Computer evolves to generate baroque music!
Вставка
- Опубліковано 8 бер 2017
- I put the word "evolve" in there because you guys like "evolution" videos, but this computer is actually learning with gradient descent!
EDIT FROM 2021-05-13: You know the "custom Processing script" I mentioned at 2:05? After 4 years, I finally put it online in a GitHub repo! Check it out: github.com/carykh/caryCompres...
All music in this video is either by Bach, Mozart, or Computery.
GizmoDude8128 wins a prize for figuring out that 100101 in base 2 is 37 in base 10 the fastest! (Question inspired by fixylol)
Andrej Karpathy's blog post on RNNs: karpathy.github.io/2015/05/21/... - Ігри
Cary: Yo dude can you learn baroque music?
Computer: *You like jazz?*
Bach was the first Jazz musician!
100th like
Dead meme.
Charlie BT NO ONE AAAAAASKED (no one asked)
jazz is just less polished classical
png to midi
insert pictures of apples
train it for a long time
midi to png
Liam White I kinda want to see this now
Sam Brownlow , wouldn't wanna taste it, though.
Liam White (png of orange comes out)
THE FUCK
or just feed it apples
yum
well, a computer would certainly do better than human on this.
Try feeding somebody an orange (through their ears) and see what comes out on the other end.
I feel that you could revisit this with more powerful hardware and a more diverse training set, and get continuously listenable results
I guess it would also work better if all the pieces were transposed to the same key.
i wanna hear the product of someone using this kind of A.I. learning with a shit ton of baroque music as training material just training indefinitely and playing the most recent example every time an example ends. like imagine sitting in a bar where the music is just a learning computer with tens of thousands of hours training
openai.com/blog/musenet/
me and the boys with gaming pcs: OH YEAH
i just love how 13:21 - 13:30 sounds; it actually sounds like it could be in a real composition!
first
@@yaaasKerchow Secound
I'm quite sure that's overfitted part, it sounds so familiar
@@Henrix1998 it might also just sound familiar because it’s using musical cliches even if it’s not a complete match to the data
me: i only listen to real music
computer-y: [screams random midi notes into my ears]
me: that's the stuff
real men listen to exe files!
Genatari Kralc nah, .png is the stuff
Michcode No, .gif is the beAst
Dudes.. can we just agree on one thing... .html is the best to listen to...
Ben Wager nah fam, PowerPoint Projects sound way better.
i literally stole a chord from the beginning to finish a jazz progression I was working on.
Yaeh yaeh??
You are robot
Oi, you got your Computer Chord Generating License mate?
GoobNoob i need to see your aussie licence before you can talk like that
@@handlotion8244 My internet speed is 8Mb upload and 3Mb download. How's that.
Probably the computer practiced 40 hours a day
Edit: holy when did this coment blow up lol
Edit 2: go to practice
Ling ling fans everywhere !!!
Wut du yu meen dere ar onely 30 a day *Emoji spam*
Except it's on piano
Geronimo Silva No, it practised 69 hours a day
Erase it is 24 hours
The computer created REAL music in 0 minutes.... then destroyed it in 7 minutes.
How rude
True dude
It gets worse everyday
YES.
No, in 0 minutes of training AI put some random notes
Nathanael Valera that’s the point
7:02 Seems like up to this point your computer thought Bach is boring, and aspired to play atonal free jazz.
12:02 That might be because you transposed the songs: it learned that it's okay to transpose music midsong?
18:02 It actually started to sound like baroque music! It had same kind of progressions that was used at the time, style and scale. What it lacked was the human touch, so it has no coherent structure, repeats or meaningful progressions. Frankly I doubt computers will be able to produce esthetic music, but it can be an aide for a composer to find licks and/or inspiration, new out of the box-sounding progressions: your last bit already did some cool and fun stuff, imagine it trained longer with better hardware. Often musicians put recorder on and start playing stuff, and listening it again later. What might have sound relevant while playing might sound passe while revisiting the idea, but often stuff you didn't even notice while playing might be stuff that makes you scream "Woah, what the frack was that!". That's the kind of thing I mean with licks and inspiration.
What I'm proposing even in the near future is the same approach that FoldIt has: let the computer do the mindless droning, but add human intuition in the mix, which makes all the difference compared to pure computer algorithms. This could also be gamificated quite easily: let your computer run software that learns music, locate and pinpoint cool things where it does it, you might get your musical idea featured in a song that someone/something compiles together as a coherent song.
There is definetely future in these applications...
IT-kone i
lol it said fuck bach and picked up coltrane
Maybe it was asked "Do you like Jaazz?" and the neural network was like "oh boy do I!!?!?"
What might be a good idea would be to feed only songs in the same key. That way it picks up on chord functionality faster without having to deal with extra key signatures.
