Thanks for sending the drawings over. I went through them this morning and the layout looks good to me.
0:00
Try itHold the key below and watch what gets typed.
Press and hold it, or hold F8 on your keyboard.
What happens when you hold the key
F8
0:02
“couldwemeetonthursdayquestionmark”
On your computer
Could we meet on Thursday question mark
Could we meet on Thursday?
Message
Could we meet on Thursday?
1
Hold F8
Or whatever key you picked. A little indicator pops up so you know it's listening, and if music or a video is playing it pauses.
2
Talk
Just talk normally, at your own pace. Stopping to think is fine, it keeps listening as long as you hold the key.
3
Let go
The speech model turns what you said into text right there on your computer. Then TuxTalk tidies it up, so "question mark" turns into an actual question mark.
4
It types it for you
Your words get typed wherever your cursor is, a browser or a terminal or whatever has the focus. On a graphics card a sentence takes about a fifth of a second.
It all runs off a few keys
F8 and F9 are both on all the time, so you just use whichever one fits what you're doing. Don't like those keys? Pick others in Preferences, from F1 to F12 or the four below.
F8
Hold to talk
Hold it while you talk. When you let go, TuxTalk types what you said wherever your cursor is.
F9
Tap to start, tap to stop
For the longer stuff. Double-tap it and it goes hands-free, then stops by itself when you go quiet, even with a fan or the dryer running. Say "press enter" at the end and your message gets sent too.
Ctrl+Alt+P
Type it again
Something grabbed the focus and your words went into the wrong window? Click where they belong and press this.
Shift+F8
Voice editing Labs
Select some text, hold it and say what to change, like "make this more formal". It's off until you turn it on.
Left Ctrl+Space
Left Ctrl+Space
Hold Left Ctrl, then Space, talk, and let go of either one. Somebody asked for it in Discussions. On KDE, GNOME with the TuxTalk extension, COSMIC and X11 your apps don't get Ctrl+Space while dictation's on, and Preferences warns you if something like an input method already uses it.
Right CtrlorRight AltorScroll Lock
Keys that don't need Fn
On a lot of laptops the F-keys only work while you hold Fn, which makes holding F8 a two-handed job. Pick one of these in Preferences, General and you don't need Fn. Right Ctrl and Right Alt stop working as Ctrl and Alt once they're TuxTalk's, and if Right Alt is your AltGr key the list tells you which characters you'd lose, like @ and €. TuxTalk may ask to update its hotkey helper first. That's your password once and no restart.
Say it the way you'd say it out loud
Punctuation goes in on its own. Say "comma" or "new paragraph" when you want one in a certain spot, and email addresses come out written the way they should be.
You sayturns intoTuxTalk types
"send the plans to john at gmail dot com"
Send the plans to john@gmail.com.
"new paragraph could we meet on Thursday question mark"
Could we meet on Thursday?on a new paragraph
"I have three cats and twenty five chickens"
I have 3 cats and 25 chickens.with Write numbers as digits turned on
"git status", in a terminal
git statusno capital and no period, since it's a terminal
"see you at noon press enter"
See you at noon.and then a real Enter, so the message goes
"let's say three o'clock, no wait, make that four o'clock"
Let's say four o'clock.with the AI corrections, which are optional and off until you turn them on
The AI that catches "no wait" and "scratch that" is a small model that runs on your computer. If it changes anything besides the correction, your words get typed exactly the way you said them. There are voice commands for fixing things too, like "undo last word" or "delete last three sentences".
It picks a setup for your computer
The first time you start it, TuxTalk looks at your graphics card and the language you speak and picks a setup for you. One button shows every other choice, with how big each download is.
The default speech model is NVIDIA's Parakeet. It's fast even on an ordinary processor and handles English plus 24 European languages. For anything else there's OpenAI's Whisper, which knows 99 languages, from Tiny all the way up to Large-v3.
Got a graphics card? NVIDIA, AMD or Intel, TuxTalk can use it. That's a 22 MB engine plus a 669 MB copy of Parakeet made for the graphics card. It times your card against your processor first and only switches if the card is really faster. On an NVIDIA card there's also CUDA, a 585 MB download, but that one only runs the Whisper models.
