Hold a key, talk, let go. Your words show up wherever your cursor is.
Free voice dictation for Linux. It runs offline, and it works on Wayland.
TuxTalk used to be called TalkType. Same app, new name.
Website • Install • New in 0.18 • What it does • All the features • How it's made • What's new
Want to hear it too? Here's the one-minute version, with sound (GitHub starts it muted, so click the speaker icon):
TuxTalk-demo.mp4
That works on GNOME, KDE, COSMIC, Sway, Hyprland and X11 with no setup, because TuxTalk's little hotkey helper reads the key straight from the keyboard instead of asking the desktop for it, so there's no shortcut to set up in your desktop's settings. It's the only part that reads the keyboard, so other programs you run can't read your typing.
It's free, it runs completely offline, and nothing you say ever leaves your computer.
Claude, Anthropic's AI, writes the code. I test it and use it every day, and I hold the code to a standard a human developer would respect. How it's made says more.
It won't type into password boxes. If your cursor is in a password field when you let go of F8, nothing gets typed, and a note over the window says why. Those words stay out of Recent Dictations too. Saying just "press enter" still works there, so you can log in after typing the password yourself. TuxTalk can only tell when the app says it's a password box: GTK apps do, so most of GNOME, Cinnamon and Xfce. KDE apps and Firefox do with accessibility switched on, and Chrome-based apps like Discord mostly don't. The desktop's own box asking for your admin password counts too, on GNOME, KDE, Cinnamon, COSMIC and Xfce.
Typing is about six times faster, and ydotool is gone. TuxTalk types with its own virtual keyboard now, so there's no background program to keep running. A 25 letter sentence shows up in about a tenth of a second instead of six tenths. You'll notice it most on X11, like Mint, where TuxTalk types everything.
COSMIC works now. On Pop!_OS 24.04 TuxTalk knows which window you're in, so a terminal gets plain text, F8 doesn't leave ^[[19~ in the terminal, and typing follows the keyboard layout set in COSMIC Settings. The same terminal fix reaches Mint, Xfce and MATE.
Fixes worth knowing about. On a keyboard that isn't US, a shortcut like Ctrl+Alt+P is the key with that letter printed on it. Alt+F8 resizes a window without starting a dictation too. And every window fits a small laptop screen with its buttons in view, dock or no dock, and on two monitors of different sizes. Emoji show up even on a computer with no emoji font, where they used to be little boxes. The changelog has the rest.
It fixes it when you change your mind. Say "no wait", "scratch that" or "actually" and correct yourself, and a small AI model on your own computer drops the wrong part. Every fix gets checked, and if the AI changed anything else, your words are typed exactly as you said them. It's optional, and it's off until you turn it on.
Commands and addresses come out right. Say "comma", "new paragraph" or "question mark" and you get the punctuation. Say an email or a web address the way you'd say it out loud and it's written the way it should be.
 |
First run looks at your graphics card and your language and picks a setup for you. On a graphics card, a sentence is typed out in about a fifth of a second. No graphics card is fine too: the default model is fast on an ordinary processor. Want something else? One button shows every choice, with what it costs you in download size and speed. Later on, Use Recommended in Preferences gets you back to it, and What works with what? shows which models run on which hardware. |
|
Names, brands, the part numbers you use every day. Put them in Your Words, in the Words tab, and they come out spelled your way, with Parakeet or Whisper, even when the model hears them split up, like "tux talk". If one still comes out wrong, say "fix that word" or pick it out of a recent dictation with Fix a Word, type it the right way, and it's fixed from then on. Each fix is kept under its word. Snippets are your own shortcuts, like "my email" for your address. |
 |
 |
Start and stop dictation, switch between Recommended, Lightest and Battery saver, or grab one of your last 20 dictations back if the text landed in the wrong window. Those are kept in memory only and gone when you log out. |
|
Your Stats shows the words you've dictated today, this week and all time, and how long it would've taken to type them. Only the numbers are kept, never what you said. |
 |
 |
Save your settings, voice commands, Fix a Word words and stats to one file, and load it on a new computer or after a reinstall. TuxTalk also keeps your last 10 settings on its own, so if a change or an update goes wrong you can just go back. |
|
TuxTalk Labs is at the bottom of Preferences, Advanced. It's for features that work but aren't finished, like Type at pause, voice editing ("edit that, make it shorter") and an indicator that follows the window you're dictating into. They're off until you turn them on, and there's a link right there to tell me how they're working for you. |
 |
I'm Ron, and I don't write code. TuxTalk is written by Claude, Anthropic's AI, and every commit says so. My part is deciding what it should do, testing it and living with it. I dictate almost everything I write with it.