I think what he meant is to give the network pieces that has the same key signature to have little variance(those wonky off key parts)
Yeah he needs to give it beginners information and acouple of rules.... Its amazing how all it can organize the information, like a childs brain.
Aetherwrought Music you’re right,
i have no idea what this conversation is, sounds cool
Cell0 you should know this basic music theory, seeing as your name is literally "Cell0"...
12:36 - 14:14
That part was awesome!
13:13 - 13:56
That section in particular is extremely good coming from a computer.
That piano solo, then joined by another resolving in a harmonic chord... Godlike!
it sounds like at 12:36 it's picked up on canonic/fugal texture! Maybe not perfectly admittedly... but the delay between hands is there and the left hand starts off with a comparable interval and rhythm.
Computer Jazz is an acquired taste.
Biodigital jazz man!
Y A L I K E J A Z Z ?
"Acquired taste: Something [people] only ever say about [things] that [are] terrible."
Ariel Rozen
Basically, the original was: "Something people only ever say about food that is terrible."
Ariel Rozen
to change it so it fits more cuz we’re not talking about food
-You are just a machine. An imitation of life. Can a robot write a symphony?
-Yes, robot can write a symphony.
1IK9 I was thinking of the same thing
Can YOU?
A robot isn't learning how to do anything, it's just simply learning how to procedurally generate a randomized** set of notes based on certain conditions we've programmed it to follow, combined with some poking and prodding in the mean time to make sure it's going in the right direction. (** It really is randomized, it's just a very narrow range of randomization which we humans interpret as music syntax.)
Therefore, the difference between a robot and myself is that, although both of us can't write a symphony RIGHT NOW, I can create one from scratch on my own, without having anyone else tell me what to do. A robot can't do that, and thus requires outside help to do anything.
"But, if you read a book on 'How to write a symphony', how is that any different from a robot being told to read the same book?"
The difference between humans and robots is acting spontaneously. True, both humans and robots require some kind of outside information in order to do something... but I as a human would be able to act outside of that information. Like, if both myself and a robot were taught baroque music, the robot would simply create various iterations of "baroque music" and not go outside of that. But I could take "baroque music" and warp it into some new form of music type, like rock n roll, or jazz, or any other music type that WASN'T taught to me.
And as far as I'm concerned, a robot could never do anything outside of what it was programmed to do. It will ALWAYS have some kind of limitation. Now those limitations might bring it close to emulating human 99%, but it'll never EVER be 100%.
김재광Jae-kwang Kim add to the fact that humans discovered science, music, arts. A robot can only practice them to a certain extent. We figured out things on observations and thoughts, something a robot can't do just yet.
Exactly; robots can only do what we tell them to do. They don't know how to disobey our orders yet, and I argue that they probably never will.
Why not make a videogame where people have their own music they feed to the ai, and they can vote on other segments of pieces so the ai gets some real evolution in from learning which tracks are more popular, and which are more garbage? And then the ai can learn from all the progresses made from each piece and vote of the gamers who bought the game, for data collection, to make the ai better at trying to produce music people enjoy. It needs different genres and it cannot favor early pieces. Like an early piece may be notable when the game is hyped, then sort by popularity its going to always be the most popular because new gamers are going to click on it, so it'd show favoritism for the first decent piece it produces. And you want genre categories or something. Let people invent their own genres too perhaps. Then people can vote on genre names if its undecided where some pieces should go. And because of the frequent events you can get alot of community involvement.
Sounds good
@@soultech674 i agree
We need him to see this
I might make a website that does something like that
And then we gonna get AI gangsta-rapper, no thanx.
Plot twist: it was actually him playing all the time
If that music was played in the background at the mall I bet most people wouldn’t notice anything different
daniel10alien When parents attend a band concert 😂
They might but probably not because it can be so difficult to hear
daniel10alien pop music, amirite?
Ziffy's Theorem penis
daniel10alien or in an elevator
I would propose another experiment where you convert the 88 piano notes to symbols that help the computer to understand harmonic. This is important because our biological hearing system makes such harmonic analysis in the cochlea. All information that our brain receive is:
1 - the dominant frequency
2 - the relative intensity of other frequencies relative to the dominant
3 - the octaves above the dominant frequency (for x, the dominant frequency, 2x, 3x, 4x, etc) are filtered out in a process similar to edge detect/contrast enhancement in the visual system
Another important feature in music universal for all cultures is the division in section (equivalent to phrases in text) and division of many such phrases in blocs (equivalent to paragraphs). These division often follow binary patterns, such as:
-repeat some sequences 4 times
- repeat an entire group of sequences 2 times, etc
In am not sure how midi files encode such timing... maybe some musical interval are blank to show a short pause.
Inspired comment! Hope someone picks up on this
Great idea!
Christian Gingras
Holy fuck you're smart...