On the processor2 seconds
On the graphics card0.2 seconds
A minute of speech with Parakeet, same accuracy both ways. Timed on my RTX 4070 Super.
The first-run screen recommends a setup. Use Recommended in Preferences gets you back to it later.
Teach it the words you use
Names and part numbers, the stuff a speech model has never heard of. Put them in Your Words, in the Words tab, and they come out spelled your way. It steers Parakeet and Whisper both, on the processor or the graphics card. Parakeet kept writing my app's name as "Tuck's talk". Now it's TuxTalk.
If a word still comes out wrong, say "fix that word" right after. A window opens on what you just said, you click the word and type it the right way, and from then on it's fixed. It fixes it in your text too, and the Words tab keeps each fix under the word it belongs to.
Snippets are your own shortcuts. Say "my email" and get your whole address typed out.
Transcribe a File
Got a recording or a video? Hand it to TuxTalk from the tray menu and you get the text back. MP3, M4A, WAV, MP4, MKV and a bunch more all work.
It uses the same model you dictate with, so there's nothing new to download, and you can keep dictating while it works. On my NVIDIA card with Parakeet, three and a half minutes of talking took about nine seconds.
You can copy the text or save it as subtitles for the video. The subtitle timing can be a second or two off, which is fine for most things but I wouldn't use it for a film. It also keeps your last 10, so closing the window doesn't lose anything.
shop-meeting.mp33 min 31 s of talking
0:00Okay, first thing on the list is the cabinet by the back door.
0:31The wiring can go behind the panel, so we'll run it before the walls go in.
1:12Let's get the parts ordered this week.
shop-meeting.txtshop-meeting.srt
An example, not a screenshot
TuxTalk Labs
Labs is where new stuff goes while I'm still working on it. It works, it's just not finished, so it stays off until you turn it on. You'll find it at the bottom of Preferences, Advanced, with a link right there to tell me how it's going.
While you talk
Pick one thing to happen while you're still talking. "Show the words in a box" puts what you've said so far in a little box by the recording indicator, so you can watch TuxTalk keep up. It needs the speech model on a graphics card, and what gets typed when you let go is the same as always.
"Type each phrase when you pause" types each phrase when you pause, instead of all at once when you're done. Put it together with hands-free on F9 and you can say a message, look it over, say "edit that, make it friendlier", and say "press enter" to send it. No keyboard at all. guelz built this one.
The box came out of the poll in Discussion #24. If you try it, tell me there how it's going.
Voice editing
Select some text, hold Shift+F8 and say what you want changed. Or skip the selecting: right after you dictate, say "edit that, make it shorter". There's no fixed wording, and Ctrl+Z puts it back if you don't like it. Jobs like all caps or taking out commas are done exactly in code. Rewording goes to a small AI model that runs on your computer, and that one's a download you only get if you turn it on.
Select and move by voice
Say "select the last sentence" or "go to the end" on its own and TuxTalk does it instead of typing it. Put that together with voice editing and you can select your last sentence and say "make it more formal" without reaching for the mouse.
Your apps, windows and desktop by voice
Tick "Apps, windows, keys and the mouse by voice" and a dictation that's only a command does it instead of typing it. "open Firefox", "switch to Firefox", "snap left", "close window". In the browser it's "new tab", "go back a page", "go to youtube dot com" or "web search cheap tires". Then there's "volume down", "pause music", "lock screen" and "take a screenshot", and if you forget one, "what can I say" brings up the list. "I opened Firefox yesterday" still gets typed like always, and "literal" in front of a command types the words.
KDE gets all of it, GNOME too with the TuxTalk extension, and X11 desktops like Mint get nearly all of it. On COSMIC and Sway the window commands can't work yet and TuxTalk tells you so.
Click with your voice
Say "mouse grid" and the screen splits into nine numbered parts. Say a number and that part gets its own grid, and on KDE, GNOME and Cinnamon the desktop zooms in on it so you can actually see the button you're after. "left 3" nudges the yellow dot a few pixels, then say "click". If TuxTalk can't put the pointer exactly where the grid shows, it doesn't click at all.
