An AI created a death metal album so even music isn’t safe from robots

Impact

Rife with often-unintelligible vocals and brutal instrument solos that sound like the guitarist’s fingers have melted off onto the fretboard, death metal can sometimes make you wonder how human musicians even pull it off. Well, in this case, they didn’t. Back in September 2018, music fans and technologists Zack Zukowski and CJ Carr used a deep learning algorithm, aka a form of AI, to create a death metal album with stunningly realistic compositions. Now, that same algorithm is streaming computer-generated metal music daily on YouTube, as TheNextWeb reported on April 22.

According to TheNextWeb, Zukowski and Carr created a recurrent neural network called SampleRNN to spit out an AI’s approximation of what an album pulled straight from the death metal genre might sound like. The result? A record called Puzzlomaly made by the fictional band Dadabots, featuring disturbingly intense rhythms and sometimes unclear vocals but still feeling as if it could’ve been made by real metal artists.

To create Puzzlomaly, the machine learning algorithm listened to the band Psyopus’ 2004 album Ideas of Reference 29 times over the course of several days, and generated over 435 minutes of audio from its “learning” session. Afterward, a human editor listened to the machine-generated audio, took various clippings from the results, and formulated it into a 10-minute album split into three tracks. Each song off the album is titled based on a mathematical system, which is why the titles (”MirrorrorriM,” “Imogen’s Puzzle e29 i79751,” and ”I.................”) don’t make a lot of sense to humans – even if they do to robots.

Puzzlomaly is not the first album Zukowski and Carr have created through AI. Dadabots has released over 10 records via the music streaming site Bandcamp, each utilizing an AI that learns from the type of content it’s fed – including tunes from bands like The Beatles, NOFX, and Atlas. None of the songs off Puzzlomaly, for instance, are perfect; the words are meaningless syllables generated from vocals fed into the AI, and the sound is far heavier than most metal. But the songs are still strikingly similar in terms of their musical style, to an extent that a casual metal fan might assume they came from a real artist.

If you want to listen to more death metal creations from Zukowski and Carr, you can go to their Youtube channel, which launched on March 24 and features a livestream of nonstop AI-generated death metal music that was trained on tracks from the band Archspire. There’s also the album the men co-released with Portuguese black metal band Cavemaster back in 2018 called Nun de la parte Del, mas Deilha (la Maquina) (“Not from him, but from her (the machine.”). As noted on the album’s Bandcamp page, Cavemaster was one of the first bands to utilize neural net-generated raw audio trained on their own music to make the album — basically, they used the same script that powers Dadabots to take recordings of their own music and had the AI “listen to it” and create songs based on those recordings. They they chopped up and edited down the results, making it into an album.

Carr and Zukowski discussed their goals with these AI projects in a 2017 paper, noting that they were originally simply aiming to achieve a “realistic recreation of the original data.” However they were surprised to find “aesthetic merit” in the things that made the music “imperfect,” like the intense tempo and nonsense lyrics. Flawed or not, it’s surreal to think that an AI can create albums that sound so close to those of real musicians. With this technology, plus AI singer-songwriters like Yona making waves, it looks like we’re closer to an AI-filled pop landscape than we realize.