The research side of "minimum modifications required" is with not quite there yet [0]. Audio source separation for now is probably the way to go (see other child comment). Although I imagine it would encounter similar problems with "over the air" type environments.

Black box transfers require large, audibly noticeable modifications. see "Transfer attacks against YouTube’s Content ID" audio example in [0].

Most work has focused on speech recognition, so see final appendix in [1], the authors of which are currently looking at the problem in more detail (stay tuned for results).

Also, over-the-air attacks require a lot more modification than attacks against image classifiers [2].

A lot of recent work (for voice at least) has found signal processing tricks are simple and efficient ways to deal with it [3]. But these untargeted attacks always run the risk of copyright detection flagging it as a different song instead of "not copyrighted".

So, basically, more research is required to get to the Fawkes "for audio" [4].

As an aside: I would not want to be "that guy" who invented the technology to defeat audio copyright detection systems. I think copyright is a good thing and would only work on something like that to help improve a given system. This specific issue should be handled by through some form of legal proceeding.

[0]: https://www.cs.umd.edu/~tomg/projects/copyrightattack/

[1]: https://arxiv.org/abs/2007.06622

[2]: https://arxiv.org/abs/1908.01551v1

[3]: https://arxiv.org/abs/1910.05262

[4]: https://sandlab.cs.uchicago.edu/fawkes/

Edit: phone went mad and wouldn't let me type things.

Put a very hard filter around the vocal range of sound, hopefully that would be enough to distort the music so that the algorithms can’t identify it.

I wouldn't think it to be exceptionally hard to do: a song is a fairly static piece of audio, with broadly available references. Using that reference to isolate and remove the song from the audio stream might not be easy, but it should be possible.