Gigantic digital library Project Gutenberg needs to step aside for a second. Welcome Project Gucciberg, a humorous spin on the former’s name. The work of viral content creator MSCHF, Project Gucciberg "features" trap legend Gucci Mane's voice reading classical literature, including Franz Kafka's Metamorphosis, thanks to the capabilities of machine learning.
It’s especially amusing that MSCHF chose Mane's voice for the project given it’s often hard-to-decipher, but then peculiarities make up a good deal of MSCHF’s content. In the past, the internet collective incentivized bashing Big Tech by offering thousands of dollars to TikTok users who had to use MSCHF’s pre-made sounds in their public videos. Its targets at the time were Palantir, Facebook, Amazon, among others.
Another MSCHF project launched "Birkenstock" sandals from leather sourced from Hermès Birkin bags which cost a fortune. And most recently, it used a Spot robot dog from Boston Dynamics with a paintball gun attached to it to let fans remotely redecorate an art gallery. So perhaps we shouldn't be too surprised by this latest project.
Text-to-speech-to-Gregor-Samsa — With Project Gucciberg, MSCHF is using Mane's voice to read Alice's Adventures In Wonderland, Pride And Prejudice, Little Women, A Modest Proposal, Frankenstein, Dracula, War of The Worlds, Anna Karenina, Anthem, Heart of Darkness, Beowulf, Jekyll And Hyde, The Adventures of Sherlock Holmes, The Prince, The Jungle, The Importance of Being Earnest, and other classic literary works.
The artificial intelligence version of Mane wades through this literature with ease and grace, even if it’s hard to hear him sometimes. MSCHF’s Dan Greenberg told The Verge the collective studied Mane's podcasts, interviews, and other audio clips that turned into a set of sample data covering at least six hours. That's six hours of Mane's style of speaking, argot, verbal idiosyncrasies, and the like.
This data was then turned into transcripts, which helped with the text-to-speech processing for machine learning. For example, the MSCHF model was taught to understand when Mane says "talm 'bout," he means "talking about.” The quality of the resultant clips isn't crystal clear — there are points where Mane's voice is muffled or the microphone sounds muffled. But it's realistic. A little too realistic.
Forgiveness not permission — Given how believable this audio deepfake sounds, we have to wonder about its ethical implications. According to The Verge, nobody asked for Mane's permission to use his voice. But does MSCHF even need his consent? It's hard to say whether these AI fakes constitute copyright infringement since they are not being used to scam people (at least to our knowledge).
But it’s certainly unsettling to know that recordings of your voice can be used to make convincing subsequent recordings of you saying things you didn’t. Like, perhaps, “One morning, upon awakening from agitated dreams, Gregor Samsa found himself, in his bed, transformed into a monstrous vermin... a situation he thought was anything but Gucci.”