Showing posts with label MACHINE LEARNING. Show all posts
Showing posts with label MACHINE LEARNING. Show all posts

Tuesday, September 26, 2023

"cluttered with regurgitations"

A piece by Chal Ravens at DJ mag asks "what is the future of sampling"?

The first half  is more about the present (the craze for edits on the dance scene and bootleg tracks that rip from well-known pop songs) and even more so about the past of sampling. Indeed it's a bit of a samplescape itself, with loads of quotes (including one from Kieran!) and some fairly familiar history (stuff about the Akai MPC, court cases in the early '90s).

Gets really interesting in the second half when it looks at the artistic potential and cultural implications of AI.

Like the STEMS technology that enables producers and deejays to cleanly separate the instruments within a sampled stretch of music. (Similar AI demixing technology enabled the remixing of Beatles albums where the four-track technology of the era had meant that tracks had been bounced down, smushing together instrumental parts inseparably - until now)

Then there's software that can "generate entire songs based on text descriptions" like Riffusion . "Through text prompts, the model can find the midpoint between otherwise unrelated sounds — what’s between, say, Goa trance and Don Cherry’s trumpet? Ask Riffusion and it’ll spit out a sample that imagines exactly that, despite the impossibility of making the “real” thing.". 

Then there's deepfake raps, where a producer can generate a convincing simulacrum of a guest feature by a star artist - not quoting an existing performance and redeploying (as with jungle borrowing bits of dancehall and gangsta rap) but creating an all-new performance in another's style. 

I wondered though - do either of those really count as sampling? Aren't they more like AI-assisted pastiche, or AI-assisted impersonation, or AI-assisted identity-theft? 

The text-to-audio stuff seems like it's outsourcing the kind of zany hybridization that supergroup initiators like Bill Laswell liked to do. Or wackily eclectic genre-colliders like. I don't know, Primus, that kind of band. The creativity reduces to thinking up the goofy idea;  the craft of implementing it is left to the technology.  But as Jaron Lanier says piquantly in the piece, the struggle to realise an unlikely or adventurously preposterous idea is the point:  “Would you want robots to have sex for you so you don’t have to? I mean, what is life for?”  

Likewise the things that Holly Herndon have done with training a vocal entity by feeding it human singing seem so far removed in process, intent and outcome from sampling as audio-quote that it doesn't really make sense to think of Spawn et al in the same breath as mash-ups and collages. 

Ditto patten's deployment of " text-to-audio AI samples from Riffusion, a model based on a database of sonograms" on Mirage FM. This feels more like a convoluted form of synthesis than a citational mode in the tradition of Pop Art / Appropriation Art. A couple of steps beyond Todd Edwards's sample choir into effectively sourceless ethereality. 



^^^^^^^^^^^^^^^^^

One thing worth noting is that until recently - the rash of interpolations and quotes in pop songs, driven by publishing companies looking to exploit assets to the utmost - overt sampling had become rather rare in mainstream pop. It was particularly noticeable with hip hop during the 2000s and 2010s: the whole historical branch of it based around the deejay-as-crate-digger became a niche underground aesthetic (Dilla being the most visible exponent). Probably the only major mainstream rap auteur still basing his thing around recognisable sample quotes during this period was Kanye. Whereas trap, for instance, hardly ever featured samples. When it did - Future's "Mask Off" with the use of an obscure soundtrack song, Migos clumsily deploying "Papa Was A Rolling Stone" -  it was striking, had a pronounced throwback feel. For the most part trap - and mainstream rap generally - in the 21st Century wasn't about breaks and loops;  it  involved programming beats, synths and keybs, and other elements assembled and organized within digital audio workstations. Beat-makers might still sometimes call themselves "DJ", but what they did had little relation to the deejay skill set. 

Thursday, June 29, 2023

retrotalk2023: recreative garbage

Substack called Anglophone Xenotrope and person who goes as Thunder At Twilight riffs on the "conjunction of futuristic technology and retro" thing I mentioned in the last post. (Strictly speaking retrotalk2022 as the post - titled "The End of Culture" - is from the very end of last year). 

"Retreads, reboots and remakes were ascendant in the first decade of this new century but the rise of the Marvel Cinematic Universe and the return of the Star Wars franchise—combined with the digital       de-aging technology, pioneered by Industrial Light & Magic—added considerable fuel to the nostalgia fire.

"When Disney bought Lucasfilm and ILM, they were not just buying Star Wars; the work that Lucas had done in bringing dead actors back to the silver screen in the form of digital composites was arguably even more significant. To that end, the corporate giant has made integrating these creations into existing brands a core part of their corporate strategy.

Examples include: 

"Rogue One: A Star Wars Story, featuring actors wearing digital masks of the deceased Peter Cushing and the (then living) young Carrie Fisher"

and Marvel's use of 

"the technology to provide a curious kind of continuity, allowing much older actors to play younger incarnations of their characters in earlier eras."

and most recently (at the time of writing):

"the trailer for the new Indiana Jones film featuring a very respectable depiction of a much younger Harrison Ford."


Twilight at Thunder speculates:

"Seeing cutting edge digital rendering technology used as a means of rehashing and remaking the past leads one to reflect on possibly the ultimate manifestation of the mash-up: AI systems such as GPT and Stable Diffusion. Soon we might be able to look forward to machine created films, featuring digital recreations of characters from existing franchises.

".... AI models cannot create, even if their outputs are novel. By definition, these systems are synthesising based on statistical inferences."