Offline dictation
Dictation that works with the internet switched off.
Koegaki runs the whole speech model on your computer. Once it is installed, dictation works on a plane, on a train, behind a firewall, and on a document that must not leave the building.
macOS 14+ · Apple Silicon (M1 or later) · Download for Windows · Download for LinuxWindows 10 & 11 · 64-bit · Download for Mac · Download for LinuxDebian 12 or Ubuntu 22.04 or newer · x86_64 · Install steps · Download for Mac · Download for Windows
Why most dictation stops working offline
Most dictation tools are a microphone connected to a server. They record you, upload the audio, and wait for a transcript to come back. Lose the connection and they stop. Keep the connection and every word you say passes through someone else's computer.
Koegaki does the recognition itself. The speech model is downloaded once, on first launch, and from then on the app never needs the network to turn speech into text. When an app update brings a new speech model, that model downloads once, in the background, while the current one keeps transcribing. That background download waits while your computer is on a connection it treats as costly, such as an iPhone's Personal Hotspot on a Mac, a network you have set as metered in Windows, or one NetworkManager treats as metered on Linux. On a Mac it also does not start while the Mac is on battery below 20 percent, and goes ahead once the Mac is plugged in.
Where offline dictation matters
Legal, medical and financial documents that policy says may not be pasted into a web service. Client work under an NDA. Flights and trains. Offices where the firewall blocks unknown services. A home connection that drops. In every case Koegaki keeps typing what you say.
It also means there is nothing to sign up for. There is no account, no login, and no usage cap, because there is no server counting your words.
What still uses the network
A few things, none of them involving audio or text: the speech model download on first launch and the cleanup model download that follows it (on a Mac, only when the app runs natively on Apple Silicon), activating a licence key and a quiet daily licence check, a check with koegaki.com for a new app version, a daily check for newer models, the download of an update or a model when you choose one, and the background download of a new speech model when an app update brings one, while the current model keeps working. That background download waits while your computer is on a connection it treats as costly, such as an iPhone's Personal Hotspot on a Mac, a network you have set as metered in Windows, or one NetworkManager treats as metered on Linux. The cleanup model download waits the same way. On a Mac, neither download starts while the Mac is on battery below 20 percent, and each goes ahead once the Mac is plugged in. The Mac looks for a new app version once a day; Windows and Linux look when the app starts. Each model downloads once. If you are offline when a check would run, nothing happens; a paid licence is never revoked for being offline.
Mac, Windows and Linux, one licence
Koegaki runs on Apple Silicon Macs with macOS 14 or later, on windows 10 & 11 · 64-bit, and on Debian 12 or Ubuntu 22.04 or newer (x86_64). It costs $30 once, covers all three platforms, and includes every future update. Try it free for 30 days.
FAQ
Questions, answered.
Does dictation really work with no internet at all?+
Yes. After the first-launch model download, you can disconnect from every network and Koegaki keeps transcribing. There is no fallback to a cloud service, because there is no cloud service.
How accurate is offline recognition?+
Koegaki uses the Parakeet Ultra model on every platform: on the Mac and on Windows since version 1.7.0, and on Linux from its first release. On Mac it runs on the Apple Neural Engine; on Windows and Linux on the CPU. It handles natural speech at a normal pace; pauses are fine.
Can I use it in a corporate environment that blocks unknown services?+
Yes, once it is set up. Dictation itself makes no network requests. First launch needs to download a speech model and activation needs to reach our licensing provider, so those hosts must be allowed once. The app connects out only for licence activation and a daily licence check with Lemon Squeezy (our licensing provider), version and model checks with koegaki.com, and downloads of updates from our hosting storage, and model downloads: the list of available models comes from koegaki.com. On Mac every model file comes from our model mirror at koegaki-models.vishutdhar.workers.dev and nowhere else. On Windows and Linux the English speech, voice-detection and cleanup models come from the same mirror, and the multilingual speech model comes from GitHub releases. None of these carry audio or text.
Is there a word limit?+
No. There is no account and no server, so there is nothing to meter. Dictate as much as you like.
Stop typing what you could say.
Try Koegaki free for 30 days. Private, instant, on your own machine.
macOS 14+ · Apple Silicon (M1 or later) · Download for Windows · Download for LinuxWindows 10 & 11 · 64-bit · Download for Mac · Download for LinuxDebian 12 or Ubuntu 22.04 or newer · x86_64 · Install steps · Download for Mac · Download for Windows