I know what a lot of people think of AI-written apps, and plenty of the time they've earned it. Thing is, I don't want TuxTalk to be one of those. Keeping the code clean matters to me as much as the app working. That means nothing left lying around that isn't used, no two pieces of code doing the same job, comments that explain why something is there, and a written checklist every change has to meet before it goes in. The goal is that an experienced developer can open any file and think it's well put together, not just that it happens to work.
And it's not only my word. Over 3,700 automated tests run on every change. Before a release it gets tested in virtual machines on Ubuntu, Linux Mint, Fedora Silverblue, Fedora KDE, Zorin, Pop!_OS, openSUSE Tumbleweed and Debian, plus Nobara, which I use it on every day. The hotkey helper had a security review from another developer, guelz, who went on to contribute a feature too. If you read the code and something looks off, open an issue. I'd rather hear it.
There's more about who's behind TuxTalk on the About page.
yay -S tuxtalk-appimage
# or
paru -S tuxtalk-appimageInstalled it and it works? A vote on the AUR page helps other Arch users find it — one click if you're logged in.
Download tuxtalk_*_amd64.deb from Releases:
sudo apt install ./tuxtalk_*_amd64.debDownload tuxtalk-*.x86_64.rpm from Releases:
sudo dnf install ./tuxtalk-*.x86_64.rpmBoth packages install to /opt/talktype (the folder kept its old name so
upgrades from TalkType replace files in place), add a tuxtalk command and an
Applications menu entry, and pull in the few system libraries they need.
Download the latest AppImage from Releases:
chmod +x TuxTalk-v*.AppImage
./TuxTalk-v*.AppImageThe AppImage includes everything needed - just download and run!
Note: The AppImage doesn't need libfuse2 anymore. It does need
fusermount, which nearly every desktop already has. If it won't start, install thefuse3package:
- Fedora/RHEL:
sudo dnf install fuse3- Ubuntu/Debian:
sudo apt install fuse3- Arch/Manjaro:
sudo pacman -S fuse3- openSUSE:
sudo zypper install fuse3
There's a sandboxed Flatpak build in packaging/flatpak, but it isn't published
for download yet, and it hasn't been rebuilt since the rename to TuxTalk.
Want a Flatpak version?
👉 Vote in the poll —
enough interest and I'll make it happen.
| Requirement | Details |
|---|---|
| OS | Linux, on Wayland or X11 (GNOME, KDE, COSMIC, Sway, Hyprland and more) |
| Dependencies | None to install by hand. TuxTalk types with its own virtual keyboard, and wl-clipboard ships inside it |
| Permissions | First run asks for your admin password once, to set up the hotkey helper and let TuxTalk type into other apps. No restart needed (without systemd it uses the input group, and then it does) |
| Audio | Working microphone |
| GPU (optional) | Any NVIDIA, AMD or Intel graphics chip (Vulkan), or CUDA on NVIDIA |
Open Preferences, go to Updates & Backup and click Uninstall TuxTalk… at the bottom. It removes TuxTalk and everything it set up on your computer: the program itself, the hotkey helper and the typing permission, the speech models and everything else it downloaded (CUDA libraries included), start at login, the menu entry and the GNOME extension. It asks for your password once. Your settings and word list stay unless you tick the box to delete them too, so installing again picks up where you left off.
