patten Artist and musician Damien Roach collates AI-generated samples to map the infinite complexities of the experience of sound By Emily Bick Photography by Yu Fujiwara

The Dream Factory

Last year, DALL-E, Midjourney and Stable Diffusion first rolled out their beta versions, with their text to image software that could, when prompted, generate any visual image that could be imagined and described in a few words. Their renderings appeared like magic, conjured by speech acts, incantations. (You can also find curses in the magic – early results of human faces appeared like Francis Bacon horror swirls, and in the months since, faces have improved, but hands still often sport extra digits at unusual angles). Now Damien Roach, the artist, graphic and AV designer and cultural researcher who records as patten, has released the first album ever made from the results of a similar AI-driven text to sound generator, patched and reworked together into a suite of songs.

Mirage FM is a tantalising rush of textures and production tics: chopped and screwed R&B wooze, a familiar smooth pulse of hairdresser house cut with electronic skips and scrapes; the keening of pop divas and rappers, emoting syllables with no content, no relation to language. It’s familiar as the spirit of radio, guiding an endless cruise, mirage-like in the sense of the infinite scroll of a million grids and feeds, in the intangibility and ephemerality of images drifting through the cloud.

Roach describes the project as “crate digging in latent space”, using a record collector’s dogged sensibility to sift through all of the possibilities of the improbable geometries of latent space – the topographical model representing compressed data classified by machine learning algorithms. “I wanted to find ways to communicate the process in a way that was quite understandable and not sort of tech-y,” he says. “You could think of them as a sort of sound library of notional music that hasn’t existed before. And on top of that, I suppose there is this strange idea with latent space. Like an almost geometric way of thinking about data that has been accumulated. And within that space, there’s a mass that forms a kind of archaeology or an archive of human thought, what’s been made in the past, human creativity. And also, inside that space it’s possible to find wormholes between points that haven’t necessarily been uncovered by things people have actually made already. There are these wild conceptual possibilities within the realm of all of the data that’s been produced already.”

That particular phrase, crate digging in latent space, is also interesting because of the clash between modes of categorising and searching for information. Crate diggers have their own heuristics for flicking through the bins: they might be looking for a label, a style of design, certain players lurking in the credits – I’ve been swayed by the odd juicy typeface. It’s hard to describe the appeal of a lucky dip record: you know it when you see it. But how does that square with machine learning systems that have their own weird systems of making patterns and connections, their own unknowable heuristics? “I mean, suppose in practical terms, in both of these systems that you’re talking about, the main interface is language. Words. And it’s funny actually,” he says, gesturing at records lying around the office of The Wire, where the interview is taking place, “seeing this categorisation and again thinking about the record shop, libraries, archives and how information is organised. And the way the information is organised obviously has an impact on how it’s perceived and what it can possibly be, right? So if you categorise something in a specific way, it impacts what the thing is. It’s like alchemy, isn’t it? Linguistic alchemy, to name or to categorise is to cloak or to reveal or position something… I guess that the Venn diagram is language and words themselves. Which is interesting.

“I think that if you’re working with a system that operates in a slightly mysterious way, where your usual framework of understanding and orientation breaks down, then there’s an opportunity for failure, a failure of precision that could afford the uncovering of spaces – again to use that spatial analogy – that you weren’t necessarily intending to seek out. Like the weird potential of a broken map.” Mirage FM was created on Riffusion, an offshoot of Stable Diffusion that adapted the latter’s text to image modelling capacity to create a text to sound generating system. As a hobby project, programmers Seth Forsgren and Hayk Martiros trained the Stable Diffusion model on images of audio spectrograms of sound clips paired with text, and built a website interface where anyone can enter a string of words and generate sound samples.

Riffusion only exists because, unlike DALL-E, Midjourney

The Wire / patten

36