The Scrybe

On-device voice dictation for Linux — speak anywhere, paste anywhere.

Press a global hotkey, an animated overlay appears with a live voice visualizer, you talk, and your words are pasted straight into whatever app you're using. Speech recognition runs locally by default. Configuring remote whisper.cpp or Ollama endpoints sends audio or text to those services.

---

Install

curl -fsSL https://mrelmida.dev/scrybe/install.sh | bash

The installer detects your GPU and only pulls what your hardware needs — it skips the Intel/OpenVINO stack on NVIDIA or AMD machines, for example. It builds Scrybe and sets it up as a desktop app with a global hotkey and login autostart. Then launch Scrybe from your app menu (it lives in the system tray) and press Meta+Alt+D to dictate.

The online installer requires Git. It prefers the newest stable version tag when available, otherwise main, and resolves the fetched revision to one commit before building in a separate checkout. Set SCRYBE_REF to a tag or commit, or SCRYBE_BRANCH=main to explicitly follow development. Existing dirty or unrelated SCRYBE_SRC paths are refused; clean previous sources are retained beside the checkout instead of deleted.

Executable, clipboard helper, and Python sidecar are published together under ~/.local/share/scrybe/runtime (SCRYBE_PREFIX/share/scrybe/runtime for a custom prefix). A failed fetch/build keeps the current app active. Run ~/.local/bin/scrybe-rollback and restart Scrybe to restore the previous bundle. Older bundles and source backups remain available until you remove them. Rollback covers app files; system packages, shared Python environments, models, and configuration are not reverted.

Supported distributions: Fedora, Arch, Debian/Ubuntu, openSUSE (KDE Plasma 6 / Wayland). Works on Intel, NVIDIA, AMD, or CPU-only machines.

Features

• 🎙️ Push-to-talk dictation with a global hotkey (default Meta+Alt+D).
• 🏝️ Animated overlay — a translucent "island" with a live, audio-reactive voice visualizer, anchored to the top or bottom of the screen.
• 📝 Live transcription as you speak (optional).
• ⌨️ Enter to send · Esc to cancel.
• 📋 Clipboard paste into the focused app — no simulated keystrokes; your clipboard is restored afterward.
• 🧠 Optional LLM formatting (local, via Ollama) — clean-up, Markdown structuring, summarizing/shortening, or your own custom style presets.
• ⚙️ Settings window to pick the backend, model, language, and formatting style, and to create custom presets.
• 🔌 Runs on any hardware — Intel, NVIDIA, AMD/Vulkan GPUs, or CPU. The installer only builds the backend your machine needs, and you can add others later from Settings ▸ Hardware.
• ⬆️ Automatic updates — checks GitHub on launch and updates in place.
• 💤 Lightweight — the model loads on demand and unloads when idle.
• 🐧 Native C++/Qt 6 (QML) for KDE Plasma 6 / Wayland.

Hardware support

Scrybe has pluggable speech-to-text backends and picks the best one for your machine automatically (stt/backend=auto):

Backend	Best for	Engine	Installed by default on
openvino	Intel iGPU / Arc / NPU + CPU	OpenVINO GenAI	Intel GPUs
faster-whisper	NVIDIA (CUDA) + CPU	CTranslate2	NVIDIA / CPU-only
whispercpp	Vulkan / AMD / any GPU + CPU	whisper.cpp	on request

auto resolves to faster-whisper on NVIDIA, OpenVINO on Intel, and faster-whisper on CPU elsewhere. The OpenVINO backend is compiled in only when an Intel GPU is present (or forced with SCRYBE_WITH_OPENVINO=1), so the app builds and runs on any machine. Any backend can be forced in the config, and missing backends can be installed on demand from Settings ▸ Hardware (or scripts/install-backend.sh openvino|faster-whisper|whispercpp).

Usage

1. Launch Scrybe (app menu) — it runs in the tray and autostarts on login.
2. Press Meta+Alt+D. The island appears and starts listening.
3. Speak. With live preview on, the text fills in as you talk.
4. Press Enter to finalize and paste, or Esc to cancel.

Open Settings from the tray (or scrybe --settings) to choose the backend, model, language, and formatting style, and to manage custom presets. Rebind the hotkey in System Settings → Shortcuts → Scrybe.

Configuration

Settings live in ~/.config/scrybe/scrybe.conf and most are exposed in the tray.

Key	Values	Default	Meaning
stt/backend	auto·openvino·faster-whisper·whispercpp	auto	recognition engine
stt/model	tiny·base·small·medium·turbo·large-v3·distil	small	model size
stt/device	AUTO:GPU,CPU·GPU·CPU·NPU·cuda·cpu	AUTO:GPU,CPU	compute device
stt/language	auto or a code (en, de, tr, …)	auto	dictation language
stt/vad	true·false	true	skip recordings without detected voice activity
stt/autoSendSecs	seconds; 0 disables	0	finalize after detected speech followed by silence
audio/device	base64 device ID; empty for default	empty	microphone selected in Settings ▸ Microphone
audio/gain	1–25	9	meter/overlay display scale; does not change captured audio or VAD
ui/preview	true·false	true	live transcription preview
island/position	top·bottom	top	overlay anchor
paste/restoreClipboard	true·false	true	restore clipboard after paste
paste/restoreDelayMs	ms	1000	grace period before restoring the clipboard
paste/shortcut	ctrl+v·ctrl+shift+v	ctrl+v	paste shortcut (ctrl+shift+v for terminals)
llm/model	Ollama model	qwen2.5:1.5b	cleanup model
llm/endpoint	URL	http://localhost:11434	Ollama endpoint
presetTemps/<name>	0–1	0.3	custom formatting preset temperature
whispercpp/endpoint	URL	http://127.0.0.1:8080	whisper-server endpoint
update/versionUrl	URL	GitHub VERSION	where to check for updates
update/autoCheck	true·false	true	check for updates on launch

