Fluentry
Speech that stays on your machine
Whisper and Parakeet run on your own hardware — no account, no upload, no per-minute bill. A short phrase comes back in about 300 ms on a plain CPU, in dozens of languages.
Version 1.0.1 is out
Say it once.
Say it once.
Fluentry types it.
You think faster than you type. Hold a key, say what you mean, and the words appear in whatever app you are already in — your editor, your browser, your chat window. The speech model runs on your own machine, so it is quick and nothing you say leaves the computer.
You say
fix the um parser bug literal comma then ship it
Fluentry types
fix the parser bug, then ship it
What it does
Built on the parts of Linux you already run.
Your microphone through PipeWire, your shortcuts read straight from the keyboard, your API keys in your keyring — and speech models that never phone home.
Teach it your words
Colleagues’ names, project code names, the jargon your field runs on. Tell Fluentry once what it keeps mishearing and it stops. Leave the replacement blank to cut a word entirely.
Say the punctuation
“literal comma” becomes a comma. Switch on filler-word removal and the “um”s and “uh”s go with it, and a closing phrase can press Return, so a whole message goes out without your hands leaving the desk.
Polish, if you want it
Send a transcript through a language model to tidy the grammar — running locally under Ollama or LM Studio, or wherever you like. Off unless you turn it on, and you can switch it off for the apps where raw text is what you want.
Looks like the rest of your desktop
Picks up your colour scheme, accent colour, icon theme and UI font, so it sits alongside your other windows instead of announcing itself. Tray icon, recording overlay, and settings laid out the Adwaita way.
In your language
The interface itself speaks eleven languages — English, Spanish, French, German, Portuguese, Italian, Japanese, Korean, Chinese, Hindi and Arabic — and follows your system locale on first run, with a picker in Settings if you want another. Arabic lays the whole window out right-to-left.
Works from the shell too
Transcribe a recording with one command, or switch on a loopback HTTP API and let your own tools read the history, manage the dictionary and ask for transcripts.
Screenshots
Six screens and a pill that floats.
Adwaita’s own greys, so it sits next to GNOME Settings without announcing itself. These are rendered straight from the running app in both the light and the dark it ships with — you are seeing whichever one your system is set to.
While you talk, this is all that appears.
A pill at the edge of the screen with a level meter, and the words so far if you have asked for them. It never takes focus, so the cursor stays exactly where you left it and a screen reader stays on the document you are actually writing.
Install
Installed and dictating in a minute.
On Debian and Ubuntu the package brings everything with it, including the piece that lets Wayland remember your permission. The first launch walks you through picking a language, downloading an engine and trying a dictation.
$ sudo apt install ./fluentry_1.0.1_all.deb $ fluentry
Not on Debian or Ubuntu? pip install 'fluentry[gui,audio,input,whisper,parakeet,libei]' — but a virtualenv cannot see the system python3-gi, so create it with --system-site-packages or the permission dialog returns on every launch.
guitray, overlay, windowsaudiolow-latency captureinputglobal hotkeyswhisperthe Whisper enginesparakeetthe Parakeet enginesonnxNemotron and Coherelibeityping on Wayland
Engines
Pick the model that fits your machine.
Models download the first time you use them and live in your cache directory, so uninstalling takes them with it. Switch whenever you like.
| Engine | Runtime | Notes |
|---|---|---|
| Whisper Tiny … Large | faster-whisper or whisper.cpp | 99 languages; the runtime is fetched when you pick one, or an existing whisper.cpp binary is used |
| Parakeet TDT v3 / v2 | onnx-asr | ~640 MB int8, multilingual, ~300 ms for a short phrase |
| Nemotron, Cohere Transcribe | sherpa-onnx | their exports use that runtime’s layout |
Compatibility
Wayland, honestly.
Wayland withholds things from every application, Fluentry included. Typing into other apps is solved — through the input path the compositor sanctions, which asks your permission once. Here is what is left, and fluentry --check tells you which applies to your session before you hit it.
| Capability | X11 | Wayland | What it needs there |
|---|---|---|---|
| Typing into other apps | Works | Works | Through libei and the RemoteDesktop portal, which asks once and remembers. Needs python-libei and your distribution’s libei and python3-gi |
| Global hotkeys | Works | Needs a group | sudo usermod -aG input "$USER", then log back in |
| Knowing the focused app (per-app rules, and Ctrl+Shift+V in terminals) |
Works | Works | Hyprland, Sway and KWin expose it directly. GNOME hides it, so the package ships a small Shell extension — enable Fluentry Focus, then log out and back in |
Privacy
Where your words go: nowhere, unless you send them.
- Transcription needs no network. The models run locally; dictation works with the machine offline.
- History stays local, and can expire. A SQLite database on your disk that clears itself after a day, 7, 30 or 90 days if you ask it to.
- Keys live in your keyring. The freedesktop Secret Service, never a plain-text file.
- AI cleanup fails closed. Off by default. A provider is only contacted after you verify it, and what is stored is a hash of the endpoint-and-key pair — change either and the verification lapses.
- The local API is loopback-only and switched off until you turn it on.