FLUSS GitHub

00 Local dictation for Windows

Hold a key.
Speak.
It’s written.

Fluss turns your voice into text wherever your cursor is — on your own machine, in tens of milliseconds. Nothing leaves your PC.

Download for Windows Source on GitHub

Windows 10/11 · free · MIT · 4 MB installer

Sourcesimulated
Level−∞ dB
Decode21 ms

01 The flow

Three motions. No window to switch to.

  1. 1Hold CtrlWinanywhere, in any app
  2. 2Speaka live preview follows you
  3. 3Releasethe text is typed at your cursor

02 Continuation

A pause is not a full stop.

Every speech model guesses a period when you stop talking. When you keep going — after a breath in hands-free mode, or with a second press into the same text field — Fluss re-reads the last seconds of your previous audio together with the new audio, so the model itself decides how the two connect.

Only the punctuation at the seam is changed. Words already typed are never rewritten, and nothing is edited if you clicked or typed in between.

First press

…and the tests look good.

Second press

and then we deploy on Friday.

Typed

…and the tests look good, and then we deploy on Friday.

joint decode · 3 s of context · −1 char, +“,”

03 One-second corrections

Teach it a word without leaving the sentence.

I moved my laptop to on March 9chunk · 3 words · 0.41 s last week.

Heardon March 9
Should beOmarchy
Enter saveCtrl←→ adjustTab another spelling

I moved my laptop to Omarchy last week.

Right after a word comes out wrong, press CtrlWinR. The popup already holds what you just said — not the last word, the last spoken chunk, found from timing, pauses, audio energy and confidence. Your cursor waits in “Should be”.

Type the spelling, press Enter. The rule is saved, the dictation you just made is corrected in place, and the word is passed to the model as a hotword from then on.

04 Pre-roll and post-roll

It hears the word you started too early.

A short ring buffer keeps the last half second before you press, and listening continues briefly after you let go. Both lengths are adjustable. If you had already stopped speaking when you released, the text is typed immediately.

05 Models

Pick the engine. Keep it warm or let it sleep.

ModelDownloadRuns onLanguagesPer phrase
Phonon 2.5 English default158 MBCPUEnglish~20 ms short · ~115 ms for 10 s
Phonon 2.5 Multilingual170 MBCPU25 EuropeanCPU, as above
Parakeet TDT 0.6B v32.5 GBGPU (DirectML) or CPU25 European~40–55 ms on GPU
Custom endpoint—anywheredependsany OpenAI-compatible server

Keep the model loaded, unload it after idle minutes (20 by default), or load it only while dictating. If it is asleep when you press, recording starts anyway and the text arrives once it has loaded.

06 Local by design

Small, quiet, yours.

Audio sent anywhere
0 bytes

Recognition runs on your CPU or GPU. A cloud endpoint is used only if you configure one.

Background app
~38 MB

A Rust process with a native overlay. The settings window only runs while it is open.

Installer
4 MB

Models download during setup, into a folder you can delete. No account, no telemetry.

07 Command line

Everything is scriptable.

# what Fluss is doing right now
fluss status --watch

# change any setting
fluss config set audio.preroll_ms 300

# teach it a word
fluss rules add Omarchy --heard "on march 9"

# switch engine, install, transcribe a file
fluss model use parakeet-v3 && fluss model install
fluss transcribe meeting.wav

# diagnose
fluss doctor
  1. 1
    Download the installer

    From the latest release. No admin rights needed.

  2. 2
    Pick a microphone and a model

    Setup downloads the model and keeps it on your machine.

  3. 3
    Hold CtrlWin and speak

    Fluss starts with Windows and lives in the tray.

08 Credits

Built on good work.