kwriteconfig6 --file ~/.config/scrybe/scrybe.conf --group stt --key model turbo

Models

Speech models download on first use (pick one in the tray → Speech model), or:

scripts/download-model.sh turbo   # tiny|base|small|medium|turbo|large-v3|distil

Key	Model	Notes
tiny	whisper-tiny	fastest, least accurate
base	whisper-base	fast
small	whisper-small	balanced (default)
medium	whisper-medium	accurate
turbo	whisper-large-v3-turbo	best speed/accuracy
large-v3	whisper-large-v3	most accurate, slower
distil	distil-whisper-large-v3	fast, English-focused

Backends

faster-whisper (NVIDIA / CPU)

The installer keeps Python packages in Scrybe's private venv at ~/.local/share/scrybe/venv, so it works on PEP 668 distributions such as Arch without modifying the system Python.

kwriteconfig6 --file ~/.config/scrybe/scrybe.conf --group stt --key backend faster-whisper
kwriteconfig6 --file ~/.config/scrybe/scrybe.conf --group stt --key device cuda   # or cpu

whisper.cpp (Vulkan / any GPU)

Build the server with your backend, then point Scrybe at it:

git clone https://github.com/ggml-org/whisper.cpp && cd whisper.cpp
cmake -B build -DGGML_VULKAN=ON        # or -DGGML_CUDA=ON
cmake --build build -j --target whisper-server
./build/bin/whisper-server -m models/ggml-large-v3-turbo.bin --port 8080

kwriteconfig6 --file ~/.config/scrybe/scrybe.conf --group stt --key backend whispercpp

LLM formatting (optional)

Enable in Settings (or the tray). Before pasting, the transcript is processed by a local Ollama model. Pick a style:

• Clean formatting — fix punctuation, casing, and filler words.
• Markdown — structure into headings, lists, bold, and code blocks.
• Summarize & shorten — condense and remove duplication.
• Custom presets — write your own instruction (e.g. "Rewrite as a formal email."); saved presets appear in the style dropdown.

The prompt asks it to preserve meaning, but generated edits can still change content. It falls back to raw text if Ollama fails. Pull a model with ollama pull qwen2.5:1.5b. Settings includes connection testing, preset generation, and a sample-text preview; review generated instructions before saving them.

Updates

Scrybe checks GitHub for a newer release a few seconds after launch (it reads a plain VERSION file) and shows a tray notification when one is available. Update any time from Settings ▸ Updates (or the tray → Check for updates…) — it pulls the latest source and rebuilds in a terminal. To update manually:

curl -fsSL https://mrelmida.dev/scrybe/install.sh | bash

Build from source

git clone https://github.com/mrelmida/scrybe && cd scrybe
./build-and-setup.sh

Or manually (dependencies already installed):

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build

Common commands:

scrybe --help                        # all options
scrybe --version
scrybe --settings                    # open the settings window
scrybe --transcribe-file file.wav    # self-test the recognizer (no mic)
make build | run | install | clean   # dev convenience targets

Uninstall

scripts/uninstall.sh            # remove the app (keeps your models + config)
scripts/uninstall.sh --purge    # also remove models, config, and OpenVINO

Troubleshooting

• Paste doesn't work — ensure ydotoold is running and you're in the input group (the installer configures this; log out/in once after the first install). For terminals, set paste/shortcut to ctrl+shift+v.
• Hotkey does nothing — check System Settings → Shortcuts → Scrybe.
• Quiet or immediate-start words are skipped — turn off Speech ▸ Skip silence. Energy VAD can recover immediate speech after a brief pause, but cannot reliably distinguish continuous speech from constant background noise. Abrupt noise changes can also trigger it. The microphone meter scale affects display only; adjust actual input volume in your system sound settings.
• First dictation shows "Loading speech model…" — normal; the model loads on demand and is cached afterward.

Architecture

scrybe (C++/Qt6 daemon)
├── core/Controller     state machine (Idle→Listening→Transcribing→Beautifying→Pasting)
├── audio/AudioCapture  microphone capture + level metering
├── stt/                pluggable backends (OpenVINO · faster-whisper · whisper.cpp)
├── llm/LlmBeautifier   Ollama client
├── paste/Paster        clipboard + Ctrl+V injection
├── update/Updater      GitHub version check + in-place update
└── qml/                overlay + settings window (sidebar, hardware, updates)

Global hotkeys use KDE's KGlobalAccel; the overlay is a Wayland layer-shell surface that returns focus to your app for pasting.

Tests

cmake -S . -B build -G Ninja -DBUILD_TESTING=ON && cmake --build build
ctest --test-dir build --output-on-failure     # C++ unit tests + sidecar protocol

CI (GitHub Actions) runs the unit tests on Ubuntu, the sidecar protocol test, and a full application build on Fedora.

Roadmap

• Settings UI (in place of editing the config file)
• Hardware-aware install + on-demand backend management in the UI
• Automatic updates
• Configurable paste shortcut (terminal support via paste/shortcut)
• Energy-based voice activity detection and optional silence auto-send
• Per-app paste shortcut overrides

License

MIT — see LICENSE.