Hacker News Reader

They're all actually AI powered, generally some form of real-time RNN trained on identifying and isolating voice content from background noise or music.

rnnoise2 is an open-source model that does very well. There also are things like Waves Clarity VX, the Nvidia Broadcast (Audio Effects SDK) too, as well as plenty of other solutions like Supertone Clear, Krisp, etc etc etc.

Does that mean youtube is AI generating your voice to "add it back" after silencing that part of the video? Does it ever generate different words to what you actually said?

It is pretty hard to subtract a signal like that, if it’s really just signal C=A+B then maybe if you have exactly B you can get A back from C, but with microphones picking up radio or something like that? No shot

Also it has to be able to work on remixes that it doesn’t know about of the songs, etc.

All in all I’d say yeah this definitely requires some clever engineering and I believe them that they use AI as part of all of this.

I guess both. One to detect which waves are copyright protected and the transformation to remove it.