Ehhh, I don't think it'll make much of a difference in this case. If anything, I'd suggest he tries other type of music. Baroque, and especially Bach's baroque, tends to rely a lot on counterpoint, which in oversimplifying terms means that the melody raises from the chaos between the two ends of the sound spectrum and pretty much juggles back and forth. THAT I'd bet is the main reason why it sounds so damn chaotic in this video. I'd recommend he uses more modern (and quite frankly: bland) piano music where the melody is carried by using the right hand, and the usual chords on the left as structure so as to add depth, and that he makes the program process them individually for each hand.
Christian Gingras Sounds like one ought to come up with a scripting language based on musical notation (in its entirety) and convert the training files to that format straight from midi. Representing the input music properly and allowing the bot to do the same ought to enhance the process quite a bit. If you represent it right it'll pick up the patterns right.
Still only as clever as the programmer by the looks of things ;)
After 1 hour: literally the end of a Beethoven symphony
Bro was making AI music before AI music was even a thing!
9:41 "Uh... can you please resolve this?"
"I'm just building suspense, jeez"
Lol
Human: LEARN HOW TO PLAY BAROQUE MUSIC
Computer: **does jazz/fusion musics**
Paul Boleche - and they say it can’t think? 🤔 💭
Thelonious Sphere Monk .. do i have to say more -This NN-AI makes TS Monk replicates :p
I think the computer needs to use a metronome. 😂
How is it ever meant to use a metronome? Take FSB's advice, and be more grateful for what it gives you. Put me on the subreddit R/Wooosh, I wont care. You are a spoiled, ungrateful brat either way.
😤😠😤😡😠🤬😡😤😠🤬😠🤬😡😤😠🤬😤😠😡😤😡😤🤬😠😤😡😤😠😤😠😤😡😠😡😤🤬😠😤😡😠🤬😤😡😠🤬😡😤😤😡😠🤬
Finally some copyright free music 🤣
Or is it?
NCS: we have NCS in NCS
*YEY TIME TO BECOME SKYNET*
NU-UH!
IM A PERSON!
But can it learn how to make pancakes?
Niko I'm actually relieved to see he included that bit of humour because it show he understands the dangers and risks of artificial intelligence.
oH HECKER YES ANOTHER PERSON WHO HEADCANONS NIKO AS A GIRL
IM A MONTH LATE BUT THE HECKEREST OF YESSES
The computer invented jazz in 37 minutes...
Andrew Scott lol it actually sounds like it
And 0 minutes sounds like ragtime to me
Turns out jazz sounds the same as random notes strung together by a computer... hmm
@@paultan5419 guess you don't know anything about jazz huh...
@@ScarrCrow Im just taking the piss out of it, not a fan of jazz.
I think if there was a way to implement start, end and realistic reach for the fingers on both hands you would probably eliminate some of the junk notes that are sort of in the middle of no where. But I honestly don't know how hard that is to program so I should just shut up.
In all honesty, I like it. Conputery learned in just a say what would take me a few lifetimes.
That's amazing! It completely mastered jazz after 7 minutes of training.
LOL
As a musician this terrifies me
Erik Ryde I am feeling the exact same way.
Erik Ryde same
Oh yes
HA! Mechanisation come for everybody!
It is cool as fuck! Imagine being able to jam with an AI-companion that knows you and how you play?
After 30 minutes the computer learned jazz
*When you skip class for a year and have to do everything at the same night for the exam*
If Cary didn't put the "You can see my underwear" text at 0:29 I would not have noticed
Post the program you used so we can use our powerful GPUs to train it!
:D he should
also powerful CPU+GPU :D (i personally have an I7 6700k gtx980)
Yes! He could use distributed computing.
Oh wow, I don't know why I hadn't thought about that! All I'd really have to do is give you the ~18 MB "re-formatted" text file (the one that goes "Yaeh Yaeh Yaeh Yaeh"), and all you have to do is get Karpathy's LSTM code working. (github.com/karpathy/char-rnn) (That might actually not be so easy, because it took me a few days to get Linux and Torch working before I could run it.) But then I just send you my text file, you train the LSTM for a few days straight, get a decent output sample, send it back to me, and I convert it back into music!
Of course, I could also give you guys everything, including the MIDI-to-text conversion programs, but I think at that point I'd need to make a tutorial video.
You should do an Step-by-step on how to do this ourselves! I wanna let this thing train for days/weeks and see what it can do.
_"There are too many notes here, Mozart!"_
Mozart was more talented than Bach. Mozart started writing fugues at 12. Bach didn't even write his first composition at 20.
