Windows runtime: parakeet.cpp v0.6.1-fv3 (CPU, Vulkan, CUDA)

Speech engines for FluidVoice on Windows, built from altic-dev/parakeet.cpp branch fluidvoice at d9f7717: upstream parakeet.cpp v0.6.1 (+7 commits, 2eac2e7) plus our CUDA memory fix (#917 in altic-dev/FluidVoice).

• cpu: any x64 PC
• vulkan: AMD, Intel and NVIDIA graphics
• cuda: RTX 30/40/50 (native), other cards through PTX
• legacy cuda: GTX 900/10, Titan V, RTX 20 and similar, native code with no PTX

The CUDA zips need the cudart/cublas zip from mudler/parakeet.cpp v0.4.0 next to them. The app downloads everything it needs automatically.

Windows runtime: parakeet.cpp v0.4.0-fv2-legacy (CUDA, older GPUs)

Speech engine for FluidVoice on Windows, for older NVIDIA cards (GTX 900 / 10 series, Titan V, RTX 20 series and similar).

Same parakeet.cpp build as the current CUDA engine, but compiled with native code for these cards so it no longer depends on the driver compiling PTX at runtime. That was what crashed on drivers like 573.xx (#1235, #1281 in altic-dev/FluidVoice).

The app downloads this automatically when it needs it. You don't need to grab it yourself.

FluidVoice Linux 0.0.19

• Hyprland and Omarchy with blur turned on: the recording overlay no longer has a dark box around it.
• Faster speech recognition: a newer speech engine, the same one as Windows, transcribes a little faster with identical results. It downloads once, automatically.

FluidVoice Linux 0.0.18

• Hyprland and Omarchy: the recording overlay no longer takes focus at all, not even for a moment, so your words always paste where you were typing.
• Hyprland and every Omarchy version: one click on Add shortcuts sets up dictation, Fluid Intelligence and Paste last in your Hyprland config. On Omarchy they're Ctrl+Alt+Space, Ctrl+Alt+Shift+Space and Super+Shift+V. Keys you already bound to FluidVoice are kept, and keys Hyprland already uses are never taken.
• Omarchy 4 is now recognised. If an earlier version had you put dictation on Ctrl+Super+Space, which Omarchy also uses for backgrounds, Settings now tells you so.
• Fluid Intelligence can be turned on and tried during setup on Hyprland, Sway, niri and GNOME 46, where it used to be greyed out.

FluidVoice Linux 0.0.17

• Hyprland and Omarchy: the recording overlay opens on the monitor you're using and no longer takes focus, so your words paste into the window you were typing in.
• Hyprland: no black block above the overlay, and FluidVoice is no longer drawn at double size on 1x screens.
• Hyprland with a Lua config, such as Omarchy 4: the setup screen shows a line for bindings.lua, on Ctrl+Alt+Space on Omarchy.

FluidVoice Linux 0.0.16

• New recording overlay, like on Mac: Basic or Smart beside the waveform, and Copy Last and Insert Last under ….
• Pastes for you on Hyprland, Sway and niri, with nothing to install, and with Ctrl+Shift+V in terminals.
• Omarchy: one click makes FluidVoice Omarchy's dictation app, on Alt+Space or hold Right Alt.
• Ubuntu 24.04 and other GNOME before 48: one click adds the dictation shortcut to GNOME's keyboard settings.
• Hyprland, Sway and niri show the exact line to add to your config, and --start / --stop give push-to-talk.

FluidVoice Linux 0.0.15

• Shortcuts no longer stop working after a restart on KDE Plasma 6.7.4 and newer, where the app said they were allowed but the keys did nothing.
• Paste last transcription is on by default as Super+Shift+V, and works on Wayland. If you already allowed FluidVoice's shortcuts, press its row in Settings > Keyboard shortcuts once so your desktop can add it.
• Cancel dictation: on KDE, Esc cancels from any app, only while you're recording. Other desktops suggest Super+Shift+Esc when they ask.
• Where you dictate in Insights now names the app you dictated into. Older dictations never recorded one, so they aren't counted.
• Learn from your edits is on by default: correct a word after dictating and FluidVoice adds it to your dictionary. It uses your desktop's accessibility system, and works in apps that share their text with it, such as KWrite and Firefox. Apps opened before FluidVoice may need reopening. Turn it off in Custom Dictionary.
• Dictation goes straight into the app you're typing in, and lands on the clipboard if it can't.
• Speech and AI use your graphics card automatically, and fall back to the processor.
• Settings are saved more safely, so a crash or power cut can't reset them.
• Settings shows the key your desktop actually has, including Fluid Intelligence's.
• New commands for desktops without shortcut support: --paste-last and --cancel, alongside --toggle.
• Mute sound while dictating, off by default, in Settings > Voice & capture. Experimental, and needs pactl.
• The overlay briefly says "Loading…" when the speech model takes a moment to load.

FluidVoice Linux 0.0.14

FluidVoice for Linux, early beta.

Download the .AppImage, make it executable and open it. Later versions install from Settings > General.

• Fixed: the main window going white when dictation starts
• The recording overlay now always floats on top, with Finish and Cancel buttons
• Closing the window keeps FluidVoice running in the tray where your desktop shows one
• Folder buttons and Open Logs now open your file manager
• Clearer shortcut status in Settings, plus on-screen confirmation for copy and export

Tested on KDE Plasma. Other desktops and real graphics cards are still being tested.

Linux runtime: parakeet.cpp v0.6.1-fv3 (CPU, Vulkan)

Speech engines for FluidVoice on Linux x64, built from altic-dev/parakeet.cpp branch fluidvoice at d9f7717 on Ubuntu 22.04: upstream parakeet.cpp v0.6.1 (+7 commits) plus our persistent-arena fix. Same source as windows-runtime-parakeet-v0.6.1-fv3.

• cpu: any x64 PC
• vulkan: AMD, Intel and NVIDIA graphics

Needs glibc 2.34 or newer. The app downloads what it needs automatically.

FluidVoice Linux update feed

The address the Linux app checks for updates. Download the app from the newest linux-v release instead.