Install

macOS

Requires macOS 12 or later on a Mac with Apple silicon (M1 or newer). In Terminal:

Terminal

curl -fsSL https://raw.githubusercontent.com/baney75/silas/main/install.sh | sh

The installer copies Silas to /Applications and opens it. The first screen starts downloading the speech model (478 MB) while you grant two permissions: Microphone, and Accessibility, which lets Silas read its key and type for you.

Stable releases are macOS only, because the Mac app is the one each release is tested on. Windows and Linux get nightly builds made from the newest code.

Windows (nightly)

Requires Windows 10 or 11 on x64. In PowerShell:

PowerShell

irm https://raw.githubusercontent.com/baney75/silas/main/install.ps1 | iex

It installs for your account only, with no administrator prompt. Silas is not code-signed on Windows yet, so SmartScreen may ask you to confirm the first time.

Linux (nightly)

Requires an x64 distribution. Run the same command as on macOS. It installs the .deb or .rpm when it can, otherwise an AppImage. Silas reads its key from the keyboard device, so the installer adds a udev rule that grants that access to the signed-in user. On Wayland, install wtype so Silas can type into native apps.

Dictate

Do this	And Silas
Hold Right Command (Mac) or Right Ctrl (Windows, Linux), speak, release	types what you said where your cursor is
Tap the key once	listens hands-free until you tap it again
Press Esc while it listens	cancels
Press the key with another key, like Right Command and C	does nothing, so your shortcut still works

A small pill at the bottom of the screen shows your voice level while Silas listens. New lines are typed as Shift+Return, so chat boxes don't send early. Silas removes "um" and "uh" and collapses stutters like "the the the".

Silas types real keystrokes. It never pastes, so apps don't ask for paste permission and your clipboard is left alone. Settings, Typing speed has Steady, Fast and Instant; pick Steady if an app drops letters.

Teach it your words

Fix it where it typed. Right after Silas types, backspace over the wrong word and type the right one. A name it didn't know, like "Ghostty", is learned at once. An ordinary word is learned the second time you make the same fix, and the first sighting waits under Dictionary, Learning.
Fix it in History. Choose Teach a fix on any dictation, edit the text and save.
Words. Names and terms Silas should spell your way. It matches near misses only for rare words, never everyday English.
Corrections. "When you hear this, write that", for phrases like "ghost E" to "Ghostty".
Import and export. Dictionary, Export saves a JSON file to Downloads. Import reads a Silas export or a TypeWhisper dictionary export.

Following edits works with a US keyboard layout. With other layouts, teach fixes in History.

Models

Model	Size	Notes
Parakeet v3 (default)	478 MB	Fast and accurate. About 12 times faster than real time on an Apple silicon CPU. English and 24 European languages.
Cohere Transcribe	2 GB	Cleaner punctuation, about 5 times slower. Best for short notes.

Silas keeps the model loaded, so the first dictation after a break is as fast as the rest. Switch or remove models on the Model page.

Updates

Settings, Updates picks the channel. Stable (macOS only) follows tested releases. Nightly follows the newest build on every platform. Silas checks at launch and every four hours, and with auto-update on it installs once it has been idle for two minutes. Every update is signed, and Silas refuses one that isn't.

Mac nightly builds are not signed with a developer certificate, so macOS asks for Accessibility again after each nightly update. Stable builds keep their permissions.

Privacy

Audio never leaves your computer, and Silas listens only while the key is held or hands-free mode is on. History and the dictionary are plain JSON files in Silas's app-data folder: ~/Library/Application Support/Silas on a Mac, %APPDATA%\Silas on Windows and ~/.local/share/Silas on Linux.