Getting started
- Open Dictation in the sidebar and turn it on.
- Follow Setup:
- Speech model: download
Parakeet-TDT-0.6B-v3if it is not installed. Izwi keeps the dictation model loaded for as long as dictation is on, so the first word is as fast as the rest. - Keyboard control (macOS only): allow Izwi under System Settings › Privacy & Security › Accessibility. Izwi needs it to paste for you.
- Try it in the box on the page.
- Speech model: download
- Click into any app, hold the shortcut, speak, and release.
Izwi must be running. Closing its window keeps it in the tray, and Settings › Launch at login starts it with your computer.
Using it
Under Shortcut › How it works you can switch from holding the key to pressing it once to start and once to stop.
A small indicator near the bottom of the screen shows when Izwi is listening and when it is transcribing. A recording stops by itself after five minutes.
Choosing a shortcut
A shortcut is one key with modifiers, such asCtrl+Shift+Space, or a function key alone, such as F9. Click the shortcut box and press the keys, together or one after another. To use modifier keys alone, press them and wait a second.
On macOS and Windows it can also be modifier keys alone: hold ⌥, Ctrl+Shift, or Fn (macOS). If you press another key while holding them, Izwi treats it as a normal keyboard shortcut and discards the recording. Modifier-only shortcuts need the Accessibility permission on macOS and stop working while a password field has focus; shortcuts with a key do not have that limit.
How the text is inserted
By default Izwi puts the text on the clipboard, pastes it, and then puts your previous clipboard back. The dictated entry is marked so clipboard managers and Windows clipboard history skip it. Text › Insert text by offers two alternatives:
When Izwi cannot insert the text, it leaves it on the clipboard and the indicator says so. This happens when:
- Izwi has no Accessibility permission (macOS).
- The focused window runs as administrator (Windows). A normal app cannot send keys to it.
- The desktop offers no way to send keys (some Wayland desktops, see below).
Speech
Typical wait between releasing the key and seeing text, measured with Parakeet on an Apple Silicon Mac: about half a second for ten seconds of speech. On computers without a supported GPU the model runs on the CPU and takes longer.
Live preview
Models that transcribe while you speak (Nemotron-3.5-ASR-Streaming-0.6B, Voxtral-Mini-4B-Realtime-2602, LFM3-Audio-2B-GGUF) show the words in the indicator as you say them, and the wait after releasing the key stays short however long you spoke. Turn it off under Feedback › Live preview.
Text
Clean up with AI adds a short delay (under a second with
Qwen3.5-2B on Apple Silicon). If the model takes longer than 1.5 seconds, Izwi inserts the text as it was heard. Replacements are applied after the clean-up.
History
Izwi keeps the text of your recent dictations on this computer so you can copy one again. Audio is never stored. Turn it off under History › Keep history, or clear it there.Linux
On X11 dictation works as on other systems. Typing instead of pasting needsxdotool.
On Wayland the desktop controls shortcuts and key input:
If the Dictation page says the shortcut could not be registered, bind a key in your desktop’s keyboard settings to:
start, stop and cancel are also accepted. The command talks to the Izwi that is already running.
Without wtype, ydotool or dotool, and without the desktop’s permission, Izwi leaves the text on the clipboard for you to paste. Modifier-only shortcuts are not available on Linux.
Troubleshooting
Each dictation’s timing is written to
izwi-desktop.log in the app’s log folder, without the text.