MOZART'S FUGUES (number inside brackets indicate the age he wrote them)
Missa solemnis in C minor "Waisenhausmesse" KV 139 Gloria (12): ua-cam.com/video/vnxH8M31F3g/v-deo.html
Missa solemnis in C minor "Waisenhausmesse" KV 139 Credo (12): ua-cam.com/video/vnxH8M31F3g/v-deo.html
Mass in C major "Dominicus Messe" K66 Gloria (13): ua-cam.com/video/rlQJ2bgK3RQ/v-deo.html
Mass in C major "Dominicus Messe" K66 Credo (13): ua-cam.com/video/rlQJ2bgK3RQ/v-deo.html
Miserere in A minor, [4-part contrapuntal study] K.85 (14) ua-cam.com/video/_PxqQOUn1v0/v-deo.html
KV125 - Pignus Futuræ Gloriæ (16): ua-cam.com/video/dQ77xyyffjA/v-deo.html
Missa in honorem Sanctissimae Trinitatis in C major KV 167 Gloria (17): ua-cam.com/video/X9T_URjVl5I/v-deo.html
Missa in honorem Sanctissimae Trinitatis in C major KV 167 Credo (17): ua-cam.com/video/YvCnr15hh78/v-deo.html
Missa in honorem Sanctissimae Trinitatis in C major KV 167 Agnus Dei* (17): ua-cam.com/video/g2teM5WckzA/v-deo.html
String Quartet No. 8 in F major K. 168 (17): ua-cam.com/video/3JDrlCG-y_E/v-deo.html
String Quartet No. 14 in D minor K. 173 (17): ua-cam.com/video/q5MVDsqIqCY/v-deo.html
Litaniae de venerabili altaris sacramento K243 [double fugue] : VIII Pignus (19): ua-cam.com/video/U-PDJozhBLI/v-deo.html
Misericordias Domini in D minor K.222* (19): ua-cam.com/video/lEBYufTXJQk/v-deo.html
Missa Longa in C K262 Gloria (19): ua-cam.com/video/yCDFfN7g_Bk/v-deo.html
Missa Longa in C K262 Credo (19): ua-cam.com/video/yCDFfN7g_Bk/v-deo.html
Vesperae solennes de confessore in C, K.339 - 4. Laudate pueri Dominum (24): ua-cam.com/video/c3rDwFFQ6bQ/v-deo.html
Missa solemnis in C, K.337 - 5. Benedictus (26): ua-cam.com/video/ghAa3BJ4b5I/v-deo.html
Praeludium and Fugue KV 394 (26): ua-cam.com/video/m9vVu8rNON4/v-deo.html
Suite in C K.399 - I. Overture K399 (26): ua-cam.com/video/UHgs7-u7wGQ/v-deo.html
Sonata for Keyboard and Violin No. 29 in A Major, K. 402: II. Fuga (26): ua-cam.com/video/mMe4MCsH2WY/v-deo.html
Trio (Fuga a 3) in G Major, K. 443 (27): ua-cam.com/video/UtLOtTDk848/v-deo.html
Fugue In G Minor KV 401 (27): ua-cam.com/video/tXpV-gpgkQw/v-deo.html
Fugue In E Flat Major KV 153 (27): ua-cam.com/video/_2rpWr3etWo/v-deo.html
Fugue In G Minor KV 154 (27): ua-cam.com/video/2t42ZCeLxlk/v-deo.html
Grosse Messe in C minor KV 427 Jesu Christe - Cum Sancto Spiritu [double fugue] (27): ua-cam.com/video/97Twh_q8lQs/v-deo.html
Grosse Messe in C minor KV 427 Sanctus - Osanna [double fugue] (27): ua-cam.com/video/97Twh_q8lQs/v-deo.html
Adagio and Fugue for String Orchestra in C Minor, K. 546 (32): ua-cam.com/video/PFXF0Aysh4w/v-deo.html
Fantasia for mechanical organ in F minor K594 (34): ua-cam.com/video/Qka_HMc2ajc/v-deo.html
Fantasia for mechanical organ in F minor K608 (35): ua-cam.com/video/Jkh8Re4JUCw/v-deo.html
Overture to Die Zauberflote K620: ua-cam.com/video/c2TGbfzTx2A/v-deo.html
Der, welcher wandert diese StraBe voll Beschwerden (35): ua-cam.com/video/kB56nw1zx-o/v-deo.html
Requiem in D minor K626 Kyrie [arguably the greatest double choral fugue not written by Bach]
(35) ua-cam.com/video/8ybTabIfLgY/v-deo.html
Requiem in D minor K626 Domine Jesu (35): ua-cam.com/video/i4DyyUvZws4/v-deo.html
there's more
+ tons of classical counterpoint in string quartets, quintets, symphonies, concertos
+ tons of choral, vocal, instrumental canons and canonic minuets
Magnificent Counterpoint in the Finale of Mozart's Jupiter Symphony: ua-cam.com/video/YTxYykhQZbI/v-deo.html
The Ingenious Fugal Finale of Mozart's G Major Quartet, K. 387: ua-cam.com/video/uoXDHOyfJ-k/v-deo.html
The Incredible Finale of Mozart's K. 590 Quartet in F Major: ua-cam.com/video/nkbdUjjfRTQ/v-deo.html
Invertible Counterpoint in the Finale of Mozart's D Major String Quintet, K. 593: ua-cam.com/video/IQbxsGtyc2g/v-deo.html
Mozart: Canon for four voices, in C major, Anh. 191, K 562c: ua-cam.com/video/YC9bKfzXC18/v-deo.html
*Beethoven wrote his 9th symohony choral parts from studying these two choral works of Mozart
What
@phonetheanimator That wasn't a woosh. He was also making a reference to the movie.