If TuxTalk won't open, the same thing runs from a terminal:
tuxtalk --uninstall # .deb, .rpm and AUR
~/AppImages/TuxTalk.AppImage --uninstall # AppImageRemoving only the package (sudo apt remove tuxtalk, sudo dnf remove tuxtalk,
sudo pacman -R tuxtalk-appimage) takes the hotkey helper with it from 0.17 on,
but leaves your settings and the downloads in your home folder.
Removing everything by hand (any version, including before 0.17)
Quit TuxTalk from its menu first. Then the system parts, which need your password:
sudo systemctl disable --now tuxtalk-hotkeys.service
sudo rm -f /etc/systemd/system/tuxtalk-hotkeys.service /etc/udev/rules.d/70-tuxtalk-uinput.rules /etc/udev/rules.d/99-talktype-uinput.rules
sudo rm -rf /usr/local/libexec/tuxtalk
sudo systemctl daemon-reload
sudo udevadm control --reload-rulesThen remove the package if you installed one (sudo apt remove tuxtalk,
sudo dnf remove tuxtalk or sudo pacman -R tuxtalk-appimage), and what's in
your home folder:
rm -f ~/AppImages/TuxTalk.AppImage ~/AppImages/TalkType.AppImage
rm -f ~/.config/autostart/tuxtalk.desktop ~/.local/share/applications/tuxtalk.desktop
rm -f ~/.local/share/icons/hicolor/*/apps/tuxtalk.png
rm -rf ~/.local/share/TuxTalk ~/.config/tuxtalk
rm -rf ~/.cache/huggingface/hub/models--Systran--faster-whisper-* ~/.cache/huggingface/hub/models--istupakov--parakeet-tdt-0.6b-v3-onnx ~/.cache/huggingface/hub/models--ggml-org--parakeet-GGUF ~/.cache/huggingface/hub/models--ggerganov--whisper.cpp ~/.cache/huggingface/hub/models--Qwen--Qwen2.5-1.5B-Instruct-GGUF
gnome-extensions uninstall talktype@ronb1964.github.ioThe second-to-last line also deletes your settings and word list; leave out
~/.config/tuxtalk to keep them. The last line only matters on GNOME.
Versions before 0.16 may also have set up a typing service and put you in the
input group. 0.18 and later remove that service when they start; if it's
still there, check first that ~/.config/systemd/user/ydotoold.service says
"TuxTalk keystroke injection" (or TalkType), then:
systemctl --user disable --now ydotoold.service
rm -f ~/.config/systemd/user/ydotoold.service
sudo gpasswd -d $USER inputSkip the last line if another program you use needs the input group (a key
remapper, for instance). Log out and back in afterwards.
- Launch TuxTalk - Run the AppImage or use your app launcher
- First-run setup - TuxTalk will guide you through initial configuration
- Start dictating - Press F8 (hold to record) or F9 (tap to toggle) — both always active
- Speak naturally - Text appears where your cursor is
- Use voice commands - Say "comma", "period", "new line", etc.
| Hotkey | How it works |
|---|---|
| F8 | Hold to record, release to transcribe (hold-to-talk) |
| F9 | Press once to start, press again to stop (tap-to-toggle) |
| Ctrl+Alt+P | Types your last dictation again, wherever your cursor is |
| Shift+F8 | Voice editing: hold it and say how to change the selected text (turn it on in TuxTalk Labs first) |
Want your own shortcut, or a script? Run tuxtalk --toggle (or --start, --stop). For the AppImage, use the AppImage file, e.g. ~/AppImages/TuxTalk.AppImage --toggle. It starts TuxTalk if it isn't running and waits for the speech model to load, so the toggle is never lost.
- Two hotkeys, both always on. Hold F8 to talk, or tap F9 to start and tap again to stop. Pick any keys you like.
- Hands-free if you want it. Double-tap F9, talk, and it stops by itself when you go quiet. A single tap still works as a plain on/off switch, so a long dictation with thinking pauses never gets cut off. Say nothing for 10 seconds and it stops without typing anything.
