LinuxWhispr

Privacy-first, open-source voice dictation for Linux — a system-wide dictation tool that replicates the full experience of Wispr Flow, with AI-powered text refinement, context awareness, and seamless integration into any text field.

Features

• 🎤 System-wide dictation — works in any text field (browsers, IDEs, terminals, chat apps)
• 🔒 Privacy-first — local-first transcription via faster-whisper, no cloud by default
• 🧠 AI text refinement — removes filler words, fixes grammar, adapts to context (email/code/chat)
• ⌨️ Global hotkey — configurable hotkey works regardless of focused window (X11 + Wayland)
• 🗣️ Voice commands — "make this more formal", "summarize the selected text"
• 📚 Custom dictionary — teach it your names, technical terms, brand names
• 🔄 Adaptive learning — automatically learns corrections you make
• 🌍 99+ languages — automatic language detection via Whisper
• 🖥️ Linux native — GTK4 + libadwaita UI, PipeWire/PulseAudio audio, X11 + Wayland support

Quick Start

System Dependencies

# Ubuntu/Debian
sudo apt install python3-dev portaudio19-dev libgtk-4-dev libadwaita-1-dev \
    gobject-introspection libgirepository1.0-dev \
    wl-clipboard xdotool xclip

# Fedora
sudo dnf install python3-devel portaudio-devel gtk4-devel libadwaita-devel \
    gobject-introspection-devel \
    wl-clipboard xdotool xclip

# Arch Linux
sudo pacman -S python portaudio gtk4 libadwaita \
    gobject-introspection \
    wl-clipboard xdotool xclip

Install

pip install linux-whispr

Run

linux-whispr

On first run, a default config will be created at ~/.config/linux-whispr/config.toml. The Whisper base model will be downloaded on first dictation (~150MB).

Usage

1. Press F12 (default hotkey) to start recording
2. Speak your text
3. Press F12 again to stop (or wait for auto-stop after 2s of silence)
4. Text appears at your cursor position

Configuration

Edit ~/.config/linux-whispr/config.toml:

[hotkey]
dictation = "F12"
command = "<Ctrl><Shift>h"
mode = "toggle"  # or "push-to-talk"

[stt]
backend = "faster-whisper"
model = "base"           # tiny, base, small, medium, large-v3, large-v3-turbo
device = "auto"          # auto, cpu, cuda

[ai]
enabled = false          # enable AI text refinement
backend = "none"         # none, local, openai, anthropic, groq

[audio]
silence_duration = 2.0   # seconds of silence before auto-stop
silence_threshold = 0.5  # VAD sensitivity (0.0-1.0)

[injection]
method = "auto"          # auto, wtype, xdotool, ydotool, clipboard-only
preserve_clipboard = true

Development

# Clone
git clone https://github.com/your-username/linux-whispr.git
cd linux-whispr

# Install in development mode
pip install -e ".[dev]"

# Run tests
pytest

# Type checking
mypy src/

# Linting
ruff check src/

Architecture

Hotkey → Audio Capture → VAD (auto-stop) → STT (faster-whisper) → AI Refinement (optional) → Text Injection

Components communicate via an internal event bus. See specs/001-linux-whispr/spec.md for the full specification.

License

MIT