01 You hold Ctrl+Win (on Linux, Ctrl+Alt) and say
“um let's meet at three no actually four and I'll bring the slides for the quarterly review”
02 On your computer
03 Typed where your cursor is
Let's meet at four. I'll bring the slides for the quarterly review.
| You said | Dusk typed |
|---|---|
| okay so uh quick update the new build is out and it fixes the login bug can everyone test it before lunch | Okay, so quick update: the new build is out and it fixes the login bug. Can everyone test it before lunch? |
| hi anna thanks for the notes uh the three things I need are the budget the timeline and the list of risks | Hi Anna, thanks for the notes. The three things I need are the budget, the timeline, and the list of risks. |
| can you send me the figma link when you get a chance thanks | Can you send me the Figma link when you get a chance? Thanks. |
History shows every dictation word for word: what clean-up removed, what it changed, and how long it took. Copy it, insert it again, or add a misheard word to your dictionary in one click.
Chrome's accessibility tree is off: clicking links and selecting text on web pages won't work.
Type what you'd say, then see what a command would do or how clean-up would write it.
Your dictations and commands, newest first.
um let's meet at three no actually four and I'll bring the slides for the cooper netties review
Let's meet at four, and I'll bring the slides for the Kubernetes review
open the calculator
Actions: launch_app
[
{
"action": "launch_app",
"name": "calculator"
}
]hi anna thanks for the notes uh the three things I need are the budget the timeline and the list of risks
Hi Anna, Thanks for the notes. The three things I need are: 1. The budget 2. The timeline 3. The list of risks
search for tauri transparent window on linux
Actions: open_url[
{
"action": "open_url",
"url": "https://www.google.com/search?q=tauri+transparent+window+on+linux",
"new_tab": true
}
]fix the off by one error in the loop over the rows
Fix the off-by-one error in the loop over the rows
go to line forty two in main dot rs
Actions: open_file[
{
"action": "open_file",
"name": "main.rs",
"line": 42
}
]git status
git status
Names and words Dusk should spell your way. Add how the speech model mishears them, if you know.
5 words
Say a trigger phrase while dictating, and Dusk inserts the text instead.
ops@example.org
12 Harbour Street Springfield, OR 97477
https://meet.example.org/team-standup
Thanks, Dave
4 snippets
The models Dusk runs on this computer. Keep several, switch, and compare them.
Turns your voice into text. Required.
Downloading Moonshine Base (English).
NVIDIA's multilingual streaming speech recognition model: accurate, and transcribes while you speak. Adds some punctuation itself.
To delete this model, use another model first.
NVIDIA's English-only streaming model: transcribes while you speak, with punctuation and capitals.
Some of this model's files are missing or damaged. Re-download it to repair it.
NVIDIA's English model, among the most accurate: transcribes when you stop speaking, with punctuation and capitals.
Adds punctuation for speech models that don't.
Adds punctuation and capitals to speech models that don't (7 MB, fast on any computer).
To delete this model, use another model first.
Your speech model's own punctuation is used as is.
Cleans up your dictation and understands commands the rules don't cover.
Good clean-up and commands, fast on a GPU. The default.
To delete this model, use another model first.
Clean-up uses simple rules, and commands use the rules only.
Dusk can find them on this computer and add them without downloading again.
Turns your voice into text. Required.
A small, fast English model for slower computers: transcribes when you stop speaking, with punctuation and capitals.
Adds punctuation for speech models that don't.
You have every punctuation and capitals model in the catalog.
Cleans up your dictation and understands commands the rules don't cover.
Small and fast, for computers without a GPU. Lighter clean-up.
Better clean-up and understanding, for GPUs with 6 GB or more.
Changes save right away.
Dusk starts in the tray when you log in.
Where the small recording indicator appears.
A small dot, so you always know Dusk is running.
A soft sound when recording starts and stops.
More languages come later.
Up to 3 shortcuts for each. A shortcut can be modifier keys only (like Ctrl + Super), modifiers plus a key, a function key, Pause, Scroll Lock, Insert or a mouse button. Hold one now: its row lights up.
Hold to talk, release to insert. Tap to start hands-free.
Press to start, press again to stop.
Hold and say what to do, release to run it.
Only while recording or working: throws the session away.
Inserts your last dictation again.
