Linux has been the forgotten platform for dictation for two decades. Here is where things actually stand in 2026, and how to get modern, on-device, Whisper-powered voice typing working on any Linux desktop.
If you've tried to set up dictation on Linux in the past, you know the story. The desktop ecosystem has always trailed macOS and Windows on accessibility tooling, and voice input has been the most glaring gap. The options have historically been: wrestle with half-abandoned open-source projects, run a cloud service that sends every word to someone else's server, or just give up and type.
In 2026, that's finally changing. Whisper-class models are small enough to run on consumer laptops, whisper.cpp has matured into a production-ready inference engine, and GPU acceleration is available on basically every modern Linux machine via Vulkan and CUDA. The pieces all exist, they just hadn't been packaged into a real desktop app. Now they have.
What's been missing is a packaged, installable desktop app that runs Whisper locally with GPU acceleration, has a proper UI, and Just Works on a modern Linux install. That's the gap Wspr is built to fill.
Tools like nerd-dictation, wtype scripts around whisper.cpp, or Vosk-based tools. These can work well for a tinkerer, but they're CLI-first, require manual setup per desktop environment, and don't give you a packaged app with a UI.
Like the OpenAI Whisper API or Google Speech-to-Text, wrapped in custom scripts. Accurate, but your audio leaves your machine and you pay per minute.
Useful for some workflows but typically limited in accuracy, languages, and customization.
Commercial dictation vendors (Dragon, Wispr Flow, Superwhisper, MacWhisper) build for Mac, Windows, and in some cases mobile, because that's where the paying users are. Linux is a rounding error on their revenue spreadsheet, so it stays unsupported.
GNOME, KDE Plasma, Hyprland, Sway, XFCE, Cinnamon: same kernel, six different ways to paste text into the focused window. Any serious Linux dictation tool has to handle X11 and Wayland, ibus and fcitx, xdotool and wtype. It's non-trivial and most vendors don't bother.
ALSA, PulseAudio, PipeWire: capturing microphone input on Linux has been a moving target for years. PipeWire finally stabilized that in 2023 to 2024, which is part of why this kind of app is feasible now.
Wspr is a native Linux app. It runs the same whisper.cpp engine as on macOS and Windows, with GPU acceleration where available and CPU fallback when it isn't.
Press your configured global hotkey.
Say it the way you would say it out loud.
Release the hotkey. Wspr transcribes locally and pastes into the focused window.
Optionally trigger an AI rewrite: "make this a bullet list", "translate to French", "fix grammar". It is routed to OpenAI, Anthropic, or Gemini using your own API key.
It also does file transcription: drop an audio or video file in, get a transcript out. Good for meeting recordings, voice memos, and podcast clips.
Watch Wspr type in five different places.
Wspr ships as a Linux build for x86_64 systems. It's been validated on mainstream distros including Ubuntu and its derivatives. Other recent glibc-based distros should work too, though your mileage may vary depending on your desktop environment's input-injection behavior.
Fully supported and tested daily.
Since version 5.1.2 Wspr injects text through libei via xdg-desktop-portal and holds one portal session for the life of the process, so the compositor remembers the grant and the old “Control input devices” prompt on every launch is gone.
The global dictation hotkey goes through the GlobalShortcuts portal on Wayland, which is why it still fires when no Wspr window is focused. Your desktop owns that binding, so you rebind it in your desktop's own settings (on GNOME, under Settings then Apps).
Compositors without that portal, meaning wlroots ones such as Sway and Hyprland, fall back to xdotool, which needs XWayland present.
A bare Wayland session with neither the portal nor XWayland is not supported yet.
Audio capture goes through cpal, which speaks whatever audio system your distro exposes: PipeWire on modern setups, PulseAudio on older ones. If you run into a compatibility issue on a specific distro or compositor, open a GitHub issue and we'll look at it.
If you dictate into a terminal, the paste path is worth understanding in a bit more depth: we wrote it up in How to Dictate Prompts to Coding Agents on Linux.
A big reason people use Linux in the first place is because they don't want a vendor-controlled pipe between them and their computer. A dictation app that streams every word to a cloud service undoes that in one install.
Wspr never sends audio anywhere. The Whisper model lives on your disk, runs on your GPU, and produces text that's pasted directly into your focused application. No telemetry, no account, no cloud dependency for core transcription. AI rewriting is the only step that optionally calls out to a provider, and only the text you explicitly rewrite, using your own API key, routed directly from your machine to the provider you chose.
Whisper "small" and "medium" models run comfortably in real-time on a mid-range Linux laptop with integrated graphics. On a discrete GPU, you can run "large-v3" and get near-instant transcription of multi-minute recordings. Wspr downloads models on first run; you pick the one that fits your hardware.
Linux has needed a real dictation app for a long time. Wspr is that app. It's private, it runs Whisper locally with GPU acceleration where available, and it costs $14.99 once.
No subscription, no account, no tiered Pro that gates normal functionality. The free download includes 50 live transcriptions so you can actually test that it works on your exact setup before paying. Pro is a one-time unlock that removes the cap, and removes it forever.
X11 sessions are fully supported and tested daily. Wayland sessions that expose the RemoteDesktop portal (KDE Plasma 6, GNOME 45+) are supported natively, through libei via xdg-desktop-portal. Compositors without that portal, meaning wlroots ones such as Sway and Hyprland, fall back to xdotool, which needs XWayland present. A bare Wayland session with neither the portal nor XWayland is not supported yet.
Wspr ships as a Linux build for x86_64 systems. It has been validated on mainstream distros including Ubuntu and its derivatives. Other recent glibc-based distros should work too, though your mileage may vary depending on your desktop environment's input-injection behavior.
No. Wspr never sends audio anywhere. The Whisper model lives on your disk, runs on your GPU, and produces text that is pasted directly into your focused application. There is no telemetry, no account and no cloud dependency for core transcription.
Yes, where GPU acceleration is available, with a CPU fallback when it isn't. Wspr runs the same whisper.cpp engine as on macOS and Windows. Models are downloaded on first run and you pick the one that fits your hardware.
Audio capture goes through cpal, which speaks whatever audio system your distro exposes: PipeWire on modern setups, PulseAudio on older ones. If you run into a compatibility issue on a specific distro or compositor, open a GitHub issue.
$14.99, once. There is no subscription and no account. The free download includes 50 live transcriptions so you can test it on your exact setup before paying. Pro is a one-time unlock that removes the cap forever.