How do you use Whisper for dictation on Linux?

Whisper is OpenAI's speech recognition model, and on Linux you use it through an app that loads the model and types the text. The app does the typing, so the choice of app matters as much as the model. For dictation into other windows, Vocalinux is the route we checked most closely, since it runs Whisper's whisper.cpp engine on your own computer. The sections below cover the session type that decides whether typing works, the memory a model needs, the errors to check, and a newer Ubuntu option.

Vocalinux runs Whisper on your computer

Vocalinux uses whisper.cpp by default, with OpenAI Whisper, Faster Whisper, VOSK and Parakeet as other engines ¹. It is a free, AGPL-3.0-licensed desktop app for X11 and Wayland ². After you download a model, the speech runs on your machine.

If you want to transcribe recordings instead of dictating, check the route in our guide to running Whisper locally. That is a separate job from typing into other windows.

Your session type decides whether the typed text lands

On Ubuntu 26.04, Wayland is the default session across modern graphics cards ³. The xdotool README says its typing does not work on Wayland, and that it is unclear whether it could ⁴. A tool that depends on xdotool for typing would fail in that Wayland session.

Check your session before you install anything. Run echo "$XDG_SESSION_TYPE" in a terminal. Our guide to speech to text on Linux covers the typing steps for each session type.

Model size sets the memory you need

Whisper's largest model, large, has 1,550 million parameters and needs about 10 GB of memory ⁵. Smaller models need less, so the choice is a trade between accuracy and what your computer can hold. Start with a smaller model, and move up only if the errors are too frequent. We did not measure how fast any model runs on a particular machine, so test the speed on yours. The question on running Whisper without a GPU covers the CPU route.

Whisper makes errors you have to check

On one benchmark track, Whisper Large v3 averaged a 7.44 percent word error rate ⁶. Word error rate explains what that figure counts. A benchmark is not a promise for your voice in your room.

The bigger risk is invented text. In a 2024 study, roughly 1 percent of Whisper transcriptions contained entire phrases or sentences that were not in the audio ⁷. Read dictated text before you send it, and take extra care with records or legal documents.

Ubuntu users can also try Canonical's Myna

A September 2026 how-to says the current Ubuntu 26.04 release can also try Myna, a voice typing app from Canonical ⁸. The article is from a news site, not from Canonical, and we did not install Myna. Treat it as an option to test beside Vocalinux, not as a settled choice.

It depends on your desktop and your job

• Dictation into any window on X11 or Wayland: Vocalinux, which we checked most closely for this answer.
• Transcribing recordings: A different route, covered in our guide to running Whisper locally.
• Limited memory: A smaller model, with more errors to check.
• Ubuntu 26.04: Myna is worth a test, with the caveats above.

What to do

1. Run echo "$XDG_SESSION_TYPE" and note whether you are on X11 or Wayland.
2. Install Vocalinux and follow the install guide. On Wayland, use the typing setup in that guide.
3. Start with a smaller model, and check that the typed text lands in the window you expect.
4. Dictate a fixed test paragraph, and count the errors and any invented sentences.
5. Re-test after an Ubuntu or Mint upgrade, since session and typing rules can change.

Yap, which we make, runs on Mac, Windows and phones, not on Linux. If you need offline dictation on a Mac or Windows PC, see the download page. For Whisper on those systems, see Whisper on Mac and installing Whisper on Windows.