Runs better on computers with a small or older graphics card, ships a newer speech engine, and keeps the pauses you speak.
New
- Computers with a small graphics card, or none, get a lighter speech model that runs on the processor and leaves one core free for your other programs. If you already use a heavier model there, the first screen offers the switch and changes nothing until you agree.
- When the graphics card fails to start, dictation carries on with the processor. On a computer where speech never started, the app fetches the lighter model by itself and tells you.
- A line on the first screen tells you when the ready-made dictionary updated itself.
Improved
- Newer speech engine (whisper.cpp 1.9.5), with fixes for AMD, Intel and NVIDIA cards.
- When the graphics card is busy, long dictations go to the engine in fewer pieces and the text arrives sooner.
- Faster paste in programs that read the clipboard on their own, such as remote desktop and some chat apps.
- Model downloads are steadier: a stalled download stops after a minute and leaves no half file behind.
- If you keep recordings, they are deleted after 30 days. The text in History stays.
Fixed
- A pause you speak stays as "..." where the engine heard it, and the next word keeps its small letter.
- A rising voice adds a question mark only to a short sentence or to one with a question word.
- Quitting while a dictation is being written waits for it to reach the cursor and History.
- Smaller fixes: a question mark at the end of an English sentence stays, a web address keeps its small letters at the start of a sentence, History names the engine of every dictation, and two-core computers use both cores.
Known issues
- This version is for Windows. Ubuntu Linux stays on 0.9.12 for now.
- The first dictation after the update can take a few seconds longer while the new engine prepares your graphics card.
Ready-made dictionaries for your language, an Edit button for suggestions, and more accurate punctuation.
New
- Ready-made dictionaries. The app can download a hand-checked list of words the speech engine often mishears in your language. Each list is signed, updates itself once a day, and any rule you change or delete stays the way you left it. It is offered on the Home page and under Dictionary, Look for my language.
- Dictionary sharing. The corrections in your Dictionary help build the ready-made dictionaries. What you dictate, your History and your voice stay on your computer. Sharing is on from the start. Switch it off, or delete what was sent, in Settings, Privacy.
- A new install asks which language you speak, so dictation starts in the right language.
- Suggestions have an Edit button. Correct a suggestion that is nearly right, then accept it.
- A card in the menu invites you to rate the project on GitHub. It goes away once you press it.
Improved
- The names you say most often reach the speech engine first, which helps it recognize them.
- Edit in History opens the whole transcript at once and grows as you type.
- Punctuation: the word after e.g. or i.e. keeps its small letter, and a domain ending said on its own, as in "on .ai", keeps its space.
- English filler sounds such as "um" are left alone in languages where they are real words.
Fixed
- When a paste is lost, History now shows it as failed.
- Smaller fixes: Suggestions keep ordinary words out of the name list, one name keeps one spelling, and Delete all data says what stays as it is.
Known issues
- This version is for Windows. Ubuntu Linux stays on 0.9.12 for now.
Send a problem or an idea from inside the app. Also cleaner punctuation, better short endings, and one paste per dictation in Chromium and Electron apps.
New
- Problem or idea? A new line in the menu opens a short form that reaches the maker. You can add a reply address and the diagnostic log, which you can read before sending. The log holds app events and errors. Your dictated words are never in it.
- The menu now shows who makes the app, Luram AI Agency, with a link to its site.
Improved
- Short last words after a pause come out right: the engine hears them together with the words before them.
- Cleaner punctuation. A pause inside a sentence stays a pause, a sentence split between two pieces joins without a stray stop, and a filler word at the start of a sentence leaves without its comma.
- A learned rule that rewrites a word you keep as it is elsewhere is switched off.
Fixed
- In Chrome, Claude, Slack and other Chromium or Electron apps, each dictation is pasted exactly once. If a window does not take it, the text waits on the clipboard.
- Smaller fixes: the Home "today" row shows today's time saved, the menu fits short windows, and the memory panel hides on narrow windows.
Known issues
- This version is for Windows. Ubuntu Linux stays on 0.9.12 for now.
Now on Ubuntu Linux. Also smarter learning from your edits, Dictionary rules that know their language, and the speech engine open to your own programs.
New
- Fuck You Flow runs on Ubuntu Linux 22.04 or newer, as a .deb package or an AppImage, under X11 and Wayland. It updates itself like the Windows version.
- A local endpoint for your own programs, in Settings, Privacy. Scripts and bots on this computer can send audio and get the text back from the speech engine. It stays off until you turn it on and answers only programs on this computer.
Improved
- Learning from your edits now works on long transcripts too. A correction is learned when you make it twice or when it is a name or a technical term, and a word you leave as it is elsewhere never becomes a rule. Your earlier edits are read again once, so new suggestions may be waiting under Suggestions.
- Each Dictionary rule belongs to the language of its letters and applies to text in that language. With two languages chosen, the rules of both work together, English words inside the other language included. The form picks the language as you type.
Known issues
- Linux: the AppImage needs a one-time keyboard permission, and the app shows the command to run. The .deb sets it up during install.
- Linux under Wayland: per-app styles are skipped, because the app cannot see which program is in front. Dictation works as usual.
A new look called Carbon, and your words stay on the clipboard after pasting into browsers and the apps built on them.
New
- Carbon, the new default look of the dark theme, with panels that have depth. It adds no load on the graphics card. The classic flat look is in Settings, General.
Fixed
- In Chrome, Edge and apps built on them, such as Claude and Slack, the words stay on the clipboard after pasting. If they did not arrive, Ctrl+V brings them back.
Pasting that checks the words arrived, a guide that picks the right model, and steadier results.
