Home » Done and Liked, Featured

No composer, no musicians — the science of crowd-sourced music evolution

18 Juni 2012 3,094 views No Comment

“Britain doesn’t need talent” is what Bob MacCallum and Armand Leroi, both at Imperial College, called the DarwinTunes demo at the Imperial Festival just a few weeks ago. What we weren’t allowed to tell anyone back then was that we actually have scientific evidence: we can evolve music from noise without composers or musicians, just by public choice. This week the paper describing our experiments will be out and available from the Proceedings of the National Academy of Sciences (PNAS): “Evolution of Music by Public Choice“.

How to evolve music

Bob, the main author, and Armand had already done a lot of work when they asked me to join the project. Bob had built a software program that stores a bunch of data structures (“genes”), which it can turn into little sound loops. The crucial thing is that these loops can have sex — and they produce new “baby” loops by combining their genes and adding a touch of random mutation. They’d also managed to get a lot of people to vote for these little loops on the darwintunes.org website. So many, in fact, that the tunes evolved over thousands of generations, with the least popular ones “dying” and the most popular ones having lots of sex and making ever more musical babies.

Here’s an amusing podcast, in which Bob explains all that more graphically.

Music informatics analysis shows the evolution

Bob and Armand could already show that the people liked later generations much better, but they didn’t know yet how to show what the evolution did to the audio—the phenotype. So they asked me what we could do to measure its musicality. Just having released the chord estimator Chordino, I figured we could use Chordino’s model likelihood as a measure of chordal clarity. This did indeed turn out to be true, and we can show a dramatic improvement in chordal clarity over the generations. Likewise Bob came up with a rhythmic complexity measure based on rhythm patterns that shows a great increase in line with the popularity ratings.

It’s science, try to beat that as an application for music informatics research!

Well, actually, it’s only science once you can empirically explain what happens…

How to explain what happens

So there we had it, a rich data source that evolutionary biologists can usually only dream of: we had the genes, the phenotypes (the audio loops), the selection statistics and even all the inheritance tree! Now making sense of it was another matter. That’s where Austin Burt came in. He and Armand finally came up with some very nifty mathematics (including the infamous Price equation) that explain the main dynamics of the evolution we observed: the loops get more and more popular on average at first, but then the average lingers at a medium high level of popularity—despite the existence of very popular loops! It turned out that fidelity of transmission was the crux: the most advanced loops were elusive “prodigies” whose clever features wouldn’t rub off to their children.

DarwinTunes keeps evolving live at darwintunes.org.

Comments are closed.