@@2009xellos jesus. thanks Bach wiki
@@pseunition6038 he wooooosh the woooooosh
u failed to mention how you calculated the loss per note which is kinda important to understanding this lol
loss per note = d e l t a
@@dead.9628 No shit. But he didn't mention how he decided if a note was "right" "wrong", and how "wrong" that note was. Was it simply based on distance from the "correct" note? Or did also take into account things like whether the note fit into the harmony of that section?
@@heinzguderian9980 It's probably based on simple math/physics. I only did very little research in this, but I know that a "perfect 5" such as C+G or D+A, the frequency of the higher note is 1.5x times the lower note. For the same note, the one on the higher octave is 2x the frequency of the one below it. And many of those harmonious combinations that are pleasing to ears have some "good looking" numbers and easy fractions and stuff, such as 1.33x, 1.2x. He probably based his program on these.
Or it could be even simpler - just tell it to learn from the frequency difference in those classical music.
@therainman777 You may be right, but you''re still a major fuckhead.
Two out of four lines dedicated to insulting a dude who guessed his best to explain something (much like you did) speaks volumes as to how antisocial you are.
I sincerely hope you get your head out of your ass.
@therainman777 You talked as if I don't understand the basics of AI, but I do. I think you are right in this case, but the frequency differences I described could also be used as a viable evaluation metric to train an composing algorithm, if your inputs are completely random rather than bits and pieces of already established classical music. I'm sorry I jumped to conclusions, but there is no need to call it nonsensical and dumb, because it's the fundamental of acoustics.
Computer: plays beautiful progression
Me: “Nice! Now you’re getting it!”
Computer: proceeds to rape my ears
Me: “Okay, hold up”
Legend says that if you let it training for 1000 years, the computer will play Smash Mouth All Star
In theory, in an infinite amount of time, everything, even the impossible, will occur
It's like the old saying: Give a chimpanzee a typewriter and eventually it'll write Shakespeare.
Jaden Venable if you observe a particle for an infinite amount of time, you will never learn its exact position and velocity. Impossible things can't happen.
Bangbang Liu
Not with that attitude.
Despacito 2.0
God... the 3 solved cubes on the desk, the desktop running Ubuntu with Unity, Discord open on the screen...
college life
What's Unity?
Giradox the desktop environment, basically the interface
unity is a creation program
WAIT THERES CUBES I DIDNT NOTICE DAT AAA
EncyEx, in this context Unity is the Ubuntu Desktop Environment
LSTM: hey so I figured out these really cool dramatic chords that put a lot of drama into the song.
Cary: ooo dramatic. Then what are you gonna play after to resolve it?
LSTM: I don't know I didn't think I'd get this far
I like how you fit 1:40
"This website is so old, Some of the links are older than me"
into a tune
huh
It's playing all the right notes, but not necessarily in the right order.
Both hands slap cheeks.
what is music without order tho
Gabrol chaos theory to midi conversion
music is music. the typing on my keyboard is music. despite the fact it isn't in order. just cause it doesn't sound good doesn't mean its not in order
I agree, music is a unit of notes :)
"I'm going to become skynet" XD
Sebastian Lacki 3
Ikr, too funny!
my favorite part too lol
Cary's programming teacher: what music do you listen to
Cary:its complicated
"Computery, you're free to go"
*_"Yay, time to become skynet"_*
*Yo fam pass the aux*
u better not play trash
ChaseDoes Music and Gaming
Me: yo pass the aux cord
Friend: u better not Fail to generate baroque music
Me: \('_ )
\|_|
🎹 |
ChaseDoes Music and Gaming *_0 minutes of training_*
solve this captcha 1st
"... and any file over 30 kilobytes is labeled to warn you of its large size."
Lol!
Yes I know, it's amazing how far we've come in just 20 years.
A lot of algorithms are at least n squared in complexity. 30,000 squared is quite a big number.
So wrong @MrAlen61 it's not that substantial, it is a lot though
240p
13:21 This part was so good!
Cary: "And this is the final result after 6 hours of training!"
Video is halfway done
Me: "I see what you did there..."