New
- Help me choose: two questions and your graphics card pick the model that suits you. One click downloads it and switches to it.
- Every model and engine explains in plain words who it is for and what it asks of your computer.
- The bar warns within six seconds when the microphone sends no sound.
Fixed
- Pasting checks that the program took the words and tries once more when it did not. If that fails too, the words wait on the clipboard.
- Words could go missing in browsers and Electron apps that moved the focus to their menu just before the paste.
- The same recording now gives the same text every time.
- Learned corrections match only the exact word you fixed and leave web addresses alone. An edit that fixes two words teaches both.
- Subtitle credits that the engine sometimes added after speech are removed.
Known issues
- In the Claude desktop app a dictation occasionally does not arrive. The words stay on the clipboard for Ctrl+V.
Better recovery when recognition drifts out of your languages, plus fixes to startup, shortcuts and pasting.
Improved
- Imported recordings are always left out of the saved-audio cleanup.
Fixed
- When recognition returns a language you did not choose, the recording is kept so you can try again. The same goes for imported audio.
- Startup shows a retry button when the settings take too long to load.
- A new cloud provider address or model takes effect at once.
- Smaller fixes: the stop key, dragging the bar, shortcut fields, dollar signs in snippets, sentence punctuation, the newest log in Diagnostics and the screen reader language.
Known issues
- A blank white window sometimes appears above the listening indicator.
- Text in a third language written in Latin letters can still slip through.
The language lock now checks every part of a dictation.
Fixed
- The language lock from 0.9.7 now works. Each part of a dictation is checked and heard again in your language when it strays, and one word in another alphabet is enough to catch it.
Known issues
- When the last part of a dictation has to be heard again, the text arrives up to a second and a half later.
- A single English word said on its own can come out in the letters of your language.
Your dictation language stays yours, thirty-three languages to choose from, a bar you can place anywhere, and a crash fix.
New
- Choose your language in Settings, Language, from thirty-three. The modes are your language with English, your language only, English only, or auto-detect. A new install starts with the language of Windows.
- Drag the bar to any edge of the screen and it stays there. Settings, Overlay can show it, so you can move it without dictating.
Improved
- The discreet style is a single dot: red while listening, green when done, amber when the words wait on the clipboard.
Fixed
- In the mixed mode a short phrase could come out in an unrelated language. Anything outside your two languages is now heard again in yours.
- The app could close during the first hour after Windows started.
- The Windows clipboard history (Win+V) could leave a paste empty.
Updates that announce themselves, a faster paste, mouse button shortcuts, a steadier hotkey and many fixes. It includes the unreleased 0.9.4 and 0.9.5.
New
- The app checks for a new version shortly after it opens and shows an Update now card, waiting until you finish if you are dictating. Check for updates is also in the tray menu.
- The side buttons of a mouse (Mouse4, Mouse5) can be shortcuts, alone or with Ctrl, Shift and Alt.
- Diagnostics has a Key check that shows live what your hotkey sends.
Improved
- Pasting is faster and more reliable: the words are on the clipboard before the key press.
- Dictated words stay in the Windows clipboard history (Win+V) and are kept out of the cloud clipboard.
- The hotkey recovers within a second when a busy Windows drops it.
- The program, its folders and its files carry the product name. Your settings, history and models move over by themselves.
Fixed
- On keyboards where the right Alt types accented letters and symbols, typing them no longer starts a dictation.
- Deleting a History entry that came from an audio file keeps your original file.
- Pictures, files and formatted text on the clipboard come through a dictation unchanged.
- If the last part of a long dictation fails, the rest is kept.
- Installing over an older version closes the old one first, and your settings survive a file that cannot be read for a moment. Delete all data also removes the backups.
- Smaller fixes in History (corrections, Paste, Copy), in Settings (shortcuts, the Save button, the startup switch, retention) and in the note window, which now follows your language and wraps its text.
Two small fixes to updating.
Fixed
- The new version bar repeated a word.
- An automatic check could interrupt an update that was already downloading.
Updates from inside the app, recordings turned into text, and a window that keeps your words when a paste cannot land.
New
- The app checks for a newer version at each start. Nothing downloads until you press the button.
- Turn a recording you already have into text: Home, Sound file to text, up to 90 minutes.
- When a dictation cannot reach the program you were in, a small window shows the words. You can switch it off in Settings, Insertion.
- A graphics card left switched off is turned on when it can be used.
Improved
- Speech models move to your data folder on the first run, which keeps updates near 70 MB. They stay when you uninstall.
- Crashes are written to the logs, so they can be looked into.
Fixed
- A rare crash while reading the clipboard.
- Smaller fixes: the note window's buttons and theme, the Windows startup entry, and the Settings page keeping what you type.
Clear reporting when a paste goes nowhere, and an interface that follows your language everywhere.
New
- A new install picks up the language of Windows. Both the interface and the dictation language can be changed in Settings, Language.
- Settings, Speech models shows which device the engine runs on.
- An event log keeps seven days of what happened and what failed. It never holds the text you dictate.
Fixed
- A dictation aimed at the desktop or the taskbar is reported: the text stays on the clipboard and the app says which window was in front.
- Model descriptions follow the interface language.
The first public release: one key, two languages in one sentence, everything on your own computer.
New
- Hold the right Alt, speak, press it again. The text appears where your cursor is, in any program, and two languages can share one sentence.
- The installer includes everything, with graphics acceleration on AMD, Intel and NVIDIA cards. Nothing downloads later.
- The first start looks at your computer and picks the model and settings that fit.
Back to the download
Fuck You Flow