Over the last few decades, some truly great black metal albums have been released. Mayhem’s seminal Deathcrush, Bathory’s self-titled 1984 release, and even Emperor’s Anthems To The Welkin At Dusk. But soon, that list may be joined by an album created completely by a neural network.
Neural net creators Christopher James Carr and Zack Zukowski presented their work at the Neural Information Processing Systems conference in Long Beach, California, this past week. There they discussed how their unique neural network, which they’ve dubbed Databots, was able to create a five-song EP of black metal music. The EP has since been released under the title Coditany of Timeness, and can be found via a Bandcamp page.
Through the release of completely computer-created music is something to get excited about, the neural network first had to have something to work from. To achieve this, Carr and Zukowski first, piece by piece, fed the neural net an album by the band Krallice. By splicing the original Krallice songs into small segments, the researchers were able to make the neural net guess what sounds might come next. If the neural net was able to correctly (or as close as possible) fill in the blanks, it was told it was right. If it created a wildly different sound, it was told it was wrong. By continuously repeating this process, the neural net was eventually trained in the musical styling of Krallice, and was then prompted to create its own music.
To build on this, even the album art and song titles were created by a computer. Title tracks were generated via a Markhov chain (which came up with gems such as Wisdom Trippin), and the cover was the result of neural style transfer. And you know what? The music isn’t all that bad.
Images: Wikimedia, Krallice, Databots
Source: The Register