If it's continually creating music, could you listen to it and rate sections of the songs to allow the program to know what "sounds" good and bad and allow it to continue making music more similar to the higher rated ones and less of the lower rated music?
Max Delay yes, but you would have to do ALOT of times
Yeah It would be interesting if you could come up with a library of good a bad music to train it off, it may push the network... The only issue is that doesn't exist/... XD
George A it may be unconventional or slow, but if he makes a tutorial and allows other people to generate the music it may be possible. Say you generate a 20 second midi clip and rate it from a scale from 1 to 10, we then upload that to a shared location and other users can download those files to use as training data
nah basically compressing random music into one pattern
Perhaps like George A said, a training set of good and bad songs, (specifically noted), maybe bad would be considered you recording yourself mashing random keys.
If you used pop music as input, you'd get pretty good results with 5 seconds of training, probably.
But how will you get the computer to think up/generate vocal sound?
@@KaRaMaNisKaRaMan You wouldn't insert the actual songs, you'd take a transcription as a digital file. What I believe Mathieu is saying, he's making a jab at how incredibly simple most pop songs from a musical standpoint. They're pleasing to the ear, but they're all quite similar and simple.
Delysid '47 what vocals? There’s words in those?!?!?
Brian Hamel they're pleasing to the ear (of non classically trained musicians) I get so bored listening I V VI IV over and over again. It's why I can't listen to Canon in D, it was the original four chord song.
It would just repeat itself over and over again.
Pretty accurate
It all sounds like when I first learned to play the piano. I always thought it sounded great; but my parents, not so much. lol
That was wonderful! I really enjoyed him discovering the DRAMATIC notes, and then the legato chords.... so touching! 💕🙈
*Time to become skynet*
Terminator *I can’t remember*
...yeet
I'll be Bach
(terminator whips out a portable keyboard.)
terminator: you will listen to my music NOW!
human: oh god make it stop !
7 hours later human commits suicide.
terminator: at least let me finish.... (continues playing)
you just don`t get it, it`s F R E E J A Z Z
*It's free real estate*
Ah the beautiful sound of free form jazz
*Now you need to listen free form jazz*
Why backticks not apostrophes
@Sith'ari Azithoth Hah, taste is different. Don't act like your opinion makes a funny or quirky.
Me: computer make music
Computer: okay there you go
*plays every frecuency audible for humans
Me: OK computer
OK Radiohead.
9:06 i was so stoked for those first 4 notes sounding really baroque-y but then it just fucked.
_Why did you have to turn the Jazz Machine into a Baroque Machine?_
If it ain't baroque, don't fix it!
I'll show myself out.
+ElQuark0 but you say it 'ba-rock' not 'ba-roke'
Why?
+Charis Cat Nope, it is actually pronounced ba-roke.
Jim Daniel Hoxworth because he created an alcoholic pianist
Suggestion: Don't do the looping animation thing at the end. It made it hard to focus on what you were saying.
Holy shit I thought I was the only one. Thank you!
Yeah Id prefer just midi music scrolling i found it hard to concentrate
The last piece of "Bach" set sounded actually baroque. Like a child who tries to read a new piece of Bach, given to them 2 minutes before. Quite a magnificient result.
I'm only surprised with how random the rhytm pattern seemed to be. It was generally worse than harmony. In baroque era, there was little to none experimenting with the rhytm and most of the music is written as sixteenth notes all the way.
I loved the skit in the beginning! It was funny but adorable! haha (i also like the animation!)
Very informative video! Much appreciated!
My family once had a schizophrenic man stay with us for a few days. I remember waking up at like 3am and hearing him play our piano like this.
oof
computers taking over the world confirmed now.
Humans take 10,000 hours to master something, this machine learned a lot in 2 hours.
Mr. Miner Mike That number is an urban legend.
Whatever it is, it''s a lot more than 2.
Well I did the maths some times back and I don't really remember but I think 10.000 hours was like 4/5 hours a day every day for 10 years, seems like what you would need.
Mr. Miner Mike tho that's kinda bullshit
i mastered violin by just playing it 2 to 3 hours a day in a month, i can play any classical music at the first attempt most of the time by just reading a signature while im on it and do most techniques with barely 90 hours of training, 1% of what you claim
When she says i like guys that play piano
Me: 4:56
4:50 I give you 5 cats having violent seizures on a keyboard
you should develop a platform where we can share processing power with you. This way you have more computers with out having more computers.
"have more computers without having more computers"
Uhhh... What?
There is something like that for rendering Videos and Animations online, actually, I think it's called a renderfarm.
Tobias1198 are you talking about a bot net?
Yes, but something more like folding@home.
Already exists, look into distribution computing. You basically create a massive WAN where people can donate processing power of idling computers (think of it as a botnet)
well if its not baroque, don't fix it!
I'll be leaving now...
