← All posts

Offline speech-to-text on Linux with nerd-dictation

linuxtools
Download source Contribute

If you want offline speech-to-text on a Wayland desktop, nerd-dictation is a great choice: it records from your microphone, transcribes locally (no cloud), and types the result into the focused window. This post documents my working setup on GNOME (PureOS/Librem 13), with the Italian Vosk model and ydotool for Wayland typing.

What you need

  • Linux with PipeWire (input through pw-cat)
  • Python 3.11+ with python3.11-venv
  • A Vosk model in your language, e.g. vosk-model-small-it-0.4 for Italian (~40 MB)
  • ydotool to inject keystrokes on Wayland

Install

git clone https://github.com/ideasman42/nerd-dictation.git
python3 -m venv nerd-dictation/.venv
nerd-dictation/.venv/bin/pip install vosk

mkdir -p ~/vosk-models
cd ~/vosk-models
curl -O https://alphacephei.com/vosk/models/vosk-model-small-it-0.4.zip
unzip vosk-model-small-it-0.4.zip

Vosk is language-specific: pick the model matching your speech, there is no language option.

Give ydotool write access to /dev/uinput

nerd-dictation needs a helper to send keystrokes. On Wayland, wtype will not work on GNOME (the compositor does not expose the virtual-keyboard protocol), but ydotool injects keys through /dev/uinput, which works on every Wayland compositor. It only needs permission, granted once with a udev rule:

sudo apt install ydotool
echo 'KERNEL=="uinput", GROUP="input", MODE="0660"' | sudo tee /etc/udev/rules.d/99-input-uinput.rules
sudo udevadm control --reload && sudo udevadm trigger
sudo usermod -aG input $USER

Group membership only takes effect after logout/login.

Dictate

Start and stop dictation with the begin and end commands. end types everything recognized into the window that has focus:

nerd-dictation/.venv/bin/python nerd-dictation/nerd-dictation begin \
  --input PW-CAT \
  --simulate-input-tool YDOTOOL \
  --vosk-model-dir ~/vosk-models/vosk-model-small-it-0.4

# ... speak ...

nerd-dictation/.venv/bin/python nerd-dictation/nerd-dictation end

Note on the Debian/Ubuntu ydotool package (version 0.1.8): it uses named keys, so two lines of nerd-dictation’s simulate_typing_with_ydotool() must be adapted, otherwise it types literal “1” characters instead of backspaces and never deletes previous text:

  • ydotool type --next-delay 5 -- <text> becomes ydotool type --key-delay 5 -- <text>
  • ydotool key -- ... 14:1 14:0 ... becomes ydotool key --repeat N Backspace

One-key enable/disable with sound feedback

Put a toggle script on $PATH (here ~/.local/bin/nerd-toggle) and bind it to a global shortcut. It plays a different chime on start and stop and shows a notification, so you know whether dictation is active (chimes from Handy’s resources, any short wav works):

#!/bin/bash
PY=$HOME/nerd-dictation/.venv/bin/python
ND=$HOME/nerd-dictation/nerd-dictation
MODEL=$HOME/vosk-models/vosk-model-small-it-0.4

if pgrep -f "$PY .*nerd-dictation begin" > /dev/null; then
    pw-play /usr/lib/Handy/resources/marimba_stop.wav > /dev/null 2>&1
    notify-send -i audio-input-microphone "Dictation stopped" > /dev/null 2>&1
    sg input -c "$PY $ND end"
else
    pw-play /usr/lib/Handy/resources/marimba_start.wav > /dev/null 2>&1
    notify-send -i audio-input-microphone "Dictation started" > /dev/null 2>&1
    sg input -c "setsid nohup $PY $ND begin --input PW-CAT --output SIMULATE_INPUT --simulate-input-tool YDOTOOL --vosk-model-dir $MODEL > /tmp/nerd-dictation.log 2>&1 &"
fi

(The sg input -c wrapper is only needed until the current login session has the input group.)

Register the shortcut in GNOME:

gsettings set org.gnome.settings-daemon.plugins.media-keys custom-keybindings "['/org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/nerd/']"
gsettings set org.gnome.settings-daemon.plugins.media-keys.custom-keybinding:/org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/nerd/ name 'Nerd Dictation Toggle'
gsettings set org.gnome.settings-daemon.plugins.media-keys.custom-keybinding:/org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/nerd/ command "$HOME/.local/bin/nerd-toggle"
gsettings set org.gnome.settings-daemon.plugins.media-keys.custom-keybinding:/org/gnome/settings-daemon/plugins/media-keys/custom-keybindings/nerd/ binding '<Control><Alt>space'

That is all: press Ctrl+Alt+Space, speak, press again, and the text lands in the focused window.