It could be a buzzword but I think it'd be fair to assume that different covers of the same song are slightly out of sync at different points. So the algorithm (AI or not) would look for signature beats in the original and match them with beats in a close range in the other songs.

I mostly got videos with the original song or overlaid at least. I am curious how well it could snyc actual 'covers'.

https://www.youtube.com/watch?v=lu1vdDhz-_c

https://www.youtube.com/watch?v=ff60IrtLogU

https://www.youtube.com/watch?v=QG0xMK1ciaA

https://www.youtube.com/watch?v=ByWD2cdIyIk

https://www.youtube.com/watch?v=p_RdZwWJzm4

https://www.youtube.com/watch?v=qfn_9sz1yaQ

https://www.youtube.com/watch?v=A2MhpkYjHLw (exception)

I would bet it's using NLP (AI) and looking for lyric matches based on the captions. You could likely make a cover with an incorrect beat but the same lyrics and the AI would match to it.

I bet not. I've been doing NLP for just a few months but my experience is that its fickle at best. Especially with a song that is so repetitive you'd be looking for (n) instance of (n) words.

Also, most of these videos arn't going to have human transcribed lyrics which would further increase the difficulty.

Why not just take all the data they have and throw it into a model to predict time.

On second run it looks like you're right. They have categories that have no lyrics at all.

These days "with the help of computers" doesn't have the same cachet. I did notice that the videos tend to follow instruments and styles. So it may be a bit more than just temporal synchronization.