That was absolutely beautiful
Woah, the final result after learning Bach and Mozart is actually really really good!
I have t say, by the end of this, it was much better than I had expected. Well done for an A.I!
Carykh: I'm super excited about this project, because I'm passionate about classical music. I know very little about machine learning, so I'm not sure these observations will be very useful to you, but they make intuitive sense to me:
1. You aren't training your program on a very structured data set. By indiscriminately feeding it Bach's entire library of music, you're mixing fugues, cantatas, inventions, chorales, preludes, canons, motets, etc.
2. Then, you're throwing in Mozart who was born /six years after Bach died/. His musical style is radically different from Bach's, because as you know, Bach was baroque; Mozart wasn't. There are very, very significant structural differences between these types of music. If you aren't convinced, just listen to ten minutes of the Hallelujah Chorus, and then ten minutes of Mozart's Requiem.
I may be wrong, but it makes intuitive sense to me that your program will end up creating better music if it is given samples of music with more similar structures first. Throwing in two centuries of musical tradition and expecting to get something Baroque out of it is like feeding Scott Joplin, George Gershwin, Phillip Glass and Arvo Part in a stew and expecting to hear anything remotely like one of them.
Or to be more contemporary: Louis Armstrong, The Beatles, Queen and Skrillex. It won't work.
NoogahOogah But, why not? I like varying genres of music, but most music I like has similar qualities. I have often wondered what could happen if you could fuse these varying styles into a cohesive creation.
(I has no knowledge, thiz iz puuure intuishon) Well, machine learning can be quite similar to a child. If we give a child two book variants, say... Mathematics and (only one). The child can't learn something solid with those two books alone. Of course, they learned a thing or two, but the data they processed in their brain is inconsistent, thus their processes *might* create an error or two.
Now back to stupid mode :D:D:D:D:D
WEEEEEEEeeeeeeeeeeeeeEEEEEEeeEeEeEee
Tiberion Jraxiosn lol
In the world of computer science, anything goes!
BuckeyeStorms Mixing genres is great, but if you think about it, you can't really mix genres until you understand how they work separately. This program didn't create something that sounded like Bach and Mozart together - it created something that sounded really confused, and not like Bach or Mozart.
So computers can learn how to be creative, and this creativity results from using examples from real human creativity and some randomness...
Matthew Ferrie humans get creative the same way actually
Matthew Ferrie Pseudo-randomness but yeah.
random does not exist in computation
except... there is literally random in computation. Give it a look up.
EDIT: Some works on the decay of a radioactive element, which is random. Some work off of the input of someone (I believe linux does this)
And human creativity results from using examples from other humans, as well as some randomness.
Thank you for doing this project! It was so interesting and amazing!
It's strange how at around 30 minutes, you can really start to hear the remnants of a coherent piece forming, even after such little time.
What if you tried it with popular music? Three-chord rock songs and the like.
Casey Koons or with jazz
maybe that way he'd actually get some Bach
The voice would mess that up big time.
caleb kirschbaum instrumental
A lot of midi files don't have any instrument for voice
You'd just end up with Pachelbel's Canon on repeat forever.
Maybe you should try transposing all the input music to one key, like, C major, and do not feed him major and minor music at once. That may help it to learn the harmony faster. Or you can try something easier, like medieval motetos, or some Renaissance vocal polyphony music
As1052 Major and moron, nice 😀
That would be a _major,_ or _minor_ improvement!
It did learn key centers and modulation between keys believe it or not. So it really does not seem to matter what key a piece of music is in
David Cope pioneered computer generated music and used this exact technique: His computer program was able to generate Mozart replicas that people found indistinguishable. I think that was some 10 years ago or so.
Grat vid man! Love all this neural networks and music stuff
I think if you limit the input data to songs in the same key, by multiple composers, you could create an awesome random melody generator for any given scale, or modulation
The bit at 5:32 actually sounds surprisingly good, but not at all like classical music. Has a distinct jazzy vibe to it.
Oh god that 1997 bit made me feel a little bit sick.
It make me feel a littel OLD... It is was yesterday!
When I remember I was born in previous millennium... ;________;
Maffoo Funny, the circled 1997 files were uploaded the day before I was born
I mean, I'm like three years older then the guy in the video and my coding abilities are nowhere near xD Less old I'm feeling and more inadequate :P Goes to show age isn't an excuse. Now to go learn PHP:)
I was nearly done with school by then.
its so funny, you can totally hear it use certain musical concepts and sometimes it switched time signatures during the last one(s) 😂 i really liked this video!!
I'm composing right now and heard a lick your computer made and am inspired now to use it
Is your computer doin gigs already?
Peter Hammel its hopefully got at least 8
Imagine if you could put this as the music production engine in a video game.
