SoylentNews is people
SoylentNews
from the how-could-they-not-call-it-Botty-McBotBand? dept.
In less than two years, the band Dadabots has ten albums to its name and has been streaming live 24x7 for a month without stop. Death Metal bands are weird. But this one is different.
Dadabots has no humans in it.
Dadabots is an AI that creates technical Death Metal genre music, and does it well enough that its creators have now decided that the algorithmically generated music is of sufficient quality that it can be streamed in real time at creation. Give it a listen here. Love it or hate it, it's true to the genre and has been streaming non-stop since last month.
Dadabots is the brainchild of CJ Carr and Zack Zukowski, a couple of musicians with a strong interest in algorithmically-generated music. For a couple of years now the duo has been working on developing a recurrent neural network that can produce original compositions after being trained on specific datasets from singular musical genres. Early experiments incorporated a variety of different genres, before the duo discovered metal and punk in particular seemed to be better suited to the erratic and often random nature of the algorithm.
"We observed that electronic music and hip-hop instrumentals did not seem to train as well as organic, electro-acoustic ensembles of musicians," Carr and Zukowski write in their most recent paper. "Music genres like metal and punk seem to work better, perhaps because the strange artifacts of neural synthesis (noise, chaos, grotesque mutations of voice) are aesthetically pleasing in these styles. Furthermore, their fast tempos and creative use of loose performance techniques translate well to SampleRNN's rhythmic distortions."
Dadabot's free albums are available here. The album covers and song titles were also, naturally, algorithmically created.
(Score: 4, Interesting) by RamiK on Friday April 26 2019, @09:30AM
The (current?) live stream [youtube.com]'s title "RELENTLESS DOPPELGANGER" is a reference to Archspire's Involuntary Doppelgänger [youtube.com] which, as mentioned in the description, they trained on.
Listening to it for a few minutes, some of the breakdowns are actually better than Archspire's.
Cool stuff.
compiling...
(Score: 4, Informative) by GreatAuntAnesthesia on Friday April 26 2019, @10:09AM (2 children)
https://soylentnews.org/comments.pl?noupdate=1&sid=7154&cid=178508#commentwrap [soylentnews.org]
(Score: 1, Informative) by Anonymous Coward on Friday April 26 2019, @10:38AM (1 child)
https://soylentnews.org/comments.pl?noupdate=1&sid=11298&cid=280561 [soylentnews.org]
https://soylentnews.org/comments.pl?noupdate=1&sid=21868&page=1&cid=576596 [soylentnews.org]
https://soylentnews.org/comments.pl?noupdate=1&sid=24774&page=1&cid=658997 [soylentnews.org]
https://soylentnews.org/breakingnews/comments.pl?noupdate=1&sid=23948&page=1&cid=633814 [soylentnews.org]
(Score: 2) by c0lo on Friday April 26 2019, @10:42AM
Life still has quite a bit to catch up, heh?
https://www.youtube.com/watch?v=aoFiw2jMy-0
(Score: 5, Touché) by pkrasimirov on Friday April 26 2019, @11:18AM
> Dadabots has no humans in it.
Finally a true Death Metal!
(Score: 1, Insightful) by Anonymous Coward on Friday April 26 2019, @12:13PM (2 children)
who do great stuff (just randomly browse SoundCloud) , we don’t need generated music. On the other hand it is doing what many musicians also do; steal a riff here, a melody there, and re-arrange it.
(Score: 4, Interesting) by RamiK on Friday April 26 2019, @03:02PM
Scales... Modes... Chords... Cadences... Rhythm patterns... Counter-points... Scale-degrees... Grammars... Composed music is generated music. Sometimes by a meatbag through memorization and practice. Sometimes by silicon. It's all algorithms and a little randomness.
Besides, there's more agreeable usages for software as training aid for musicians. Personally I've wrote a dumb (pure randomness, not even markov chains) music sheet generator script for sight reading practice since I run out of material I didn't inadvertently memorized as I'm learning the fretboard. But for real musicians you have stuff like https://github.com/Impro-Visor/ [github.com] to learn and practice jazz improvisation that has an AI that processes sheets and produces a grammar that it then goes on to train you in: https://www.youtube.com/watch?v=YOOjKNHS9A8 [youtube.com]
compiling...
