It's pretty great, works decently, but the sad thing is the author put it out a few years ago, trained it on proprietary datasets, wrote a paper and then moved onto something else, never to be seen again. Now it's pretty much unmaintained since it requires some very specific ML and DSP knowledge to hack on.
https://github.com/xiph/rnnoise
NVIDIA Broadcast in my opinion is better, but rnnoise will do if you can't run it.
Turns out the background noise removal was turned on for my output too.
I know that Slack and Google are slowly baking these features in, but NVIDIA are in a league of their own. AMD not having them made my GPU brand choice easy last time I upgraded.
Sound doesn't stop at the property line.
And babies? That was during lockdowns when my then wife was looking after the young one (we took turns; obviously this applies to when I was in my meetings). Didn't stop the 2-3y child from being very noisy at times.
Regardless of the source, Broadcast does an amazing job of filtering and cleaning it up. Recipients could not hear any of the background commotion.