Local GPU dictation that actually knows your codebase.
The open-source Superwhisper & Wispr Flow experience for Linux and Wayland. Parakeet, Whisper, or Cohere on your own GPU, grounded by what's on your screen, cleaned up without rewriting what you meant.
curl -fsSL https://shuush.kitsunekode.in/install.sh | bashTry the Shuush hotkeys
Four bindings are the whole interface. Press them on your keyboard, or click a key to watch the pill react.
Measured on a daily-driver laptop, not a slide.
Telemetry from real dictation, not a synthetic test. When a take is already clean, Shuush skips the LLM entirely and delivers in 24 ms.
Intel i5-13450HX · RTX 4050 Laptop 6 GB · 23 GiB RAM · Arch + Hyprland · ~8.1 hours saved vs 40 WPM typing
Native GTK4. Live diffs. One switch to go offline.
No Electron. Dictations stream in through Gio.FileMonitor, every repaired symbol is highlighted, and grounded jargon syncs straight into voxtype's daemon replacements.
Every take streams in live. Open one to see the word-level Heard → Cleaned diff, replay the audio, or teach a fix with Ctrl+Enter.

Feels native to Wayland. Never rewrites what you meant.
Recorded at 60 fps on Arch + Hyprland with gpu-screen-recorder. Every cleanup passes a deterministic span gate, so the model can fix mishearings from your screen without paraphrasing you.
The pill, as it really runs
Listening, live word wash, polishing, delivered. Quickshell, GPU-recorded, unedited.
- ListeningLive PipeWire waveform + target window
- Live previewStaggered word-by-word color wash
- PolishingAnti-paraphrase cleanup, ~500 ms
- DeliveredGathered checkmark, shrink-to-dot exit
Five context sources, one gate
App, AT-SPI screen and draft, selection, a ≤15 s clipboard, and the date. Pick a scenario.
Wispr Flow vs Superwhisper vs Shuush + voxtype
| Capability | Wispr Flow | Superwhisper (Omarchy) | Shuush + voxtype |
|---|---|---|---|
| Price & word cap | $15/mo | 10,000 free words/mo | Free · MIT · unlimited |
| Native Linux / Wayland | No (macOS / Windows) | Omarchy / macOS | Hyprland · Sway · Niri · Omarchy · Arch |
| GPU & ASR freedom | Cloud only | Apple Silicon / cloud | NVIDIA CUDA · AMD/Intel Vulkan · CPU |
| Context sources | App + screen | App + selection + clipboard | App + AT-SPI screen & draft + selection + ≤15 s clipboard + date |
| Offline privacy mode | No | Paid local tier | 1-click air-gapped Ollama + AT-SPI password bypass |
| Self-learning dictionary | Proprietary | Manual / AI audit | Phonetic auto-mine + reject list + daemon [text] sync |
Three steps on stock voxtype-bin.
Shuush sits on top of voxtype instead of replacing it. Each project owns one job, so nothing is duplicated and nothing drifts.
# 0 · optional pre-flight check $ curl -fsSL https://shuush.kitsunekode.in/doctor.sh | bash # 1 · voxtype + Wayland deps $ yay -S voxtype-bin wtype wl-clipboard jq python-gobject libadwaita gtk4 at-spi2-core $ voxtype setup # 2 · shuush on top (1 line) $ curl -fsSL https://shuush.kitsunekode.in/install.sh | bash # 3 · verify, then open Studio $ shuush doctor && shuush studio
Runs where you do
- Compositors
- Hyprland · Sway · Niri · Omarchy
- GPUs
- NVIDIA CUDA · AMD ROCm/Vulkan · Intel Vulkan · CPU
- Engines
- Parakeet · Whisper · Cohere GGUF · Moonshine
- Cleanup
- Local Ollama, or Groq → OpenRouter failover
voxtype
by Peter Jackson · MIT
- PipeWire audio capture
- CUDA / Vulkan / CPU speech recognition
- Wayland typing
shuush
this project · MIT
- 5-source context pipeline
- Anti-paraphrase span gate
- Phonetic self-learning
- KitsuneHud pill & GTK4 Studio
kuwai
your rice, or any config
- Hyprland binds & window rules
- Optional island OSD hosting
- First-party auto-discovery of
shuush/libexec
Shuush wouldn't exist without voxtype. It's a companion layer, not a fork: zero drift, works with stock voxtype-bin, and the MIT credit stays where it belongs.
Direct answers before you install.
What is the best open-source Superwhisper or Wispr Flow alternative for Linux?
Shuush, built on stock voxtype, is a free, MIT-licensed Linux/Wayland alternative to Superwhisper and Wispr Flow. It runs local GPU speech recognition (Parakeet, Whisper, or Cohere GGUF) with no word caps, adds 5-source AT-SPI screen, selection, and clipboard context, anti-paraphrase LLM cleanup (100% offline Ollama or ~500 ms Groq failover), and a native GTK4 Studio.
Does Shuush work on Omarchy, Hyprland, Sway, Niri, and non-NVIDIA GPUs?
Yes. Shuush installs on top of stock voxtype-bin across NVIDIA CUDA, AMD ROCm/Vulkan, Intel Vulkan, and CPU-only laptops. Studio reads whichever voxtype engine you configured (parakeet, whisper, cohere, or moonshine), with first-party kuwai rice auto-discovery and standard Hyprland, Sway, and Niri support.
Can Shuush run 100% offline without sending audio or text to the cloud?
Yes. Speech recognition always runs locally via voxtype, and Shuush Studio has a 1-click air-gapped offline mode that pins cleanup to local Ollama (127.0.0.1:11434). Password fields detected via AT-SPI bypass LLM cleanup entirely and are never written to history.
What are the default Shuush hotkeys?
Super+Ctrl+D toggles dictation. Hold F9 for push-to-talk, or tap it (under 300 ms) for hands-free. Super+Ctrl+E is voice command, which rewrites your highlighted selection in place. Super+Alt+D opens Shuush Studio, and a double-tap of Esc cancels a take.
Why build Shuush on top of voxtype instead of forking it?
voxtype already nails PipeWire audio capture, CUDA/Vulkan/CPU inference (Parakeet, Whisper, Cohere GGUF), and Wayland typing. Keeping Shuush as a companion layer means zero fork drift, instant compatibility with stock voxtype-bin from the Arch AUR and Omarchy, and proper credit to Peter Jackson's MIT daemon.
How does the anti-paraphrase span gate prevent LLM hallucinations?
voxtype_pp_gate.py aligns every cleaned reply back against the raw recognizer transcript. Ordinary English words are never rewritten unless grounded by a phonetic match in your vocabulary, the visible AT-SPI screen, your highlighted selection, or a fresh (15 seconds or newer) clipboard.
Say it once. It lands exact.
Free, local, and yours to fork. Two minutes from clone to your first clean take.
curl -fsSL https://shuush.kitsunekode.in/install.sh | bash

