Guide · Privacy
Private dictation that stays on your computer
Dusk turns speech into text with models that run on your own computer. Your voice, your text and your history never leave it: there's no account, no telemetry and no cloud processing. Dusk connects to the internet for four things only, listed below, and none of them carries what you say.
Last updated October 27, 2026
What runs on your computer
- The speech model writes down your words while you speak.
- The language model cleans up the text: it removes filler words, applies corrections, and adds punctuation and paragraphs. It also plans voice commands.
- The guard compares the cleaned-up text with what you said. If clean-up added or dropped words, Dusk types your words as you said them.
- History keeps each dictation for 30 days by default, readable only by your user account. You can shorten it, keep it forever, turn it off or clear it.
- Word counts for the free plan's weekly limit are kept on your computer too.
The models run on your graphics card when there is one, and on the processor otherwise. Speech is transcribed while you speak, so the text is ready a moment after you let go of the shortcut.
Every connection Dusk makes
All of them go to Dusk's own server, api.duskdictation.com, which keeps no logs and stores
nothing. It forwards license requests to Polar, our reseller, and redirects downloads to where the files are
hosted.
| Connection | When | What it sends |
|---|---|---|
| Model download | When you choose a model | Which model file to download |
| Update check | Once a day. Turn it off in Settings | A request for the newest version and the list of models |
| License activation | When you enter a license key, or remove a computer from it | The key, your computer's name and system, and Dusk's version |
| License check | Once a week, while you have a license | The key and the computer's activation |
No connection carries audio, text, history or word counts. If the license check can't reach the server, the license keeps working and Dusk tries again later, so you're never locked out by the network.
Who it's for
Anyone who'd rather not send their voice to a company's servers, and people whose work doesn't allow it: lawyers, therapists, clinicians, journalists, and anyone under a confidentiality agreement. Dusk doesn't by itself make you compliant with a regulation, but nothing you dictate is sent to us or to anyone else.
Questions
Does Dusk work without an internet connection?
Yes, once its models are downloaded. Dictation, clean-up, commands and history all work offline.
Can I check what Dusk sends?
Yes. Every connection goes to api.duskdictation.com, so a firewall or network monitor shows each
one. Blocking that address leaves dictation working; only downloads, the update check and license requests
stop.
Which models does it use?
Open speech models from NVIDIA (Nemotron, Parakeet) and Moonshine, and Qwen3.5 language models. See the models and their licenses.
Where's the full privacy policy?
In the privacy policy, which also covers the website and buying a license.