"mouse grid", then a number or two, then "click".
You selectand sayIt changes to
hey can u send me the report by friday thx
make this more formal
Could you please send me the report by Friday, thank you.the AI reworded it
milk eggs bread
make this a bulleted list
- milk - eggs - breaddone exactly, no AI
the deadline is tomorrow
make this all caps
THE DEADLINE IS TOMORROWdone exactly, no AI
And a bunch of smaller stuff
Fits the sentence you're in
Click into the middle of a sentence and dictate, and it carries on: no capital, a space where one's missing, no period stuck in front of the next word. It works in apps that tell your desktop's accessibility system what's around the cursor, and TuxTalk asks before it turns that on.
Hears your voice, not the room
TuxTalk listens for a voice now, not just for sound. A fan, a dryer or dishes in the sink don't keep hands-free going, and a cough after you finish mostly doesn't get typed as "Yeah." It's built in, so there's nothing extra to download.
Changed your mind?
Press Esc while TuxTalk is still working and whatever hasn't gone in yet is dropped, a "press enter" included. On a fast graphics card a short sentence is often in before you can reach Esc, so say "delete last sentence" to take it back. Anything already in the box stays put.
Quotes
Say "he said quote hello quote" and hello goes in quotes. "I want to quote him on that" stays words.
Pauses your music
Music and videos pause while you dictate and pick back up when you're done. I had a YouTube video going once and it turned "isn't it" into "Venezuela".
Turns other sounds down
For what pausing can't reach, like a Discord call or a game. It's off until you turn it on, and it only changes what you hear. A call still hears your mic.
Says when the mic is silent
Muted mic, unplugged headset, wrong input picked? TuxTalk tells you it heard nothing but silence instead of just typing nothing.
Sends the message for you
Turn on Send each dictation automatically in Preferences, Advanced and TuxTalk presses Enter after a short countdown, so a chat message just goes. Tap F8 during the countdown if you want to fix something first. It won't send in a text editor, where Enter only starts a new line, or in a terminal, where it would run whatever you said, and if you switch windows before it's done counting nothing gets sent.
Takes out the ums
There's a setting in Preferences, Advanced that drops "um" and "uh" and the words you say twice by accident, so "um, so the the van is ready" comes out "So the van is ready." It leaves the ones you mean alone, like "no, no, no".
Easy on a laptop battery
On a laptop it lets go of the speech model when you haven't dictated for a while. F8 still works, the first dictation after just takes a few seconds longer. On a desktop that's off.
Recent Dictations
Your last 20 are in the tray menu, one click to copy. They're only kept in memory and they're gone when you log out.
Your Stats
How many words you've dictated and about how much typing that saved you. Only the numbers get kept, never what you said.
Settings backup
Save all your settings and words to one file and load it on another computer. It keeps your last 10 settings on its own too, in case a change goes wrong.
An indicator you can style
Orb, waveform, frequency bars or radial, any color. It shows up in the window you're talking into, or stays in one spot on the screen if you'd rather. It moves with your voice so you know it's hearing you.
Only the languages you speak
Whisper picks from 99 languages and on a short dictation it sometimes guesses wrong. Tick the ones you actually speak and it sticks to those.
Stays out of password boxes
If your cursor is in a password field, TuxTalk types nothing and says why. That works in most GTK apps, in KDE apps, Firefox, Chrome and apps built on Chrome with accessibility on, and in your desktop's own admin password prompt.
Help you can search
Type in the box at the top of Help, or just start typing, and you get every card and voice command that matches, with your own keys in them. Help's sorted by situation, so this is how you find a feature you only know by its name.
It tells you why
If a dictation can't be typed into a window, TuxTalk says why and what fixes it instead of just doing nothing.
Report a Problem
Help, Problems has a button that puts together what I'd ask you for anyway, your system and settings, the end of TuxTalk's log and the hotkey helper's own log. Your username's taken out and none of what you dictated is in it. You see all of it first, then you can copy it or save it, or open a GitHub issue with it already filled in. Nothing gets sent unless you do that yourself.
No tray? Still works
On a desktop with no system tray, like plain GNOME, opening TuxTalk from your apps brings up its menu in a small window.
