Fuck You Flow
Free, open-source Wispr Flow alternative for Windows 10 and 11 and Ubuntu Linux 22.04
or newer. English, Greek and 97 more languages, detected automatically, or locked to your own language in Settings.
Press Right Alt, speak, then press it again to insert text into a compatible
Windows text field. The default Whisper engine processes speech on your own PC.
A cleanup step, on by default, removes hesitation sounds such as “um” and “uh”
and applies spoken self-corrections before the text lands.
Local dictation works offline after setup, with no account, subscription or word
cap. An optional remote speech provider sends audio to the provider you configure;
its fees and privacy terms apply. The app also checks online for updates, and
destination apps can store or upload text you insert into them.
Made by Luram AI Agency in Thessaloniki, Greece. Beta, under MIT.
The latest release
and the changelog list what changed.
Why it exists
We wanted a free Windows workflow for local dictation, starting with Greek and English, with a
hotkey, editable dictionary and no recurring charge for the local engine. FU Flow
builds that workflow around openly released Whisper models and whisper.cpp.
Other dictation products offer different platforms, models and features. FU Flow
is an independent project; the sourced comparisons above explain the tradeoffs
without claiming to replace every competitor feature.
What you get
- Right Alt starts, Right Alt stops. Press, speak, press again. The text is
inserted through the normal Windows paste workflow. Escape cancels. Ctrl+Win
remains available for hold-to-talk, and shortcuts can be changed in Settings.
- Transcription while you speak. Finished phrases are processed during
recording to reduce the work left when you stop. The remaining wait depends on
the model, speech, hardware, drivers and available memory.
- 99 languages. Whisper detects the language on its own, or you pick yours from
thirty-three in Settings so it never drifts into another one. Greek and English
are the most tested, including English words inside Greek sentences. Review
names and punctuation; recognition can make mistakes.
- Help me choose. Two questions (your language, what matters most) plus your
graphics card memory pick one model, marked Best for you. The same rule is
written out in the model guide.
- A paste that checks it arrived. When nothing takes the words within a second,
Ctrl+V is pressed once more; if that fails too, the words stay on the clipboard.
In Chrome, Edge, Claude, Slack and other programs built on Chromium the words
always stay on the clipboard after pasting, one Ctrl+V away.
- Two looks for the dark theme. Carbon, the default, adds depth with still
gradients and shadows; Classic is the original flat black. Neither adds work for
the graphics card. A light theme is also available.
- Rule-based cleanup. Rules can remove filler words, apply spoken corrections
such as "Τρίτη, όχι όχι, Παρασκευή", and add question marks from the wording.
They can misinterpret speech, so review the result before sharing it.
- A dictionary you control. Save recurring corrections as editable rules.
When learning from edits is enabled, suitable History edits produce suggestions
you can accept or dismiss. This does not retrain Whisper or guarantee that a
word will always be recognized correctly.
- Snippets. Say a phrase, get a block of text.
- History, statistics and per-application styles, all on your disk.
- Local speech processing by default. The local adapter talks to an engine
process on the same PC. Selecting the optional remote provider uploads the audio
to that provider. The app attempts to reject detected password fields; some
sensitive fields may still go unrecognised.
Some applications, protected documents or elevated windows can refuse automatic
paste. Copy the transcript from History and paste it manually if needed. Local
history and clipboard contents can be read by someone with access to the PC;
local recognition does not make a cloud document or conversation offline.
Install
Windows
Download the installer from
Releases and run it. The
full installer is about 1.7 GB to download, including the speech models.
That is the download size. Allow additional
space for installation, local models, history and updates. The installer uses
per-user installation mode.
The beta installer is unsigned, so Windows may show a "Windows protected your PC"
dialog. Use the linked project release and compare its published SHA256 checksum
before deciding whether to continue through "More info" and "Run anyway".
The full installer includes the Vulkan speech engine, large-v3 and turbo models,
and the voice activity detector. Supported AMD, Intel and NVIDIA cards can use
Vulkan; a CPU fallback is available. Compatibility and speed vary by device and
driver. Initial setup examines the machine to select a model and thread count;
review the selection in the Speech models tab in Settings.
The app checks for available updates after startup. Downloading an update waits
for your choice. Local dictation does not require a successful update check.
Ubuntu Linux
Ubuntu 22.04 or newer, 64-bit. From the
latest release
download Fuck.You.Flow.Setup.<version>_amd64.deb and install it with
sudo apt install ./Fuck.You.Flow.Setup.<version>_amd64.deb, or download the
AppImage, make it executable and run it. Both include the speech models, about
1.6 GB. The .deb sets up the keyboard permission the Right Alt shortcut needs;
with the AppImage the app shows the one command to run. Dictation works under
X11 and Wayland. Updates arrive through the app, and the .deb asks for your
password to install them.
Speed
Four recorded maker observations on an RTX 3070 with large-v3-q5_0, measuring
the wait after finishing dictation until text appeared:
| Words | Speech | Wait |
|---|
| 32 | 17.5 s | 0.59 s |
| 108 | 82 s | 0.17 s |
| 179 | 77 s | 0.53 s |
| 312 | 152 s | 0.88 s |
These are the maker's own measurements on one PC. They make no claim about Wispr
Flow, OpenWhispr or any other product, and results on other PCs will differ.
The last phrase is one source of remaining work; total recording length, model
state, memory pressure and other running applications can also affect the wait.
See the model benchmark summary for separate recognition
accuracy and processing-time measurements on the project's test corpus.
An earlier engine comparison on that machine used one 5-second Greek phrase with
large-v3-q5_0: Vulkan 593 ms, CUDA 663 ms, processor alone 12.5 s. Those figures
describe that setup. Shipping Vulkan support does not establish equal speed on
AMD, Intel and NVIDIA cards. A smaller model or CPU fallback has its own tradeoffs;
test a short dictation on your machine before relying on it.
Speaking well
Speak at a comfortable pace and check the transcript. Microphone quality, noise,
pronunciation and model choice affect recognition. Practical things to try:
- A short pause between sentences can help the engine separate phrases. It does
not guarantee punctuation or a particular processing time.
- Pronounce the last word of a sentence fully, especially Greek word endings.
- Names and foreign words: say them clearly and in one piece.
- Check question marks after dictation. You can try saying "ερωτηματικό" as an
explicit punctuation command, then review the result.
- With self-correction cleanup enabled, try "όχι όχι" followed by the right word.
Check that the intended correction was applied.
Build it yourself
Requirements: Rust stable, Node 20 or newer, pnpm, the Visual Studio C++ build
tools. See docs/DEV-SETUP.md.
pnpm install
pnpm tauri dev
Always build through the Tauri CLI. A plain cargo build --release produces a
development binary that looks for the interface on a dev server.
Documentation
Contributing
Issues and pull requests are welcome. The parts that most need other people's
machines: applications that refuse a paste, keyboard layouts, older processors,
and Greek words the dictionary should know.
License
MIT. See LICENSE. The speech engine and the models carry their own
licenses, listed in the third-party notices.
FU Flow is independent of Wispr Flow and OpenWhispr. Third-party components retain
their own licenses and attribution requirements.
