OpenFlow

OpenFlow is a local-first, privacy-focused dictation app for Linux. Hold a hotkey to talk, release to transcribe on-device, optionally clean up the text, and paste into the active field without clobbering your clipboard.

• OS support: Linux only (Wayland + X11)
• CPU arch: x86_64 only
• Display servers: Wayland + X11 (X11 supported)
• HUD: built-in overlay window on most setups; optional GNOME Shell HUD extension for GNOME Wayland

Tested / supported distro baselines:

• Ubuntu 24.04+
• Debian 12+
• Fedora 40+
• Arch (rolling)

Install (single command)

curl -fsSL https://github.com/logabell/OpenFlow/releases/latest/download/install.sh -o /tmp/openflow-install.sh && bash /tmp/openflow-install.sh

Non-interactive (CI-friendly):

curl -fsSL https://github.com/logabell/OpenFlow/releases/latest/download/install.sh -o /tmp/openflow-install.sh && bash /tmp/openflow-install.sh --yes

Uninstall:

curl -fsSL https://github.com/logabell/OpenFlow/releases/latest/download/install.sh -o /tmp/openflow-install.sh && bash /tmp/openflow-install.sh --uninstall

What The Installer Does

The installer is a tarball-based Linux install designed for predictable system integration.

• Installs OpenFlow under /opt/openflow/
• Adds an openflow launcher at /usr/local/bin/openflow
• Installs a desktop entry and icons
• Installs runtime dependencies via your distro package manager (apt/dnf/pacman/zypper where available)
• Configures permissions for global hotkeys + paste injection (see Linux notes below)
• Downloads models (defaults to Parakeet ASR + Silero VAD unless you choose otherwise)

Note: It only uses sudo for system changes; do not run the installer as root.

How It Works

At runtime, OpenFlow follows this pipeline:

1. Capture microphone audio (16kHz mono)
2. Preprocess audio (echo/noise control via WebRTC APM when enabled)
3. Detect speech (Silero VAD when available; otherwise energy fallback)
4. Transcribe on-device (Parakeet by default; Whisper optional)
5. Optional deterministic cleanup (Tier-1 autoclean)
6. Paste into the active field while preserving your clipboard

HUD (Visual Dictation Status)

OpenFlow can show a small on-screen HUD orb to indicate warming / listening / processing.

• GNOME Wayland: ships an optional GNOME Shell extension (OpenFlow HUD, UUID openflow-hud@openflow) for compositor-native rendering
• Other Wayland compositors + X11: uses a regular overlay window (best-effort; some tiling/fullscreen setups may hide or constrain it)

You can toggle the HUD in Settings -> HUD Overlay.

Linux Permissions (Wayland + X11)

On Wayland, apps cannot reliably capture global hotkeys or inject keystrokes via the compositor. OpenFlow uses Linux kernel input devices for a compositor-agnostic workflow:

• Global hotkeys: reads /dev/input/event* (requires access via the input group)
• Paste injection: creates a virtual keyboard via /dev/uinput

Recommended:

1. Start the app
2. Open Settings -> Linux Setup
3. Click Enable (admin)
4. Log out and back in (required for group membership to take effect)

Security note:

Membership in the input group and access to /dev/uinput allows reading global key events and injecting input. Only enable this on machines you trust.

Models (What To Download)

OpenFlow can run with different on-device ASR engines. You can manage models in Settings -> Models.

Model	What it's good for	Tradeoffs	Notes
Silero VAD (required)	Reliable speech detection (start/stop trimming + diagnostics)	Small download, minimal CPU	If it fails or isn't installed, OpenFlow falls back to an energy-based VAD
Parakeet ASR (default)	Low latency dictation on CPU; great default	Slightly less accurate than large Whisper models	Good "always-on" workflow; recommended for most users
Whisper CT2 (Accuracy-first)	Strong accuracy on CPU (especially "small"/"medium")	Larger downloads; higher latency on laptops	Uses CTranslate2 / faster-whisper model formats; compute type follows the precision setting
Whisper ONNX (Advanced)	Sherpa-based Whisper; choose int8 vs float	More variants; can be heavy	int8 is faster; float is usually higher quality but slower

Whisper variants (recommended starting points):

• small + int8: best balance on most CPUs
• medium: higher accuracy, often noticeably slower
• large-v3 / large-v3-turbo: best accuracy (or best speed among large), highest resource use

Language notes:

• en variants are smaller and optimized for English
• multi supports multilingual (and is forced for the large-v3* models)

Default Settings

These are the defaults shipped in the app:

Setting	Default
Talk mode	Hold-to-talk (hotkeyMode=hold)
Hotkey	RightAlt
ASR engine	Parakeet (asrFamily=parakeet)
Whisper backend	ct2
Whisper model	small
Whisper language	multi
Whisper precision	int8
VAD sensitivity	medium
Paste shortcut	ctrl-shift-v
Language	auto + autoDetectLanguage=true
Autoclean	fast

Development (Linux)

All commands run from app/:

yarn install
yarn tauri dev

Notes:

• Rust toolchain: Rust 1.78+ recommended
• Whisper CT2 backend dependency: SentencePiece + pkg-config
  • Debian/Ubuntu: sudo apt install libsentencepiece-dev pkg-config
  • Fedora: sudo dnf install sentencepiece-devel pkgconf-pkg-config
  • Arch: sudo pacman -S sentencepiece pkgconf

Repo Structure

• app/: Frontend (React + TypeScript + Vite) and Tauri configuration
• app/src-tauri/: Rust backend (audio, ASR, VAD, models, output injection)
• scripts/: Packaging helpers
• docs/: Architecture notes