- Your choice of speech model. NVIDIA's Parakeet is the default: fast on any processor, for English and 24 European languages. OpenAI's Whisper covers 99 languages, from Tiny up to Large-v3. Not sure which? What works with what? in Preferences lays it out.
- Only these languages. Tick the languages you actually speak and Whisper's auto-detect sticks to those.
- Any graphics card. NVIDIA, AMD or Intel through Vulkan, a 22 MB download. On NVIDIA you can use CUDA instead. TuxTalk times your graphics chip against your processor first and only uses it if it's really faster.
- Corrections. Say "no wait", "scratch that", "I mean" or "actually" and the optional AI fixes it. It also drops "um", "uh" and stutters.
- Voice commands. "comma", "period", "new paragraph", "undo last word", "delete last three sentences" and more. See the full list below.
- Email and web addresses. "john at gmail dot com" types john@gmail.com, and "github dot com" types github.com.
- Numbers as digits, if you want them. 25 instead of twenty-five, and 7:30 and 15th too. Works in English and seven other languages.
- Plain text in terminals. No capital at the start and no period at the end when you dictate a command.
- Voice editing (Labs). Select text, hold Shift+F8 and say how to change it. Tone and rewrites go to the AI, while caps, lists, commas and removing a word are done exactly in code. Bold, italics and underline work in LibreOffice and other word processors.
- Select and move by voice (Labs). Say "select the last sentence" or "go to the end" on its own and TuxTalk does it instead of typing it.
- Snippets. Make a phrase type anything you want, like "my email" for your address.
- Your Words. The names and terms you use, spelled your way.
- Fix a Word. Teach it a name or term it keeps getting wrong. Or just say "fix that word" right after a dictation.
- Paste Last Dictation. Ctrl+Alt+P types your last dictation again, for when it landed in the wrong window.
- Transcribe a File. Turns a recording or a video into text or subtitles with the same model you dictate with, and keeps your last 10.
- Pauses your music and videos while you talk, and starts back up only what it paused.
- Recent Dictations. Your last 20, one click to copy. Kept in memory only.
- Your Stats. Words dictated and typing time saved, at your own typing speed.
- Settings backup. Save everything to one file and load it on another computer. TuxTalk also keeps your last 10 settings on its own, so you can roll back a change.
- A recording indicator you can style. Orb, waveform, frequency bars or radial, any color, anywhere on screen. It reacts to your voice so you know it's listening. In Labs it can follow the window you're dictating into.
- Swiss German spelling. Types ss instead of ß, on by itself when your system is set to Swiss German.
- Tells you why when it can't type. If a dictation can't be typed into the window, TuxTalk says why and what fixes it.
- Private. Everything runs on your computer, and what you dictate is never written to a log unless you turn that on for troubleshooting.
- GNOME extension. A native panel menu on GNOME, and a tray icon everywhere else.
- Updates the right way. Through your package manager for the .deb and .rpm, in place for the AppImage, or through your AUR helper on Arch.
More screenshots
Hold F8 and talk: the indicator shows TuxTalk is listening, and your words land wherever you're typing
First run recommends a setup • one button shows every choice
Model, graphics card and hotkeys • Dictation cleanup and AI self-corrections
Microphone test and recording indicator • The Voice Commands quick reference
The Words tab: Your Words and Snippets • Your Stats
Built-in help, including a page on TuxTalk Labs and voice editing
Voice commands, the full list
| Say This | Result |
|---|---|
| "comma" | , |
| "period" / "full stop" | . |
| "question mark" | ? |
| "exclamation point" | ! |
| "colon" | : |
| "semicolon" | ; |
| "open quote" / "close quote" | " " (smart quotes) |
| "dot dot dot" / "ellipsis" | ... |
| Say This | Result |
|---|---|
| "new line" | Line break |
| "new paragraph" | Double line break |
| "tab" | Tab character |
| Say This | Result |
|---|---|
| "undo last word" | Deletes the last word |
| "undo last sentence" | Deletes to the previous sentence |
| "undo last paragraph" | Deletes the last paragraph |
| "delete last three words" | Deletes several at once — words, sentences or paragraphs |
| "delete last 5 sentences" | Digits work too; counts above what you dictated are clamped |
| "clear everything" | Clears the entire input field (see warning below) |
Undo, delete and remove are interchangeable — say whichever comes naturally ("delete last word", "remove last two sentences"). Counts accept digits or the words one through ten.
