although my voice was very unlike the soft music of their tones
December 2, 2017 10:45 AM Subscribe
Coditany of Timeness Two technologists trained a neural network to produce a black metal album.
They started with a pre-existing recording, cut it into tiny tracks, then taught the software to arrange the clips into convincing music. Here's their paper.
They started with a pre-existing recording, cut it into tiny tracks, then taught the software to arrange the clips into convincing music. Here's their paper.
This album is part of a submission to NIPS 2017 Workshop for Machine Learning, Creativity and Design: "Generating Black Metal and Math Rock".(via)
This album was generated with a recurrent neural network trained on raw audio from the album "Diotima" by Krallice. All titles were generated by a Markov chain. The album cover was created with neural style transfer.
This site may contain copyrighted material the use of which has not always been specifically authorized by the copyright owner. We are making such material available non-commercially in an effort to educate and advance research in machine learning, generative music, music information retrieval, computational creativity, etc. We believe this constitutes a ‘fair use’ of any such copyrighted material as provided for in section 107 of the US Copyright Law.
The copyright issues are interesting. To what extent is the output of a neural network trained on X a derived work of X? How large an amalgamation of works does X have to be before each is diluted to the point where copyright is extinguished?
posted by acb at 11:51 AM on December 2, 2017 [2 favorites]
The copyright issues are interesting. To what extent is the output of a neural network trained on X a derived work of X? How large an amalgamation of works does X have to be before each is diluted to the point where copyright is extinguished?
posted by acb at 11:51 AM on December 2, 2017 [2 favorites]
Be sure to also check out Calculating Calculating Infinity.
posted by kenko at 11:59 AM on December 2, 2017
posted by kenko at 11:59 AM on December 2, 2017
In all seriousness, it is as Nelson says above: Lou Reed's "Metal Machine Music" absolutely to a T. You have to laugh, somehow. Lou Reed -- damn!
posted by Modest House at 1:47 PM on December 2, 2017 [2 favorites]
posted by Modest House at 1:47 PM on December 2, 2017 [2 favorites]
Metal Metal Machine Music
posted by acb at 2:55 PM on December 2, 2017 [1 favorite]
posted by acb at 2:55 PM on December 2, 2017 [1 favorite]
I had the same thought as everyone else here. Eventually some drums do kick in and make it sound a little more black metal though.
posted by atoxyl at 5:30 PM on December 2, 2017
posted by atoxyl at 5:30 PM on December 2, 2017
They started with a pre-existing recording, cut it into tiny tracks, then taught the software to arrange the clips into convincing music.
People are still better at that.
posted by flabdablet at 6:46 PM on December 2, 2017 [3 favorites]
People are still better at that.
posted by flabdablet at 6:46 PM on December 2, 2017 [3 favorites]
Serious question. What's the difference between neural network machine learning and Markov chains?
Yes. The Rammstein clip was far far better.
posted by njohnson23 at 7:59 PM on December 2, 2017
Yes. The Rammstein clip was far far better.
posted by njohnson23 at 7:59 PM on December 2, 2017
Tagged as "deathmetal"? Ooooh...heads are gonna roll. Or, more to the point, churches gonna burn.
posted by adamgreenfield at 12:11 AM on December 3, 2017
posted by adamgreenfield at 12:11 AM on December 3, 2017
What's the difference between neural network machine learning and Markov chains?
At some 50,000 foot level they're similar but they are completely different mathematical models. Neural networks have more complicated functions embedded in them and often more complex structures. The theory is they can learn to do more phenomena more accurately and learn them faster. More reading here and here.
posted by Nelson at 7:40 AM on December 3, 2017 [2 favorites]
At some 50,000 foot level they're similar but they are completely different mathematical models. Neural networks have more complicated functions embedded in them and often more complex structures. The theory is they can learn to do more phenomena more accurately and learn them faster. More reading here and here.
posted by Nelson at 7:40 AM on December 3, 2017 [2 favorites]
« Older In 1973, the Goat faced an uncertain destiny. | How gun violence affects kids across America Newer »
This thread has been archived and is closed to new comments
posted by Nelson at 11:26 AM on December 2, 2017 [5 favorites]