Hold the dictation shortcut at least this long for push-to-talk. A shorter tap starts hands-free dictation.
Hold the front button for a command, and the rear one for dictation (click it for hands-free). They stop working as back and forward.
Turning it off frees the dictation shortcuts.
How much Dusk tidies what you say before inserting it.
Also applies your corrections (“no, actually…”) and formats lists and numbers.
“Three, no, actually four” becomes “four”.
“First…, second…” becomes a numbered list.
Say “comma” or “period” to insert them.
Chat messages look more natural without it.
When you fix a word right after dictating, it's added to your dictionary.
Words appear while you talk, and each phrase is cleaned up when you pause. Off: the text appears when you finish.
Shows the words as you speak.
Fast. Uses the clipboard for a moment.
Puts back what you had copied.
Some applications need a different method. Use the window class, like “gnome-terminal-server”, or the program’s name on Windows, like “WindowsTerminal”.
Clean-up adapts to the kind of application: casual in chat, paragraphs in email, exact in code editors and terminals.
Hold the command shortcut and say what to do. Turning it off frees the shortcut.
When the built-in rules don't understand a command, the language model plans it.
Dusk works better with these applications after a one-time setup.
VS Code is installed. Dusk's key bindings aren't added yet.
Chrome is installed. Its accessibility tree is off, so commands can't see links and buttons.
A microphone takes a moment to wake up, which can cut off your first word. With this on it stays on while Dusk runs (your system shows it in use), but nothing is recorded until you press a shortcut.
Everything runs on this computer. Dusk only connects to the internet to download models you choose and to check for new versions.
Your dictations and commands, so you can find and insert them again.
Older entries are deleted automatically.
Deletes every entry now.
Words dictated, speaking pace and time saved, shown on Home, start again from zero. History is kept.
Keeps the audio of your last 20 sessions on this computer, to troubleshoot and to compare speech models.
Once a day, asks GitHub whether a newer version of Dusk is out, and shows it on the About page. Nothing about you or your dictations is sent.
Where the language model runs. Automatic picks the best one.
Automatic uses what's best for this computer (16 cores).
Saves memory, but the first dictation after a pause takes a few seconds longer.
Time without use before the memory is freed.
Unloads the speech and language models. They load again when you next dictate, which takes a few seconds.
/home/you/.local/share/dusk/models
Logs your dictated text in full, to help find problems. Normally only its length is logged.
/home/you/.cache/dusk/logs
Back to the defaults. Your models, history, dictionary and snippets stay.
Speak to dictate and to control your computer.
Version 0.1.0-preview
Everything runs on this computer. Dusk only connects to the internet to download models you choose and to check for new versions.
Models aren't part of Dusk. Each one you download comes with its own license.
The open source software and fonts included in Dusk, with their licenses.
How it works
Every step runs on your computer, and you can see and undo each one.
01
Ctrl+Win for dictation, add Alt for a command. On Linux, Ctrl+Alt, and add Shift. Tap it instead to keep talking hands-free.
02
A small pill at the bottom of the screen shows your voice and the words so far. It never takes focus.
03
Chat stays casual, email gets paragraphs, and code editors and terminals keep file names and commands exactly as you said them.
04
Typed or pasted into the app you were using. Changed your mind? The command “erase everything” clears the field.
so the plan for next week is
Commands
Hold Ctrl+Win+Alt (on Linux, Ctrl+Alt+Shift) and say what you want done, in your own words. There are no commands to memorize: common ones run instantly, and the language model on your computer works out the rest from your open windows.
Open the calculator
launch_app calculator
Search in Chrome for bread recipes
web_search "bread recipes"
Go to x com
open_url https://x.com
Close this tab
key ctrl+w
Erase everything
key ctrl+a BackSpace
Type git push dash dash force
type "git push --force"
Make this more formal
rewrite the selected text
The action Dusk planned for each command, from dusk command.
Dictionary
Add names and jargon to the dictionary, with how the speech model mishears them. Snippets expand a short phrase into a full paragraph.
Names and words Dusk should spell your way. Add how the speech model mishears them, if you know.
5 words
Models
Setup checks your memory and graphics card and recommends speech and language models that run well on them. Each model's license is shown before you download it.
The models Dusk runs on this computer. Keep several, switch, and compare them.
Turns your voice into text. Required.
