Speech to Text on Linux: Dictation on Ubuntu and Mint

Ubuntu and Linux Mint do not include a dictation feature that we could find in their current release documentation. To dictate on Ubuntu or Mint, you install an open-source app. That app runs a speech engine on your own computer, usually Whisper or Vosk, and types the recognized words into whatever window has focus. The engine is the easier part. Vosk's portable language models are about 50 MB each ¹, and Whisper's small model needs about 2 GB of graphics memory ². The harder part is the typing step. On Ubuntu 26.04 LTS, Wayland is the default session on modern graphics hardware ³, and xdotool, the classic typing tool, does not work there ⁴.

So the answer depends on your session. On a Wayland desktop such as Ubuntu 26.04 with GNOME, use a tool that types through ydotool or lists it as a method. On an X11 session, the xdotool-based tools work. This guide covers the tools we could verify, the install route for each, and the setup that a tested Ubuntu 26.04 configuration needed.

Yap, which we make, runs on Mac, Windows and phones, not on Linux, so the tools below are open-source projects made by other people. If you only need dictation in a web browser, the free dictation page in your browser works without an install.

Linux desktops have no built-in dictation, so you add an open-source tool

The roadmap for Ubuntu 26.04 LTS and the Linux Mint 22.3 release page do not describe a dictation feature, so the dictation tools in this guide are third-party apps.

Linux Mint 22.3 is the release we checked, and it has an Ubuntu Noble package base ⁵. Because Mint shares that package base, Ubuntu-built packages usually install on Mint 22.3 too. The desktop matters more than the packages. Cinnamon, Mint's desktop, is due to get fully supported Wayland sessions in Mint 23.0, which is planned for Christmas 2026 ⁶. Until then, Wayland on Mint is the less settled choice.

Dragon NaturallySpeaking has no Linux edition on Nuance's pages

Dragon is a widely known dictation product, so people ask about it first. Nuance's Dragon Professional page describes version 16 as optimized for Windows 11 and backwards-compatible to Windows 10 ⁷. We found no Linux edition on Nuance's pages. Our full answer, with the alternatives, is in the question on Dragon on Linux. For a wider comparison of free options, see the guide to open-source dictation software.

Offline engines like Whisper and Vosk do the recognition on your own machine

An offline engine runs the speech model on your computer, so the audio does not have to leave the machine once the model is downloaded. Two engines appear most often in the open-source tools covered here.

Whisper is OpenAI's speech recognition model. Whisper Large v3 had an average word error rate of 7.44 percent on one track of the Open ASR Leaderboard ⁸. That figure comes from one benchmark, and your own voice in your own room may score differently. Word error rate is explained in our guide to word error rate.

Whisper also has a known failure mode. In a 2024 study of recordings from speakers with and without aphasia, using a 2023 version of Whisper, roughly 1 percent of audio transcriptions contained entire phrases or sentences that did not exist in the audio ⁹. Read your dictated text before you send it.

Vosk is a smaller engine from Alpha Cephei. Its portable per-language models are about 50 MB each, and the project also offers much bigger server models ¹. The small size is why Vosk suits modest hardware. The trade-off is accuracy. The Vosk home page gives no word error rate, so we cannot put a number on Vosk's accuracy here.

Whisper's cost depends on the model you choose. The table sets the engines side by side. The graphics memory figures come from the Whisper model table and are approximate.

Typing the text into apps is harder on Wayland than on X11

Recognition works on Linux. Typing the result into the focused app is where dictation most often breaks, and the reason is the display system. X11 lets one program send keystrokes to another. Wayland does not offer that general mechanism, so a dictation tool needs a different route.

The xdotool README is direct about this. It says that on Wayland, "Typing, window searching, and many other functions of xdotool do not work, and it is unclear if they could ever work" ⁴. The nerd-dictation project uses xdotool as its default input tool ¹⁰, so that tool fits an X11 session and not a Wayland one.

ydotool takes a different route. It emulates an input device through the Linux kernel's uinput framework, so it works below the display server, and its daemon needs access to /dev/uinput ¹¹. Vocalinux, a dictation app covered below, lists five text-insertion methods for both X11 and Wayland: xdotool, IBus, wtype, ydotool and a clipboard fallback ¹².

A tested Ubuntu 26.04 setup uses ydotool for typing

Handy, a dictation app covered below, publishes a troubleshooting guide for Ubuntu 26.04. The guide was tested on Ubuntu 26.04.1 LTS with GNOME on Wayland and Handy 0.9.8 ¹³. Its fix is to use ydotool for typing and handy_keys for shortcuts ¹⁴. The steps below follow that guide. The full guide on GitHub has the exact wording.

1. Confirm that you are in a Wayland session. Run echo "$XDG_SESSION_TYPE" in a terminal. The answer should be wayland.
2. Add your user to the input group, because the uinput device belongs to that group. Run sudo usermod -aG input "$USER", then log out and back in.
3. Install and start ydotool. Run sudo apt install ydotool, then systemctl --user start ydotool.
4. Test the typing. Run ydotool type "HELLO FROM YDOTOOL" with a text field focused. The words should appear in that field. If nothing appears, check the group membership in step 2 before you change anything else.
5. Configure Handy to type with ydotool. Step 3 of the guide gives the command, and the settings key can change between versions, so follow the guide's wording for your release.

The input group matters for security. Membership gives your user access to the input device that ydotool writes to, so add only accounts you trust.

