murmel

Push-to-talk dictation for Linux/Wayland. Runs locally on your GPU.

Bind a hotkey, hold to talk, release to transcribe. murmel captures audio, transcribes it with NVIDIA Parakeet (via NeMo), and types the result at your cursor. No cloud services, no API keys.

1. A daemon listens for SIGUSR1. Bind any key to murmel toggle.
2. Hold the key to record, release to stop.
3. Audio is transcribed locally (GPU or CPU).
4. Text is injected at the cursor via wtype (Wayland), xdotool (X11), or clipboard fallback.

git clone <this-repo>
cd murmel
chmod +x install.sh
./install.sh

The installer handles pacman packages, sets up a Python 3.13 venv with uv, installs NeMo, and configures the systemd service.

# System deps
sudo pacman -S --needed wtype libnotify portaudio

# Create venv with Python 3.13 (NeMo doesn't fully support 3.14 yet)
uv venv --python 3.13 .venv

# Install Python deps (pulls in PyTorch, ~2 GB first time)
uv pip install sounddevice soundfile numpy 'nemo_toolkit[asr]'

# Install wrapper script
mkdir -p ~/.local/bin
cat > ~/.local/bin/murmel <<EOF
#!/bin/sh
exec "$(pwd)/.venv/bin/python3" "$(pwd)/murmel.py" "\$@"
EOF
chmod +x ~/.local/bin/murmel

Add to your Hyprland config (e.g. ~/.config/hypr/bindings.conf):

# Push-to-talk: hold Super+D to record, release to transcribe
bindd = SUPER, D, Start dictation, exec, ~/.local/bin/murmel toggle
bindrd = SUPER, D, Stop dictation, exec, ~/.local/bin/murmel toggle

# Or simple toggle with F9 (press to start, press again to stop)
bind = , F9, exec, ~/.local/bin/murmel toggle

Note: Use the full path (~/.local/bin/murmel) since Hyprland's exec environment may not include ~/.local/bin in PATH.

Start the daemon (auto-starts on login if the service is enabled):

systemctl --user start murmel   # or: murmel start

List models and switch:

murmel models
murmel config --set model=parakeet-tdt-1.1b

You can also pass any HuggingFace model ID directly:

murmel start --model nvidia/parakeet-tdt-1.1b

murmel start [--model KEY] [--inject MODE] [--no-preload]
murmel toggle        # send SIGUSR1 to the running daemon
murmel stop          # send SIGTERM
murmel status        # show PID and current config
murmel models        # list models
murmel config        # view config
murmel config --set KEY=VALUE ...

# Examples
murmel config --set model=parakeet-tdt-1.1b
murmel config --set notify=false
murmel config --set inject_mode=clipboard
murmel config --set log_level=DEBUG

auto checks $WAYLAND_DISPLAY and picks wtype, falling back to xdotool on X11.

Show recording state in Waybar with a custom/murmel module:

~/.config/waybar/config:

"custom/murmel": {
    "exec": "[ -f /tmp/murmel.pid ] && echo '🎙' || echo ''",
    "interval": 1,
    "on-click": "murmel toggle",
    "tooltip": true
}

journalctl --user -u murmel -f     # systemd logs
tail -f ~/.config/murmel/murmel.log

NeMo uses PyTorch and will use a CUDA GPU if one is available. The 0.6B model runs in roughly 0.5-2s on CPU on modern hardware.

No audio captured: check your default input device:

.venv/bin/python3 -c "import sounddevice; print(sounddevice.query_devices())"

wtype fails in XWayland apps: switch inject mode to xdotool or clipboard:

murmel config --set inject_mode=clipboard

Model download fails: NeMo caches models in ~/.cache/huggingface/hub/. If a download is interrupted, delete the partial files and restart.

MIT