Downloading Moonshine Base (English).
NVIDIA's multilingual streaming speech recognition model: accurate, and transcribes while you speak. Adds some punctuation itself.
To delete this model, use another model first.
NVIDIA's English-only streaming model: transcribes while you speak, with punctuation and capitals.
Some of this model's files are missing or damaged. Re-download it to repair it.
NVIDIA's English model, among the most accurate: transcribes when you stop speaking, with punctuation and capitals.
Adds punctuation for speech models that don't.
Adds punctuation and capitals to speech models that don't (7 MB, fast on any computer).
To delete this model, use another model first.
Your speech model's own punctuation is used as is.
Cleans up your dictation and understands commands the rules don't cover.
Good clean-up and commands, fast on a GPU. The default.
To delete this model, use another model first.
Clean-up uses simple rules, and commands use the rules only.
Dusk can find them on this computer and add them without downloading again.
Turns your voice into text. Required.
A small, fast English model for slower computers: transcribes when you stop speaking, with punctuation and capitals.
Adds punctuation for speech models that don't.
You have every punctuation and capitals model in the catalog.
Cleans up your dictation and understands commands the rules don't cover.
Small and fast, for computers without a GPU. Lighter clean-up.
Better clean-up and understanding, for GPUs with 6 GB or more.
Compare
The well-known dictation apps are good at text. But they charge every month, send your voice to their servers, and stop at typing.
| Dusk | Cloud dictation apps | Paid local apps | Open-source transcribers | Voice control | |
|---|---|---|---|---|---|
| Price | Free plan, or $39 once | Subscription, limited free tier | Paid license or subscription | Free | Free or built in |
| Works offline | Yes | No | Yes | Yes | Some |
| Your voice stays on your computer | Yes | No, processed in the cloud | Yes | Yes | Yes |
| AI clean-up of dictation | Yes, local | Yes | Yes | Partly, or through a cloud service | No |
| Commands in plain language | Yes: windows, apps, websites, edits | Rewriting selected text | No | No | Fixed phrases or scripts |
| Windows | Yes | Yes | Some | Some | Yes |
| Linux | Yes | No | No | Yes | Some |
Based on public information about popular apps in each category in October 2026. Features change.
Privacy
Plans
The free plan has every feature, and a license only lifts the weekly word limit. Whichever plan you're on, your voice, your text and your history stay on your computer.
See pricingPricing
Every feature works on the free plan. When Dusk becomes how you write, a license removes the weekly limit for good: no subscription, no account.
$0
$39 once
Payments and sales tax are handled by Polar, our reseller. Full refund within 30 days, for any reason. Terms of sale · Privacy
Download
Download the package for your system and open it. Setup checks your computer, recommends the speech and language models that run well on it, and downloads them for you.
Run it. It installs Dusk for your user, without an administrator password, and adds it to the Start menu. Remove it in Settings, Apps.
Open it with App Center, or install it from a terminal:
sudo apt install ./Dusk_amd64.debInstall it from a terminal:
sudo dnf install ./Dusk_x86_64.rpm
Release notes, checksums and earlier versions are on the
releases page. The checksums are signed
with our signing key: gpg --verify SHA256SUMS.asc, then
sha256sum -c --ignore-missing SHA256SUMS.
FAQ
Windows 10 or 11, or Linux with GNOME on Wayland (Ubuntu's default) or an X11 session; 8 GB of memory; and about 2 GB of disk space for the recommended models, which setup downloads the first time. On Windows, Dusk's shortcuts and typing don't reach apps running as administrator.
No. A graphics card with Vulkan support makes clean-up faster, but isn't required. Setup recommends models that run well on the computer you have.
The free plan has every feature, with a limited number of words each week. A license is $39 once and removes the limit. There's no subscription: pay once and use it forever, with every update we publish in the next 12 months.
No. Your voice, your text and your history stay on your computer. Dusk connects only to download the models you choose, to check for new versions, which you can turn off in Settings, and once to check the license key you enter. Word counts stay on your computer.
Dictation and most commands work there too. Commands that switch between windows need GNOME on Wayland, or an X11 session, for now.
It depends on the speech model you choose. Nemotron 3.5 understands English, Arabic, Chinese, French, German, Japanese, Korean, Spanish, Ukrainian and Vietnamese; the other speech models are English only.