We recently ran across an Instagram post pointing out the similarities between the chorus of Dua Lipa’s 2020 hit Break My Heart and that guitar riff in Need You Tonight by INXS back in 1987. The post, which had many ❤️s, resonated with us, since the truth is that everyone imitates, no music is created ab nihilo, all artists synthesize the bits and pieces of all the music they have ever heard and floats in their brain, and this is just normal creation. And, in this case the artists involved have reported that the whole process was congenial: Dua Lipa and her collaborators noticed the similarity, got in touch with INXS, and added Andrew Farriss and Michael Hutchence as writers before the song was released. Farriss has since said that he appreciates being a part of the billion-plus streamed track.
Anyway, it got us wondering. A good deal of what we build here exists to notice exactly this sort of thing, so we decided to see if it would
Play them one after the other and you will hear it pretty quickly. The tempos are neighbors, the chords similar, and there is the clipped event on the back half of the beat.
Our Audio ID fingerprint answers a narrow question and answers it very well: is this pretty much exactly the same recording. The harder question, the one behind our Version ID work, is whether this is the same piece of music, performed again by other people in another room on other instruments: cover versions, live takes, an AI imitation.
For that we use our VIBE embedding. A window of audio goes in, and what comes out is a single point in a space of many dimensions, arranged during training so that two performances of the same composition land near one another while two unrelated recordings land far apart.
We took both original recordings, cut each of them into overlapping half minute windows, computed our embeddings for all of them, and measured the distance between every pair. This is known in the biz as a cross-similarity matrix. The y-axis is the timeline of Break My Heart and the x-axis the timeline of Need You Tonight. A darker square in the grid means that the embeddings are similar, i.e. that the music at that x-location is similar to the music at that y-location.
For instance, we can see a dark spot in the upper left of the outlined area, and it sits where a listener would prick up their ears, i.e. the chorus of Break My Heart laid against the riff of Need You Tonight. And the vertical smear underneath it is interesting too. Follow that column down and you can see the song’s structure. Each time the Dua Lipa chorus comes back, around two minutes and again near the end, the same stretch of INXS returns as its closest neighbor, and the varying darkness of the patch shows how each repeat varies musically.
Compare this to a self-similarity graph of Break My Heart, i.e. a graph that compares it to itself. This is a view that was pioneered by our friend Jonathan Foote decades ago, where everything off the diagonal shows the song’s internal structure. It’s easy here to see the chorus repeating itself, and you can also see that same pattern that we saw above along the left-hand side.
It’s pleasing that, inside the pair of songs, the geometry found the right half minute, three times over, at the place an amateur musicologist would have noticed. Remember that locating a resemblance and adjudicating one are separate problems. In this case, it’s not a strong enough match that we currently consider it a full-on cover, as too much around it is different. So…
Because it was fun, and because not everything worth investigating is a product. This was an afternoon’s detour from the Version ID work rather than a preview of anything, and we would rather put it in public as the curiosity it is than let it sit in a notebook. We are spending more of our time out in the open these days, at conferences and in papers and behind a few sponsorships, on the theory that the best questions in music technology tend to arrive from outside the building.
For more than 25 years, Audible Magic has delivered B2B solutions in content identification, for both music and TV/film, along with music catalog fulfillment and rights administration. Our systems power billions of transactions every month for customers around the world, helping platforms reduce DMCA takedown notices and administer licenses for user-generated content. The new VIBE-powered Version ID is the latest step in that work. We’d love to hear about your use case and figure out how to solve it together. Contact us to learn more.
Erling Wold, Chief Scientist and Technologist