Holly Herndon is not interested in creating her own replacement. The composer-musician saw the rise of machine intelligence in music as major labels and tech companies, hungry for cheap production, began pushing AI songwriters. But rather than fight it, Herndon decided to raise a robot bandmate herself.
Herndon, who completed her Ph.D. at Stanford’s Center for Computer Research in Music and Acoustics, trained an artificial neural network built into a “DIY souped-up” gaming computer to find its voice from scratch and sing with her. For two years, Herndon nourished the AI baby, known semi-affectionately as Spawn, with her own vocals, composing, cooking, and living alongside it. Instead of sampling directly from her voice, Spawn created its own sounds using rules it could discern from the patterns it heard. The results are spliced throughout Herndon’s recent album PROTO, on which Spawn stumbles into uncanny beauty — an alien child moaning and wailing in harmony with its mother.
The choice to include the skittering roughness of some of Spawn’s vocal contributions wasn’t rooted in low-fi fetishism, but in transparency about the nascent state of musical neural networks. For Herndon, current A.I. compositions tend to sound stuck in an “aesthetic cul de sac” because of relatively unsophisticated programs that prioritize a clean sound over uniqueness.
“So much of the AI music that we’re presented with is so glossy and perfect, and it’s really not there yet,” Herndon says. “But right now it feels a bit like smoke and mirrors.” For Herndon, AI compositions tend to sound shiny but sparkless. Instead of making hollow copies of someone else’s style, Spawn’s ungainly and imperfect voice is its own.
That transparency also extends to Herndon’s interrogation of the legal and ethical issues raised by algorithmically-generated music. Are royalties owed if A.I. is trained on the sounds of a singer like Holly — or, say, the late Aretha Franklin? Even as A.I. composers rapidly improve at emulating individual artists, copyright law seems not yet prepared to tackle this question.
But by making explicit the human labor that went into Spawn, Herndon hopes to help set the norms for A.I. music and prevent uncredited mooching off others’ music. “I really wanted to make audible the people who went into the training data,” she says. “I’m concerned for attribution, for people’s intellectual labor just to be hoovered up as soon as it is machine-legible and spit out by whoever has access to the most powerful model.”