(Score: 2) by DeathMonkey on Friday April 26 2019, @06:02PM
Hey look, it's Technology's Dan from American Gods!
(Score: 0) by Anonymous Coward on Friday April 26 2019, @12:47PM (1 child)
Will it burn the bitchurch?
(Score: 2) by Rich on Friday April 26 2019, @03:30PM
Only with a network that has been trained on black metal.
Which they already did, btw... cf. https://nips2017creativity.github.io/doc/Generating_Black_Metal.pdf [github.io].
As a future development, they could train one that out-Max-Martins Max Martin, in turn removes any value from mainstream music and breaks the system.
(Score: 2) by inertnet on Friday April 26 2019, @02:48PM
Silicon isn't really a metal. But it is dead.
(Score: 1) by Rupert Pupnick on Friday April 26 2019, @04:49PM (4 children)
It does really capture the feel of the genre, and it’s particularly impressive that it can deliver output in real time, but as a compositional engine, well, it’s completely atonal, and so misses out on a lot of potential musical complexity and beauty.
Yeah, I get that it’s death metal, but there’s been plenty of AI composition that’s been done already, scoring entire symphonies in some cases, that’s been around for at least a decade.
(Score: 2) by RamiK on Friday April 26 2019, @09:51PM (2 children)
Tonality is far more flexible in metal than other genres and you can find a lot of dissonant and 12tones pieces floating around: https://www.youtube.com/watch?v=r5eggJmOeJA [youtube.com] https://www.youtube.com/watch?v=gKkn-YjYmCA [youtube.com] https://www.youtube.com/watch?v=IVyUHFl0iB8 [youtube.com]
What's lacking is structure. e.g. Involuntary Doppelgänger is built on a modified Arch (ABCBA) form* with very distinguishable sections, courses, melodic lines and a lot of repetitive licks in-between serving to color the the chord transitions like turnarounds do in blues. While not all of those are required and indeed many of album's songs lack in a few, you at least need some sort of repetition like how Blotted Science does it.
Of course, that's only if you consider it to be a song. If you consider it as a selection of melodies for you to taste, digest, and think on how to fit into a real song, than it's some very useful stuff!
*yeah I know who could have guessed... :D
compiling...
(Score: 1) by Rupert Pupnick on Friday April 26 2019, @11:48PM (1 child)
You’re right, there’s a lot of atonal music out there, and not all of it is death metal by a long shot.
And you put your finger on it better than I did. Not only is dada atonal, it doesn’t seem to have a notion of tempo, nor are there identifiable repeating sections. No structure.
So to me it seems less of an accomplishment compared to David Cope’s who had designed software years ago (called EMI or “Emmy”) that generates these sprawling but relatively coherent compositions, and in the style of whatever composer it “trains” on. Check him out if you’re not familiar. Here’s EMI doing Vivaldi: https://m.youtube.com/watch?v=2kuY3BrmTfQ [youtube.com]
(Score: 2) by RamiK on Saturday April 27 2019, @10:38AM
I believe EMI is more procedurally generated and trained on midi samples (notation effectively) and much of the structure and grammar was manually coded (e.g. after so and so bars, inverse chords of the following degrees and add a counter point to this and that beats) while these guys are sampling the audio itself and letting the AI figure it all out through pure training.
Accomplishment wise, it depends on your goals. If you want to have an AI listen to a recording and start composing similar works or even back a live stream with percussion, bass or even a few melodies, you'd probably want a combination of the two approaches or alternatively train a huge neural net to effectively reconstruct all our music theory knowledge from scratch.
Currently Cope already won the "race" to generated music since his output is already sounds as good as any human's composition. But going forwards, there some huge neural-nets on the way ( http://www.tachyum.com/ [tachyum.com] ) that will make it clear which approach is easier under a decade.
compiling...
(Score: 0) by Anonymous Coward on Saturday April 27 2019, @01:55AM
Thanks, interesting for a few minutes.
Now, back to Sun Ra...