numen

README

Numen

Numen is voice control for handsfree computing, letting you type efficiently by saying syllables and literal words. It works system-wide on Linux and the speech recognition runs locally.

There's a short demonstration on: numenvoice.org

Install From Packages

Packages of Numen are available on:

• Alpine
• Arch (AUR)
• Nix (anpandey's flake)

and potentially other platforms.

Install From Source

go (aka golang) and scdoc are required.

A binary of the speech recognition library and an English model (about 40MB) can be installed with:

./get-vosk.sh && sudo ./get-vosk.sh install
./get-model.sh && sudo ./get-model.sh install

The dotool command which simulates the input, can be installed with:

./get-dotool.sh && sudo ./get-dotool.sh install

Finally, Numen itself can be installed with:

./build.sh && sudo ./build.sh install

Permission and Keyboard Layouts

dotool requires permission to /dev/uinput to create the virtual input devices, and a udev rule grants this to users in group input.

You can try:

echo type hello | dotool

and if need be, you could add your user to group input with:

sudo groupadd -f input
sudo usermod -a -G input $USER

and re-login and trigger the udev rule or just reboot.

If it types something other than hello, see about keyboard layouts in the manpage.

Getting Started

Once you've got a microphone, you can run it with:

numen

There shouldn't be any output, but you can try typing hey by saying "hoof each yank", and try transcribing a sentence after saying "scribe". Terminate it by pressing Ctrl+c (aka "troy cap").

If nothing happened, check it's using the right audio device with:

timeout 5 numen --verbose --audiolog=me.wav
aplay me.wav

and specify a --mic from --list-mics if not.

Now you're ready to have a go in your text editor! The default phrases are in the /etc/numen/phrases directory.

Going Further

I use Numen and the default phrases for all my computing, with keyboard-based programs like Neovim and qutebrowser. I also use a minimal desktop environment I made, called Tiles, that doesn't require a pointer device for window management, file picking, etc.

If you'd like to tweak the phrases, copy the default phrases to ~/.config/numen/phrases and edit them there. The manpage covers configuration.

Mailing List and Matrix Chat

You can send questions or patches by composing an email to ~geb/numen@lists.sr.ht.

You're also welcome to join the Matrix chat at #numen:matrix.org.

See Also

• awesome-numen - a list of Numen configs and resources
• sprec - a speech recognition command (if you're just looking for speech to text)
• Tiles - a minimal desktop environment suited to voice control
• Noggin - experimental face tracking I made for playing games

Support My Work 👀

Thank you!

License

AGPLv3 only, see LICENSE.

Copyright (c) 2022-2024 John Gebbie

Extra Responsive Hack

You can append these lines to your model's conf/model.conf to make things extra responsive:

--endpoint.rule2.min-trailing-silence=0.25
--endpoint.rule3.min-trailing-silence=0.25
--endpoint.rule4.min-trailing-silence=0.3

The default model is /usr/share/vosk-models/small-en-us, but you can edit a copy instead and specify it with the NUMEN_MODEL environment variable. We should be able to implement better with the next Vosk release.