dictate

Hold a key, speak, let go. The words are typed into whatever window has focus. Runs on your own computer, no account, no per-minute cost.

Download for Windows

Windows 10 or 11, 64-bit. About 75 MB. The first launch downloads a 140 MB speech model, so keep an internet connection for that one time.
Releases live at github.com/MikeRuf/dictate-releases.

Install

Run the installer. It installs for your user only, so it never asks for administrator rights.
Windows will say "Windows protected your PC", because the installer is not code-signed. Click More info, then Run anyway. That dialog appears once.
Leave Start dictate when I sign in checked unless you would rather launch it yourself.
When it finishes, a microphone icon appears in the tray by the clock, and a window shows the speech model downloading. That takes a minute the first time and never again.

Dictate

Click into anything that accepts typing: an email, a chat box, a code editor, a terminal.
Hold right Ctrl and talk. A small pill at the bottom of the screen turns red while it listens.
Let go. A fraction of a second later the text appears where the cursor was.

Say comma, period, question mark or new line to get those. Right-click the tray icon and choose Open for history, settings, and a box to practice in.

Settings worth knowing

Trigger	Any key your other hand never needs, or a mouse thumb button. While the app runs, that key stops doing its normal job.
Microphone	The system default unless you pick one. The level meter on the Status tab shows whether the app hears you.
Engine	faster-whisper is accurate and needs no setup. Moonshine is quicker on short phrases. Either downloads its model the first time you pick it.
Recordings	Every dictation is kept as audio, so a bad transcript can be replayed. Right-click a history row to play, open, or delete it. There is a cap on how many are kept.

Known limits

Windows only. The keyboard hook and the typing are Win32.
Programs running as administrator do not receive the text, and the trigger key does not fire while one of them has focus. Windows blocks input from a less privileged process. Run dictate as administrator if you need it in an admin console.
English by default. The large-v3-turbo model is multilingual, at 1.6 GB and a second or two per dictation.
The first launch needs internet for the model download. After that it runs offline.
Not code-signed. SmartScreen asks once (see Install). A PC with Smart App Control turned on refuses unsigned programs altogether; it is on only for some clean Windows 11 installs, and Windows Security, App & browser control, shows its state. Turning it off is permanent, so decide for yourself whether that is worth it.

If something is wrong

Open the window from the tray icon. The Status tab says whether the engine loaded and whether the microphone hears you. If the engine failed, the file errors.log next to the config (the Settings page has an Open config file button; the log sits beside it) says why. Send that file along with what you expected to happen.

Two things that look like the app being broken and are not: the tray icon hidden behind the small chevron by the clock, and Windows privacy settings not letting desktop apps use the microphone (Settings, Privacy and security, Microphone).

Microphone works elsewhere but not here? The Status tab's level meter should move when you talk. If it does not, the app is probably listening to a different device than the one you use: pick yours by name in Settings, Audio. To test the microphone without the app, save mic-test.ps1 and run it from a PowerShell window with
powershell -ExecutionPolicy Bypass -File mic-test.ps1
It shows the privacy switches, lists the input devices, and draws a live level bar from the default microphone for ten seconds. It uses the same Windows path a desktop app does, so what it hears, dictate can hear.