⚠️ "clear everything" empties the whole field, not just what you dictated. The phrases undo/delete/clear + everything/all — six in total — all do the same thing: select all and delete. That includes text you typed yourself, so it is not limited to the current dictation.
Say "literal" before any command to output the word instead:
- "literal comma" → types "comma" (not ,)
- "literal period" → types "period" (not .)
Speech models
Choose the right model for your needs in Preferences → General:
| Model | Size | Speed | Accuracy | Best For |
|---|---|---|---|---|
| tiny | 78 MB | Fastest | Basic | Quick notes |
| base | 148 MB | Fast | Good | Casual use |
| small | 486 MB | Balanced | Very Good | Any of 99 languages |
| medium | 1.5 GB | Slower | Excellent | Professional |
| large-v3 | 3.1 GB | Slowest | Best | Technical work |
| Parakeet | 670 MB | Fast, even without a GPU | Best | Recommended for English + 24 European languages |
Tip: First run recommends a setup for your computer and language. Use Recommended in Preferences → General, or tray → Performance → Recommended for this computer, puts you back on it any time. Parakeet is the default. It's a different engine that NVIDIA made, and it runs on your processor (or on your graphics card through Vulkan). It doesn't cover Chinese, Japanese, Korean, Arabic and other non-European languages. For those, first run picks Whisper: Small on the processor, or Large-v3 on a graphics card.
Graphics cards
A graphics card makes the Whisper models much faster, and with Vulkan it speeds up Parakeet too. There are two ways to use one:
| Works on | Download | Pick it in | |
|---|---|---|---|
| Vulkan (Light) | NVIDIA, AMD and Intel | 22 MB | Preferences → General → Device → Vulkan (any GPU) |
| CUDA (Full) | NVIDIA only | 1.4 GB | Preferences → General → Device → CUDA (NVIDIA) |
On an RTX 4070 Super both were equally fast: half a second for 11 seconds of speech with the Large model. First run uses any graphics card through Vulkan. CUDA is set up from Preferences by choosing it as the Device.
Before switching to Vulkan, TuxTalk times your graphics chip against your processor and only switches if the graphics chip is clearly faster. The graphics built into some desktop processors is too small to help, and in that case it tells you and keeps using the processor.
Parakeet doesn't need a graphics card, since it's already fast on the processor. With the device set to Vulkan it runs on the graphics card anyway: on an RTX 4070 Super, a minute of speech took 0.2 seconds instead of 2, with the same accuracy. That needs a second copy of Parakeet in the graphics engine's format (669 MB). The processor copy stays too, as a fallback. CUDA can't run Parakeet.
Settings file
Settings are stored in ~/.config/tuxtalk/config.toml. To keep a copy, use Save Backup in Preferences → Updates & Backup.
model = "parakeet-v3" # AI model: parakeet-v3, tiny, base, small, medium, large-v3
device = "cpu" # "cpu", "vulkan" (any GPU) or "cuda" (NVIDIA)
hotkey = "F8" # Hold-to-talk key (hold to record, release to stop)
toggle_hotkey = "F9" # Tap-to-toggle key (press once start, press again stop)
# Both hotkeys are always active simultaneously
language_mode = "auto" # "auto", or "manual" to use the language below
language = "" # language code for manual mode, e.g. "de"
beeps = true # Audio feedback sounds
smart_quotes = true # Use curly quotes " "
auto_space = true # Auto-space between utterances
auto_period = true # Add period at end of sentencesTroubleshooting
- Look in
~/.config/tuxtalk/tuxtalk.logfor "Made TuxTalk's virtual keyboard". If it says "Can't make TuxTalk's virtual keyboard" instead, open Preferences → Advanced → Show system check and click Fix Typing Permissions - If some letters go missing on a slow computer, add
key_delay = 6(or higher) to~/.config/tuxtalk/config.tomland restart TuxTalk
- Another app may be using F8/F9 - try different keys in Preferences
- Ensure TuxTalk service is running (check tray icon)
On Wayland a normal program isn't allowed to see keys pressed in other windows, so TuxTalk needs a little help to notice F8.
