While other AI forays into music have usually drawn from existing archives, Herndon wanted not only to build Spawn from the ground up but to input original audio.

“We didn’t want to get in this situation where we were glorifying the music of the past,” she says. “We wanted it to be this living, breathing thing.”

It would would mean many hours of recording, many more hours waiting for Spawn to parse the vast sound files – and then many months for her to produce anything they could use.

“The first six months were extremely frustrating because we didn’t get any interesting results at all,” says Herndon.

“There might have been something interesting from a research level but there was nothing interesting musically.” (Herndon, American-born but based in Berlin, was also researching a PhD in AI and music at Stanford University.)

They nearly gave up. And then suddenly, signs of exciting, exquisite life – albeit as kind of digitally strangled, nonsensical “speech”.

“It was sonically new to me,” Herndon says of those first sounds that would become the track Birth. “It’s kind of wild.”

After that first breakthrough, “it just snowballed”.

Herndon edited and overlaid many of Spawn’s offerings into cohesive album tracks. Yet when they trained Spawn with music by fellow electronic artist Jlin (pronounced jay-lin), the AI created a track that went unedited and played for its entirety on the album.

“We were all jumping up and down in the studio when we heard [it],” she says of Godmother, an eerie whirlwind of percussive vocalisations, akin to a wheezing beatboxer in space.

“I didn’t program or train Spawn necessarily to make any of those sounds. It felt like there were some creative decisions being made by Spawn.”


Godmother is credited to Jlin as well as Spawn and Herndon. Those credits are important in a world rife with creative theft, unattributed sampling and the “hoovering up” of anonymous data. It’s an “ethical quagmire” that machine learning will only complicate further as it advances, says Herndon.

“I’m sure there are people with much bigger CPU capacity than me who are working on a Taylor Swift vocal model or an Aretha Franklin vocal model,” she says.

“What does that mean for [artists] if we continue to reanimate their voices to entertain us today?”

It may no longer be beyond machines to mimic the creative impulses of artists, says Herndon, but “it’s really up to us humans to give something meaning and decide whether or not something has artistic value in our society”.

“I don’t see that as something that will be so easily replaced or removed.”

In the end, everything Spawn knows has been informed by humans – specifically, Herndon and the musicians she recorded with, three of whom join her on her Australian tour.

“The live show’s really an ecstatic celebration of people on stage,” says Herndon.

Holly Herndon appears at the Magic Mirrors Spiegeltent at Sydney Festival, January 16; Mona Foma, Launceston, January 18-19; and Melbourne Recital Centre, January 21.

Most Viewed in Culture


Source link