A command for your own shortcut
tuxtalk --toggle, plus --start and --stop, if you'd rather use your desktop's shortcuts or a script.
A look around
Swipe or scroll sideways for more, and click one to see it full size.
Dictating a letter, with the indicator showing it's listeningThe tray menu, with Recommended, Lightest and Battery saverPreferences: model, graphics card and hotkeysDictation cleanup and text formattingThe Words tab: Your Words and SnippetsYour StatsMicrophone test and indicator styleUpdates and backupThe voice commands, all in one placeEvery other choice, with its download size
Watch it work
Just under a minute, with sound. It dictates a letter in LibreOffice Writer, then shows the tray menu and the first-run setup. The "no wait" fix in there is the AI corrections, which are off until you turn them on and run on your computer.
No sound handy? This 30-second version has captions instead.Almost two minutes, with sound: opening apps, snapping windows, the browser and the mouse grid, all by voice. It's in Labs, off until you turn it on.
I'm Ron, and I don't write code. Claude, Anthropic's AI, writes TuxTalk. I decide what it does, test it on real machines and use it for nearly everything I type. And I hold the code to a standard a human developer would respect: nothing left lying around, no copies of the same thing, tests for what you'd actually notice.
TuxTalk turns your speech into text right there on your computer. There's no account and no cloud, and once the speech model is downloaded it works with the internet unplugged.
The privacy page has the details, including the few times TuxTalk does go online and why.
Only TuxTalk's little hotkey helper reads your keyboard, and it only passes on TuxTalk's own hotkeys. Other programs don't get to see what you type.
What you dictate doesn't go in the log file. There's a setting to log it for troubleshooting, and it's off.
The AI corrections are optional and off until you turn them on. If you do, the AI runs on your computer too.
Your Stats keeps numbers only. Recent Dictations are held in memory and gone when you log out.
Get TuxTalk
Pick whichever one matches your Linux. It's the same app in all of them, for 64-bit PCs (x86_64).
The Download buttons take you to the newest release on GitHub. Grab the file that ends in .AppImage, .deb or .rpm. What changed in each version is in the changelog.
The first time it runs, TuxTalk asks for your admin password once. That sets up its hotkey helper, the one small piece that reads the keyboard, and lets TuxTalk type into other apps. No restart needed. On a system without systemd it uses the input group instead, tells you why, and then it does need a restart.
On another distro and rather not use a package? The README has steps to install it from source, in a folder of its own. tuxtalk-appimage is the quicker install on Arch, and the two can't be installed together.
The AppImage needs fusermount, which almost every desktop already has. If it won't start, install the fuse3 package. No Flatpak yet. If you'd like one, vote in the poll.
Works on your desktop
TuxTalk's hotkey helper reads the key straight from the keyboard instead of asking the desktop for it, so hold-to-talk works on Wayland and X11 with no setup. On GNOME you get a proper panel menu, and GNOME itself draws the recording indicator so it's sharp on a scaled screen; everywhere else a tray icon, and on a desktop with neither, a small menu window.
GNOME
KDE Plasma
COSMIC
Sway
Hyprland
Cinnamon
Other X11 desktops
Uninstalling
Open Preferences, go to Updates & Backup and click Uninstall TuxTalk… at the bottom. It takes out TuxTalk and everything it set up on your computer: the program, the hotkey helper and the typing permission, its own downloads (the graphics and AI engines and the CUDA libraries), starting at login, its menu entry and the GNOME extension, and it turns your desktop's accessibility off again if TuxTalk turned it on. It may ask for your password once. The speech and AI models sit in a download folder other apps can share, so they go only if you tick that box too. Your settings and word list stay unless you tick the box to delete them.
If TuxTalk won't open, the same thing runs from a terminal. For the AppImage it's ~/AppImages/TuxTalk.AppImage --uninstall, and for the .deb, .rpm and AUR:
tuxtalk --uninstall
Removing just the package with apt, dnf or pacman takes the hotkey helper with it, but leaves your settings and the downloads in your home folder. If you'd rather do it all by hand, or you have a version older than 0.17, the step-by-step commands are in the README.