Command-line tools suit developers, and desktop apps suit everyone else

There are two kinds of Linux dictation tool. Command-line tools run from a terminal or bind to a key. Desktop apps have a window, a settings screen and an installer. The choice depends on how much setup you want to do.

nerd-dictation is the best-known command-line option. It describes itself as offline speech to text for desktop Linux, built on the VOSK-API ¹⁵. Its README lists Windows and macOS support only as a possible future item ¹⁶, so it is a Linux-only choice. Use it on an X11 session, since its default input tool is xdotool.

Handy is a desktop app for offline speech to text. Its README lists Windows, macOS and Linux as supported platforms ¹⁷.

Vocalinux is an open-source Linux app for offline dictation. After you download a model, its local engines run speech to text on your machine. The default is whisper.cpp, and OpenAI Whisper, Faster Whisper, VOSK and Parakeet are also available ¹⁸.

Speech Note is a Linux desktop app for note taking with offline speech to text. Its README gives the Flatpak route through Flathub, with the package name net.mkiol.SpeechNote ¹⁹. Its speech to text uses several engines, including Vosk and whisper.cpp with Parakeet ²⁰. You need Flatpak installed first.

The maintained tools differ by engine, desktop and install route

The table compares the tools we checked against their own README pages. Each row names what its documentation says, so you can check it before you install.

The table cells cannot carry citations, so the facts behind them are cited here. Handy's README lists its Whisper models in four sizes ²¹, and it sends you to the releases page for the download ²². OpenWhispr's README names macOS, Windows and Linux ²³. HEX's Linux beta has a smaller feature set and needs a source build or Nix ²⁴. For Murmure, the README tells you to download the Windows installer, Murmure_x64.msi or Murmure_x64-setup.exe ²⁵, so we list it as a Windows app.

Our order of preference for an Ubuntu 26.04 user on GNOME is Handy or Vocalinux. Handy has a tested guide for this exact setup, and Vocalinux lists Wayland support and ydotool among its typing methods. For a Mint 22.3 user on an X11 session, nerd-dictation is the lightest option. Speech Note suits anyone who wants a graphical app with a choice of engines. These are judgments from the documentation. We did not run every tool on a test machine, so try your choice on a short note before you rely on it.

A GPU makes large Whisper models practical, and small models run on a CPU

The Whisper model table gives the required graphics memory for each size. The figures are approximate, and they describe GPU memory, not system memory. The medium model needs about 5 GB ²⁶, the large model about 10 GB ²⁷. The same table rates the tiny model at about 10 times the relative speed of the large model, which it rates 1x ²⁸.

The chart shows the graphics memory each size needs. A GPU with less memory than the figure cannot hold that model at full size, so the tool has to fall back to a smaller model or to the CPU.

You do not need a GPU to run Whisper. whisper.cpp lists support for CPU-only inference among its features ²⁹. The table's relative speeds are ratios, not times, so they do not tell you how long a CPU run takes. The question on running Whisper without a GPU covers this in more detail, and our guide to running Whisper locally covers the setup.

Browser dictation covers Google Docs in Chrome, but not your whole desktop

Browser dictation is the simplest route when the text you need goes into a web app. Google's voice typing in Google Docs works with the latest versions of Chrome, Edge and Safari, and the browser controls the speech to text service ³⁰. So the browser's speech to text service handles the speech, not an engine you install. If you need on-device dictation, use a desktop tool from the sections above.

Google's Chrome help page says Chrome is compatible with devices that run Windows and Mac operating systems ³¹. It does not name Linux, and we did not find a Linux statement on the pages we opened. Before you rely on voice typing in Docs on Ubuntu or Mint, test it in the browser you use there. The guide to Google Docs voice typing covers the menu steps, and the guide to speech to text in Chrome lists the other options.

Browser dictation does not reach the rest of your desktop. A chat window, a terminal or a note app that is not a web page will not receive the text. The guide to dictating into any app compares the system-wide options on each platform.

Voice control on Linux is handled by Talon, which lists X11

Dictation turns speech into text. Voice commands, such as opening a window or pressing keys by voice, are a different job, and the question on voice commands on Linux covers that in full. The main option we found is Talon, whose home page lists Linux with the note X11 ³². That note means the Linux build is documented for X11, so check Talon's documentation before you use it on a Wayland session. Talon Voice is built around hands-free control, so for dictating prose, the desktop tools above are the better fit.

What to install, depending on your desktop

The right tool depends on two things: whether your session is X11 or Wayland, and whether you want a graphical app or a terminal tool.

• Ubuntu 26.04 on GNOME, Wayland: Start with Handy and the ydotool setup above, which was tested on this configuration. Vocalinux is the alternative, because it lists ydotool among its typing methods.
• Mint 22.3 on Cinnamon: Check your session type first. On an X11 session, xdotool-based tools such as nerd-dictation fit. On a Wayland session, use the ydotool route. Fully supported Wayland sessions in Cinnamon are planned for Mint 23.0 ⁶.
• A computer without a graphics card: Use Vosk through Speech Note or nerd-dictation, or whisper.cpp with a small model. Vosk's models are about 50 MB each ¹.
• Dragon or a Windows-only workflow: Dragon does not run on Linux as a supported product. See the question on Dragon on Linux, and the guide to offline speech to text on Windows if you can use a Windows machine.

Whatever you install, speak a sentence, check that the words appear in the right window, and then dictate a longer passage. Recognition errors and typing failures show up at different points, so you want to know which one you have. The word error rate guide covers how accuracy is measured if you want to test your own voice.

Linux dictation is less polished than on Mac or Windows. The engines are the same kind, but the display system, the desktop and the input permissions vary. Check the release notes of the tool you pick, and re-test after an Ubuntu or Mint upgrade, because the session and input rules can change between releases.