- Preferences → Advanced → Typing Setup → Fix Typing Permissions
- On most systems that installs TuxTalk's hotkey helper. It's a small
background service, and it's the only thing that reads the keyboard. It
tells TuxTalk about its hotkeys and nothing else, so no other program you
run gets to read what you type. No restart needed. Check it with
systemctl status tuxtalk-hotkeys - On a system without systemd it puts you in the
inputgroup instead. Then restart the computer (logging out often isn't enough) and check withgroups
Older versions put you in the input group. That works, but the group lets
every program you run read every key you press, passwords too. The helper
fixes that. It runs as a throwaway system user with no network and no access
to your files, it never reports a letter or a number, only F-keys, Esc and
shortcuts with Ctrl, Alt or Super, and only to whoever's at the screen.
systemd-analyze security tuxtalk-hotkeys rates it 0.6, "SAFE".
A shortcut that isn't an F-key needs two of Ctrl, Alt and Super, and while those stay held the helper reports two different keys at most, so a stuck Ctrl+Alt can't turn your typing into "shortcuts". The check is per user, so two logins of the same user both count as the person at the screen.
To type, TuxTalk still needs /dev/uinput, so whoever's at the screen gets
it, the same rule Steam uses for game controllers. A program can use it to
press keys but not to read them. That's still real power: any program you run
can type into any window, the lock screen included, and so can a Flatpak app
given full device access. It's less than the input group gave, which also
let programs read every key.
Some packages give your user access to the keyboard itself. OpenRGB's rule,
for one, does that for every Logitech device. Programs can then read that
keyboard directly, helper or not. Check with
getfacl /dev/input/event*: a line naming your user means that device is
open to everything you run.
To take the helper off again:
sudo systemctl disable --now tuxtalk-hotkeyssudo rm -r /etc/systemd/system/tuxtalk-hotkeys.service /etc/udev/rules.d/70-tuxtalk-uinput.rules /usr/local/libexec/tuxtalk- Use your graphics card if you have one (see Graphics cards above)
- Try Parakeet, or a smaller Whisper model (tiny or base)
- Tray menu → Performance → Recommended for this computer
- TuxTalk offers to install its GNOME extension on first run
- Or manually: Preferences → Advanced → Install Extension
Building from source
GTK's Python bindings can't be installed with pip, so they come from your
system and the rest goes in a venv. run-dev.sh puts the two together.
# Prerequisites (Fedora/Nobara; other distros have the same packages under similar names)
sudo dnf install -y portaudio-devel wl-clipboard \
python3-gobject libayatana-appindicator-gtk3 libnotify
# Clone and install
git clone https://github.com/ronb1964/TuxTalk.git
cd TuxTalk
python3 -m venv .venv
.venv/bin/pip install poetry
.venv/bin/poetry install
# Run
./run-dev.shTo type, TuxTalk needs to make a virtual keyboard through /dev/uinput. First-run
setup's Fix Typing gives it that.
To run the tests (swap in your Python version, python3 --version shows it):
PYTHONPATH=./src:/usr/lib64/python3.14/site-packages:/usr/lib/python3.14/site-packages \
.venv/bin/python -m pytest tests -qMore in DEV_SETUP.md.
MIT License. See LICENSE for details.
TuxTalk's downloads also include other open-source software, like wl-clipboard, each under its own license. THIRD_PARTY_LICENSES.md lists them.