Linux · Windows · macOS · MIT
Dictation, at your application.
Press your shortcut key, talk, and your words land back in the app where you started, even if you clicked elsewhere mid-sentence.
- #just wisper
# how it works
From voice to text in one breath
A short pipeline with no round trip to a server, the same on Linux, Windows, and macOS. You speak, a local model listens, your words land.
-
Speak
hotkey · push-to-talk / toggle
Hold the global hotkey (Super on Linux, Win on Windows, Cmd on macOS) for push-to-talk, or tap it to toggle. Your mic records locally, and the overlay pill shows a live waveform plus the icon of the app you started in.
-
Transcribe
onnx · on-device · stays loaded
Local ONNX models turn speech into text on your machine: Whisper, Parakeet, IndicConformer, Moonshine, SenseVoice, or Qwen3, with auto-detect or your own language priority. Models stay loaded so your next dictation starts instantly. Cloud keys (OpenAI, Groq, Sarvam) work too, strictly opt-in.
-
Insert
paste-back · origin app
The text lands back in the app where you pressed the hotkey, optionally polished by AI first. The exact paste method depends on your system, see setup.
# features
Everything you need to stop typing
Wisper does one job and stays out of your way. These are the controls you actually get.
-
presence
The overlay that follows you
A small pill floats on whichever monitor your cursor is on, pinned to the top or bottom, on Linux, Windows, and macOS alike. It shows the app you started dictating in, a voice-reactive waveform, a rotating radiant beam, and one button to cancel.
-
feedback
Hear it work
Subtle sounds mark recording start and text ready. Master toggle plus per-event controls. Silence it whenever you want.
-
control
Push-to-talk
Hold to record, release to transcribe. Prefer toggle? One tap. Any key or key combo works, and the Super key reads Win on Windows and Cmd on macOS.
-
privacy
Nothing leaves
Transcription runs on your machine. Cloud is strictly opt-in, and keys are saved on this device only.
-
refine
AI cleanup, your voice
An optional pass reformats and polishes the transcript. It never answers for you. Six writing styles plus custom instructions. Runs on OpenAI, Anthropic, Groq, OpenRouter, Ollama, OpenCode Go, or your own endpoint.
-
vocabulary
Your words, your way
Custom vocabulary turns shortcuts into proper terms. Say "gpt", get "GPT". Four profiles with import/export, URL sync, and per-profile toggle, plus scanning past dictations for new words worth saving.
-
input
Pick your microphone
Choose a specific input device or stay on system default. Virtual sinks are filtered out, and you test the level before you rely on it. Reduce background noise with a strength slider, and trim silence before transcription.
-
memory
Look back
Searchable history: replay the recording, re-transcribe, edit, copy. Collapsible Today shows today dictations/words/time saved. Set retention limits and choose whether recordings are kept.
-
engines
Local or cloud
Stay fully offline, or plug in a cloud speech key and switch any time. Several keys per provider rotate past rate limits automatically.
-
paste-back
Back where you started
Wisper remembers the window where you pressed the hotkey and pastes there, even if you click elsewhere mid-sentence. Native on Windows and macOS; on GNOME Wayland a tiny bundled helper does it.
-
languages
Your language order
Pick transcription languages yourself or leave auto-detect on. Enable several and drag to set priority, so a bilingual mix picks the right one first.
-
tray
Out of the way
Lives in the system tray on all three OSes. Copy the last transcription without opening the window, and launch at login starting hidden. Ctrl+W hides it, Ctrl+Q quits (Cmd on macOS).
# engines
Pick the engine that fits your machine
Twenty local ONNX models across six families, pocket-sized to max-accuracy, or bring a cloud key: OpenAI, Groq, Sarvam (5 output modes incl. translate and transliterate), or any OpenAI-compatible endpoint. One toggle in the Engine tab.
- IndicConformer11
- Parakeet2
- Whisper3
- Moonshine2
- SenseVoice1
- Qwen31
- OpenAI · Groq · Sarvam · custom
Whisper Large V3 covers 100 languages · Parakeet TDT 0.6B covers 25 · IndicConformer 600M Multi covers 22 Indian languages · Moonshine base is just 56 MB
# download
Download
Wisper v3.4.0. The site fetches the latest release when it loads. Auto-update is built in.
Loading the latest release…
Prefer the AppImage if your distro isn't DEB or RPM based. On macOS, grab the DMG for your chip: the app is unsigned, so right-click + Open on first launch. Updates are signature-verified and install in-app. View all releases →
# setup
First-run setup
Three steps and you're dictating, on Linux, Windows, or macOS.
1 Pick an engine
Open Engine and choose how speech becomes text. Same choices on every OS.
- Local. 20 models across six families, from a 56 MB English model to a 2.45 GB one covering 22 Indian languages. Nothing leaves your machine.
- Cloud. OpenAI, Groq, Sarvam, or any OpenAI-compatible endpoint. Sarvam adds five output modes, including translate and transliterate.
Nothing transcribes until one is active.
Engine tab · local or cloud
2 Pick your mic
In General, choose a specific input device or stay on the system default, then test the level meter before you rely on it.
Each system asks for microphone access differently. Your OS is picked for you below.
General tab · input device
3 Make pasting work
Your text goes back into the app where you started, even if you clicked elsewhere. How Wisper gets it there depends on your system:
Linux · Wayland
Allow the one-time "remote desktop" prompt, or install ydotool to skip it and type faster:
sudo systemctl enable --now ydotoold
sudo usermod -aG input $USER1.0.4 or newer. Older builds type about 40x slower. Setup guide
Linux · X11
Nothing to install. Direct typing works out of the box on GNOME, KDE, Hyprland, Sway, and anything else on X11.
