As the singularity approaches, neural network pens black metal album
Reeeyeeeeeese... of tha maaassssshyyyyyyunaahhhhh
RotM If Coditany of Timeness was released without the high-tech fanfare, no one in the notoriously elitist black metal scene would bat an eyelid. Perhaps popular online US music mag Pitchfork would even give it a "6/10".
The album has everything intrinsic to the sub-genre – tremolo guitars, blasting drums, barked vocals, and total disregard for lamestream songwriting conventions – but humans aren't responsible for this cacophony. At least not directly.
Because the record, as first revealed (here) by The Outline, was written by a computer. More specifically, deep learning software that regurgitates an emulation of whatever music it's fed. In this case, it's the avant-garde black metal act Krallice's 2011 album Diotima.
The project, Dadabots, is led by machine-learning boffin Christopher James Carr and music producer Zack Zukowski. The pair are due to present their paper – Generating Black Metal and Math Rock: Beyond Bach, Beethoven, and Beatles (PDF) – to the Neural Information Processing Systems conference in Long Beach, California, this week.
The abstract describes their work thus: "We use a modified SampleRNN architecture to generate music in modern genres such as black metal and math rock. Unlike MIDI and symbolic models, SampleRNN generates raw audio in the time domain. This requirement becomes increasingly important in modern music styles where timbre and space are used compositionally. Long developmental compositions with rapid transitions between sections are possible by increasing the depth of the network beyond the number used for speech datasets. We are delighted by the unique characteristic artifacts of neural synthesis."
The software cut Diotima into audio chunks and fed it through a neural network, which was then asked to guess what comes next. "Correct" suggestions would strengthen and prioritise that particular pathway to eventually make something resembling the source material.
But with such an aggressive origin, it wasn't all plain sailing. "We trained using 5-layer LSTM and GRU models," the paper stated. "The GRUs failed to learn the audio resulting in harsh noise when sampled. The LSTMs were successful in training and sounded like Krallice. We generated 20 sequences with four-minute durations."
The music wasn't the only thing generated by machine. A Markhov chain named the album and each track therein – including such gems as "Timension" and "Wisdom Trippin'" – and the cover art was "created with neural style transfer".
Coditany of Timeness is by no means unlistenable, especially if you're a hardened tr00 kvlt veteran like your correspondent here. It does sound like it was written by a computer but had that not been revealed, it could pass for something bashed out in an edgelord's bedroom. It also sounds like a drunk Krallice, of course.
Krallice isn't the only left-field victim of Dadabot's hack 'n' mash either. The Dillinger Escape Plan's seminal "math metal" debut, Calculating Infinity, was also given the treatment and somehow sounds even more incoherent.
Based on that, AI pop stars are still a fair way off. ®
Sponsored: Beyond the Data Frontier