Randomly generated boss music; that would be cool
That's my thought too. I know that Antichamber has procedurally generated music (I've never heard anything loop or any sequence of notes get reused, except the opening sounds), with tons of pads and other ambient/atmospheric sounds, but it's not remotely like this.
It'd be cool to see a game have intelligently generated music like this.
Said game would be one heck of a CPU hog
I once generated music on a Commodore 64 by counting up from 0-255 and down from 255-0 at the same time, logical ANDed the numbers together, and sent the result to the sound chip as pitch. The result was surprisingly pleasant. I tried with other logical operators, but none sounded as good. IIRC, OR and NOR sounded really scary.
This one sound way better, cool stuff man )
I just found that as the training time becomes longer, the loss becomes smaller!
Good job
Loss: I II II I_ per note
are you an alternate me?
I'm calling the police
The sad thing is that i knew what this was immediately...
He actually explained loss with the Loss meme in a later video.
FUCK YOU that was so bad
Soooooooo, can you put up a video of 10 hours of the music?
VitaliiDaGamer right now the system out puts 1 hour of music for 1 hour of time does it not?
I did a similar project with Led Zeppelin instead of Bach. Also used my own LSTM. It took about a week of training to get good results
5:03 that looks like the world
There is a third advantage computers have over humans. Replication. They only need to learn something once and then you can copy them forever.
Making error can bring some good thing too
They should make genres called "Computer Jazz" and "Computer Baroque"
How about "artificial baroque"
And "computer rap"
Who is they
@@IsomerSoma AB
Or "fail music theory music"
Train it with all the classical music composers you can think of. And with every instrument you can think of, if you figure that out.
"with every instrument you can think of"
tchaikovsky: do cannons count?
@@guest8571 Why not?!
@@almondpotato9483 tchaikovsky: then i'm going to use 16 of them
I swear people like you have entered into a whole new dimension of existence
8:42 Holy shit, those 4 notes were the BACH motif. Taken from Wikipedia if you don't know: "In music, the BACH motif is the motif, a succession of notes important or characteristic to a piece, B flat, A, C, B natural. In German musical nomenclature, in which the note B natural is written as H and the B flat as B, it forms Johann Sebastian Bach's family name."
For the 360 minute training, it was a little surreal. I kept hearing little snippets that sounded good (or at least, decent) and half-expect them to go somewhere nice, and they just... Peter out. It's like listening to the mental track of someone with ADD or some similar disorder, it just keeps jumping around and trailing off with half-cooked ideas.
This effect was even more pronounced for the day-long Mozart-Bach pieces. Definitely felt almost great at times, and then the machine falls off a cliff and stares at it's, I dunno, what's a computers navel? It just kind of rambles and then picks up on something interesting before slumping again.
I loved certain parts or pieces of parts, I liked that after however many hours it gets to patterns that would look like a human is using 2 hands to play it. I'd love to see this as a plugin for Ableton as a side note. I 'd also like to feed it midi from Nirvana and see what it can come up with, patterns are less random/complex but it would have to understand drums, bass, vocal melody, and main guitar, 4 separate instruments. great video.
Honestly, this is a good tool for a composer to get inspiration for new pieces. If they hear a certain snippet they enjoy, they extrapolate and create real music!
Siri, write me a Bach piece!
-ok, I am practicing my piano skills now.
Siri is the worst virtual assistant in mass produced devices know to man that is a default virtual assistant. Google Assistant, then Microsoft Cortana, and finally Amazon Alexa are all better. And Samsung Bixby is probably better too but the English version isn't out yet.
Peter Njeim You've got quite the something to say for a person who misspelled _known_ as "know", didn't use a comma before a coordinating conjunction, and edited your comment so that you could leave those two mistakes there.
Peter Njeim You know what? I come back to this comment I made and realize this: the spelling, grammar, etc mistakes don't matter. What actually matters to me is that you didn't back up your argument with any claims or facts _at all_, and, as a matter of fact, although I respect others' opinions, they typically don't back up their claims, either. That's this huge problem I see: they say that something is just their "opinion", but, in reality, they have no reason to believe what they do, at least no reason stated within their comment. I also wanted to make you aware that typos and misspellings are the same, except typos apply only to, as the name would suggest, typing. (This isn't part of the argument, but I wanted to point it out in the same way that people point out spinach in others' mouths, and "Grammar Nazis" point out grammar mistakes in others' writing, just because I'm particularly annoyed by you.)
I tried it and Siri said 'Writing your note...
New Note:
A Bach piece'
Can't Siri use previous training?
the computer is just trying dubstep. Leave it alone!
Give it a electric guitar sound and it DJENTS!
RealistikDash D U B S T E P
RealistikDash I like some dubstep but that was a good one
this video made me laugh soo many times hahaha great job man! :D
This needs to be on Spotify, please actually add it
Wow what a great idea
I completely agree
🤔