Windows
Nothing to install. The built-in tool pastes with Ctrl+V, and the overlay sits above the taskbar.
macOS
Nothing to install. The built-in tool pastes with Cmd+V. Grant Accessibility once, or the hotkey never fires.
Wayland: ydotool or wtype · X11, Windows, macOS: built-in
What your system needs
Everything specific to one operating system, in one place.
Install
Grab .deb for Debian or Ubuntu, .rpm for Fedora, or the AppImage for anything else. Make the AppImage executable first with chmod +x.
Pasting
Works out of the box on X11. On Wayland, either accept the one-time "remote desktop" prompt or install ydotool to skip prompts and type roughly 40x faster than the built-in tool.
Coming back to your app
On GNOME with Wayland, Wisper installs a small companion extension the first time you run it so it can return to the app you started in. You will need to log out and back in once. Every other desktop needs nothing.
Microphone
No extra grant. Wisper uses PipeWire or PulseAudio, and you can pick the input device in General.
Install
Run the -setup.exe installer. New versions arrive inside the app, so you rarely download again.
Microphone
Allow desktop apps in Settings → Privacy → Microphone. Windows does not warn you: when access is blocked it feeds silence, so the level meter in General just stays flat.
Pasting
Nothing to install. The built-in tool pastes with Ctrl+V, and the overlay floats above the taskbar.
Keys and shortcuts
The hotkey keycap reads Win on Windows. The tray icon holds the last transcription, and Ctrl+W hides the window while Ctrl+Q quits.
Install
Open the .dmg for your chip and drag Wisper to Applications. The build is unsigned, so the first launch needs right-click then Open and a confirmation. If macOS offers no Open option, clear the quarantine flag with xattr -cr /Applications/Wisper.app and launch normally. macOS 13 or newer.
Microphone access
Grant it in System Settings → Privacy & Security → Microphone. Without it, recording captures silence rather than failing.
Accessibility access
Grant it in System Settings → Privacy & Security → Accessibility. Without it the global hotkey never fires and nothing pastes.
After an update
If the hotkey stops working, remove Wisper with the minus button in Accessibility and add it again with plus, turn it on, then relaunch. The keycap reads Cmd, and pasting uses Cmd+V.
# faq
Questions people ask
The honest answers, including the platform-specific rough edges.
How do I install on Linux?
Pick your package from the download section above: deb for Debian/Ubuntu, rpm for Fedora, or the AppImage for everything else (run chmod +x Wisper-*.AppImage first). On Wayland, install ydotool (1.0.4 or newer) afterwards for prompt-free pasting; the setup section above has the commands, plus a full guide.
Does Wisper work on Windows, and how do I install?
Yes, with the same local transcription, overlay, and paste flow as Linux. Grab the NSIS installer from the download section above and run it; stable releases update in-app. If the mic meter stays flat, check Settings → Privacy → Microphone access: Windows delivers silence when access is blocked.
Why is my mic silent on Windows?
Almost always the privacy switch: open Settings → Privacy → Microphone access and allow apps to use the microphone. When blocked, Windows feeds silence instead of an error, so the mic meter in General stays flat. The built-in Test paste button helps confirm everything past the mic is working.
How do I install on macOS?
Grab the DMG for your chip (Apple Silicon or Intel) from the download section above, drag Wisper to Applications, then right-click + Open on first launch: the build is unsigned, so macOS needs that explicit open once. Then grant Microphone and Accessibility access when asked. macOS 13 or newer. If the hotkey ever stops working after an update, the fix is in the setup section above.
Is my audio sent to the cloud?
No, not by default. Transcription runs on your machine with local ONNX models. Cloud providers are optional and only used if you add a key in Engine or Process settings.
Which Linux desktops are supported?
Any Linux with X11 or Wayland, with a global hotkey that works the same in both. For pasting without a permission prompt on Wayland, install ydotool. GNOME Wayland needs a one-time logout for the bundled helper; details are in the setup section above.
Why does it ask for "remote desktop" permission?
That prompt comes from the built-in paste tool on Wayland, which routes input through the XDG Desktop Portal. Installing ydotool or wtype avoids it entirely.
Do I need to download a model?
Only if you want to stay fully local. On first launch, open the Engine tab and pick a local ONNX model: Moonshine for speed and size, Whisper, Parakeet, or IndicConformer for language coverage, plus SenseVoice and Qwen3. Prefer not to? Connect a cloud provider instead. Wisper can't transcribe until one engine is active, and you can cancel a download mid-way.
What is the cloud option?
In the Engine tab you can switch from the local model to a cloud speech provider (OpenAI, Groq, Sarvam, or a custom OpenAI-compatible endpoint) using keys you supply. Sarvam adds five output modes: Transcribe, Translate to English, Verbatim, Transliterate, and Code-mix. You can store several keys per provider for automatic rotation past rate limits. The audio then leaves your machine to that provider, and the Process tab estimates monthly spend.
Which languages can I dictate in?
Whatever your model covers, from 100 languages on Whisper Large V3 down to focused bilingual models. Cloud providers follow their own support. In General settings you can leave detection on auto, or enable several languages and drag them into priority order.
How do overlays, tray, and updates differ per OS?
They don't, much. The overlay pill, system tray, launch-at-login, history, and in-app updates work the same on Linux, Windows, and macOS. Windows pastes with Ctrl+V and macOS with Cmd+V; the hotkey keycap reads Super on Linux, Win on Windows, Cmd on macOS.
What does the AI cleanup actually change?
It re-formats and polishes the transcript through the chat model you configure in Process settings. It rewrites only what you said, never answers for you. Profiles cover auto, cleanup, email, developer, messaging, and formal writing.
Is Wisper free and open source?
Yes. The code is MIT licensed and the repository is public on GitHub. There is no account, no telemetry, and no paid tier.