[crazygringo]

Content ID is never going to be perfect (so there always needs to be an appeal mechanism).

But I'd always expected that false matches would be due to some kind of hash collision, not different recordings of the same piece.

Because YouTube doesn't know what the composition is, it only compares to other performances (which are under copyright).

And the fact that two different pianists could interpret and perform a work identically enough to generate the same signature Content ID uses is... astonishing to me, from my classical music background. If I'd had to guess, two musicians normally probably couldn't achieve that even if they tried.

So this seems to just be a crazy statistical fluke? Otherwise classical performances would be getting blocked left and right, given the hundreds/thousands of different performances of the same underlying music? Or is something more advanced that audio fingerprinting going on?

[sdenton4]

Content ID is designed for flexibility as well: it has to catch a lot of transformative that might be applied to try to bypass it. Perhaps you've noticed that actually evasive material has increasingly terrible effects layered in as the years go on... (I remember recently finding a cartoon that looked like it was a zoomed in camcorder recording from a CRT... But emulated in effects layers.) I've also seen things subtly slowed down or sped up, strangely cropped, bordered, then given animated borders once that stopped working.

Between an incessant evasion arms race and a content generation industry all to happy to profit from false claims, I just think to myself, "this is why we can't have nice things."

[word-reader]

You can tell it's not a fluke if you look at the copyright claims that are sometimes available in the "Show More" section. Frequently, it will show claims that recordings of the same piece by other performers are part of the video.

Here is one example: https://www.youtube.com/watch?v=RxDCgFFBk20

Two of the movements have correct copyright claims for the performance, but another of the movements has a claim for a recording by a different conductor and orchestra. I see this all the time.

Like another person said here, we don't know how Content ID works or what exactly it's looking for, but it's clearly not an "exact match" of sound waves or something: it has to catch covers of compositions that are still under copyright, for example [1].

[1] https://diymusician.cdbaby.com/youtube/posting-cover-songs-o...

[SamReidHughes]

I had a recording of the Revolutionary Etude, filmed on a phone (iirc), played on an out of tune Acrosonic piano (more a piece of furniture than a functioning piano), with no pedal usage, get copyright matched with a Sony published recording. I disputed the claim and that was it.

It's about as far from a professional recording as you can get on YouTube, except for some piano recital videos featuring ugly children inanimately reading their way through the sheet music.

[brokensegue]

Hash collisions? No, I don't think you understand how these content matching systems work.

[imihai1988]

well, that does not seem to be a very constructive answer. mind sharing some info/references on this specific one, or any others ? because I though it's a blackbox type of software

[brokensegue]

I mean how exactly they work isn't public knowledge. But we can be sure they don't just hash anything. It would be too easy to just add small amounts of noise and defeat them.

Here's a reference I googled for you https://www.toptal.com/algorithms/shazam-it-music-processing...

[imihai1988]

I think you misunderstood the parent comment ( or I misunderstood your reply to that). He talks about about a signature (which is pretty much what it comes down to), not a file hash as in here's a hash of the file contents. That would be meaningless.

As for the reference you provided I know how shazam works, they pretty much made a blog post about it. This is not that. Not by a long shot.

[brokensegue]

I can't think of a useful copyright matching technique that wouldn't flag different renditions of the same song. That's my point. The failures of these techniques wouldn't be hash collisions but songs that legitimately sound similar.