Click the button, speak, click again. Your words are typed wherever your cursor is. Recognition runs on your own CPU — no cloud, no account, no audio sent anywhere.
Free & open source · GPL-3.0 · No telemetry, ever
One button, three moments, about a third of a second of thinking.
Click the floating button, or the icon in your tray, from any app: email, editor, browser, terminal.
Say the thing. The button shows a live waveform while it listens, and a turning arc while it thinks.
Your words appear where the cursor is, filler words trimmed and your own spellings applied.
Speech recognition runs on your CPU using NVIDIA's open Parakeet model. After the one-time model download, the app pins itself offline — unplug the internet and Talkin will not notice. It makes exactly two kinds of network request in its whole life: fetching the model once, and checking for updates when you ask it to.
Click the small button beside it, type what it misheard and the right spelling once. Never see that mistake again. Your dictionary exports and imports as a file.
English, French, German, Spanish, Italian and Dutch. Every visible string lives in one editable file — add a column, get a language.
Claude Code has voice input built in. Talkin is clearer and more accurate in everyday use, and it is not confined to one window: it types into an email, a terminal, a form in the browser, anywhere your cursor is. Built and used daily by someone who dictates for a living.
A single AppImage. Mark it executable and run it. It updates itself from Settings, only when you press the button.
One download, two very different jobs underneath. Typing into another app is something a Wayland compositor guards closely, and rightly, so there Talkin asks through the standard desktop portal and the compositor stays in charge — no ydotool, no joining the input group, no udev rules. On X11 no permission exists to ask for, so it simply types. Talkin works out which it is on and picks its own way in; you are told plainly if a desktop cannot support it.
Keyboard shortcuts are gone. On Wayland the compositor's press and release signals proved unreliable enough to lose whole dictations, and one habit that works everywhere beats two that differ. So there is a floating button you can drag anywhere, plus the tray icon. A click has nothing to go wrong.
On X11 you can highlight a misheard word and open the correction popup with it already filled in. Wayland does not let a background app read your selection — deliberately — so there the popup asks you to type the word instead.
Talkin runs on Wayland and on X11. It is one AppImage; there is nothing to choose between.
What it needs depends on which you are in. To see which:
echo $XDG_SESSION_TYPEwayland — your desktop must provide the RemoteDesktop portal, which is how Talkin is allowed to type: GNOME 45+ or KDE Plasma 5.27+. You approve it once, on first run.x11 — any desktop at all, including Cinnamon, XFCE, MATE and bare window managers. Nothing to install and nothing to approve.Talkin says so plainly if a desktop cannot support it, rather than failing quietly. You will also need PipeWire or PulseAudio for the microphone, and a 64-bit (x86_64) machine.
Mark the file executable (right-click → Properties → “Allow executing” on most desktops), then open it. On first run it downloads the speech model once, about 600 MB, and works fully offline afterwards. Your desktop will ask once for permission to type into other windows — that is the portal doing its job.