SKYGGE - AI and music composition

Anyone else feel like AI might be getting a little too close for comfort? Phones and speakers eavesdropping on our intimate moments, social media bots fiddling about with our elections, insidious deepfakes undermining the credibility of our very eyes and ears - long gone are the days when the biggest threat we faced was being beaten at chess.

By Oscar Allan

It’s perhaps understandable if news of AI extending its reach into the sphere of music - surely one of our most hallowed creative pursuits - is met with a little scepticism then. But fear not! Benoit Carré’s collaborative AI project, SKYGGE, is not the cold-hearted, soulless affair the naysayers among you may have been expecting. Rather than expunging music’s human elements, the ‘Flow Machines’ software which Carré works offers new ways to combine, explore, maybe even enrich them, something he was keen to press home at the launch of his new release, The American Folk Songs EP.

Carré began with an explanation of the software’s methodology, which he formulated in collaboration with Spotify Labs. I won’t attempt to dig into the details (or claim to understand them) but the program essentially generates a selection of complex new harmonies between two pieces of music: one a target, the other a source. The source, which Carré equates to the traditional concepts of ‘inspiration’ or ‘style’, is chopped up and sculpted around the target melody in various different, apparently AI-informed ways, with around ten new arrangements generated for each run of the program. It then falls to the composer and the good old fashioned human ear to decide what works, what doesn’t, and stitch it together into a track.

To demonstrate the process, Carré showed how he created a section of his take on the Appalachian ballad ‘Black is the Color’, using Pete Seeger’s 1958 a capella cover as the target and a score of 60s bossa nova-style strings for the source. Despite the disparity between the two pieces, the end result is a strikingly coherent arrangement. While Seeger’s scratchy vocals remain largely intact, they’re now buoyed by swelling orchestral tones and complex chord progressions, all clearly sprung from the source material, but now coaxed and carressed into a radically new framework.

The presentation was followed by a performance of The American Folk Songs EP in its entirety, with live vocals and violin from Canadian folk singer Kyrie Kristmanson and a Funktion One immersive audio setup provided by Sova Audio. As the name suggests, the target pieces are all traditional folk ballads - a poignant choice given the contrast between their old-timey charm and the hyper-virtual process they’re subject to - with sources (projected on a screen throughout the performance) ranging from 17th Century compositions by Henry Purcell and William Byrd to 70s jazz-style electric piano. 

These influences don’t just vary between songs but within them. For ‘Amazing Grace’, Kristmanson recorded a vocal cover that was first melded with Byrd’s ‘Mass for Three Voices’ to give a stunning, intricate chorus of artificial voices. The software then reworked the same vocal track using Purcell’s ‘Now Winter Comes Slowly’ for the next verse and chorus, before the song breaks into jazzy, pop-electronica, almost Four Tet or Caribou-esque percussion, coupled with the kind of pitched-up vocal cuts you might usually find flitting about a Todd Edwards garage remix. The cumulative effect of these multifarious elements was all the more powerful for the spatial separation enabled by the immersive speaker arrangement; disembodied, disjointed vocals swirling amid currents of strings; horns wriggling free from deep sub-bass and circling the room like amiable bats.

While the ability of the Flow Machines software to smoothly integrate all these starkly different influences is impressive, it’s the more unpredictable, left-field nature of its compositional style that is perhaps its most interesting attribute. AI has no regard for musical tradition or history. It’s not daunted by the prospect of combining a Renaissance-era choral mass with an American folk classic. It just knuckles down and gives it its best shot, no matter how weird the results may sound to us mere mortals. As Carré explained in his talk, “I like to work with AI because of the unexpected surprises; chord changes and harmonies that, to me, are inspiring.” He also made clear how crucial a role the composer still had to play in transforming these computer-generated tidbits into a final product though.

This is the aspect of the SKYGGE project that raises the most pertinent, potentially unsettling questions around the future of human creativity in a world becoming more and more permeated by technology; will we always require human input to discern a ‘good’ composition from a ‘bad’ one? Is it only a matter of time before Cowell and his rapacious cronies can shove a fistful of Now CDs into an AI program and have it spit out infinite numbers of perfect pop songs? Are they doing that already?

In my opinion, Carré’s work doesn’t indicate or endorse a movement towards this grim musical dystopia. In fact, I would say it stands more in defiance against that idea. Despite the intense sophistication of the production techniques, the songs on The American Folk Songs EP retain a wonderfully - for lack of a less woolly word - ‘human’ feel. Though this is no doubt bolstered by the emotive choice of the source material, for me a crucial factor of a good song will always be the chorus of semi-perceptible errors that can only come from the clumsy hand of a Homo sapiens.

The singularity may be coming, but I reckon we’ll be the ones soundtracking it.