Most prompt-based generative AI systems are trained using some form of input and output pairs – an image and the text description of the image or the beginning of a sentence and its completion. What is being encoded is a relationship between text and output – with millions of examples, this relationship becomes nuanced and able to produce novel outputs in the desired style.
In the case of the still image (the first domain to be fully appropriated by the machine), scant concern was paid to the wholesale ingestion and regurgitation of our greatest artists. It was early days, people didn’t know how it all worked, and besides: there aren’t many lawyers out there defending the IP of long-dead painters.
Music however is different. Major record labels are famously litigious and fiercely defensive of their artists’ IP (not necessarily always for the benefit of the artists themselves). As a result, generative music systems have tended to avoid explicitly referencing the artists which they have ingested. Asking suno or udio for ‘an Elvis Presley song about squirrels’ will get you quickly denied. As a result, those working in the nascent field of ‘AI music’ have had to use careful circumspect language to fool the machine into giving them what they want.
Not so with sonauto: it appears that this generation system has thrown caution to the wind, allowing for direct reference to well known artists and producing remarkable facsimiles of both music and vocals. I can sense a horde of salivating lawyers just across the horizon…
By default, sonauto will generate some lyrics for the artist to sing – most of which are hilariously bad. However, if one prompts the machine to generate an instrumental song in the style of a singer, it frequently hallucinates the singing in a gloriously aphasic manner. It won’t do this all the time, but when it does it is delightful. Phonetically plausible nonsense, turning everyone into the Cocteau Twins.
Some examples for your listening pleasure
Elvis Presley
The Beatles
Talking Heads
Joni Mitchell
Leonard Cohen
Kate Bush
Bob Dylan
James Brown
The obvious historical parallel is this piece of ‘English-sounding’ song by Adriano Celantano from 1972 – not actually English, but close enough for the Italian audience of the time.
I find these kinds of AI edge-cases fascinating. Our AI friends can do a remarkable job of emulating our text, our art and our music, but every now and then the facade falls and their fundamental nature slips out. Whether it be AI-generated photographs of people with extra fingers or the adding of non-existent references into your chatGPT essay, these hallucinations allow us to glimpse just how fundamentally different these machines are from us, however good their impressions.
In the case of music, it will be fascinating to see where this goes. Will people be happy with an infinite supply of generated music in the general style of their favourite artist? And perhaps more pertinently, in a world of infinite AI music, from whence does the new Elvis emerge? Will the superstars of tomorrow be entirely AI generated, or will we be living in the end-stage feedback loop of Post-Modernity and only listen to facsimiles of the past?