As a holder of a degree in music AI composing is interesting and yet horrifyingly terrifying. Maybe we should all bow down to our robotic overlords before mankinds eventual extermination, after the singularity.

I think I learned more about neural networks from this video than I did from the Computerphile video. Maybe I just needed another alcoholic to explain it to me

@CopperHvH

5 жыл бұрын

alcoholics explain the best

@aurora2319

4 жыл бұрын

indeed me too.

@mr.bigglesworth2952

3 жыл бұрын

When did it become normal to NEED alcohol??

@jackmackenzie6721

3 жыл бұрын

@@mr.bigglesworth2952 last year

@chuckmaddox6725

2 жыл бұрын

@@mr.bigglesworth2952 since people

I think part of the problem is that its output is being limited to one "instrument" or sound in these examples, whereas the examples its being given are using a whole variety of sounds/instruments. If you trained it strictly on piano solos and play the melodies it comes up with on a piano it would probably sound much better. Shame this is probably going to be buried because I'd love to see another attempt with this approach.

Are you planning to release the songs made by AImadeus? I would like to listen to them, at least the final one

@yeetthelizard4307

3 жыл бұрын

Hot damn you should team up with Michael reeves

@madgear1174

3 жыл бұрын

LMAO I see what you did there! (I'm just trying to fit in I'm actually very confused but not sexually)

"My primary career is in computer science." *shows him juggling balls* Accurate Great video btw.

F*ck yeah, Rotting Revenant Ruckus is the heaviest shred I've ever heard brother!!!

OMG did carykh get a voice changer?

Honestly, I'd love to see more videos like this, maybe use less editing so it isn't as hard on you. Maybe possibly make them more frequently?

@StanleyJones365

4 жыл бұрын

Phatdoughnut fork

@motorsphere356

4 жыл бұрын

@@StanleyJones365 &$@+-#

I've seen some of your other videos, and I thought they were all really cool or interesting, but this one has truly convinced me that this channel has some magic to it. I will certainly stay tuned for the future, your big break on KZread is coming soon.

@Qxir

5 жыл бұрын

Thank you, I'm glad you like my stuff :)

@realkekz

5 жыл бұрын

Yeah, cheers from the USA, man!

@pegtopmaster8780

Жыл бұрын

Welp you were right about the big break

@firstnamelastname6216

11 ай бұрын

​@@pegtopmaster8780for real lol 😂

@realkekz

9 ай бұрын

@@Qxir Congratluations on 1 million subscribers. I was right :P

You should really consider doing more videos like this. I like the way you explain stuff. It’s like a friend explaining what going on.

Strange that it still sometimes plays discordant notes, I would think that proper cords would be one of the first things the network would pick up on.

@xadithy9116

4 жыл бұрын

Madotsuki chords are basically simple ratios between notes so a computer wouldn’t know what sounded good since to simple is overly basic and to complicated sound awful

AImadeus! AImadeus! AImadeus! Doesn’t have the same feeling.

@BoserPSN

3 жыл бұрын

Falco was too special

I love your take on the project of using ANN's for music development. I've been working on a similar project since 2009 and I've made so much progress that I'm about to patent my project that makes music from images using a custom stem library that I made. Final touches soon-ish. :)

@TT-hd3zi

5 жыл бұрын

Good luck with that man

@ashleighosborn1994

5 жыл бұрын

do you have a demo project of the ai?

@JosephFuller

5 жыл бұрын

Good luck getting past the parent trolls. You should have patented the concept years ago, you do not have to have a final working product to secure a patent. I am 98% sure that it has already been done by now. I did something similar; making music from images (no ANN) back in the 90's so... However, you should definitely keep working on it for the pure pleasure of getting it done, I just would not expect too much. Of course, you could get incredibly lucky and actually secure a patent and make some money, and I hope you do have success but my experiences have left me a bit jaded.

@anonymouscrab2013

5 жыл бұрын

Is the music on your channel made by your ai? I like it however it was created, but that'd be really impressive!

@endlessnameless8181

3 жыл бұрын

Anything come of this?

The music actual starts at 7:54

@Unknown-kw1vz

3 жыл бұрын

thx)

Hey Qxir, i recently found your channel and i wish i found it sooner, it’s genuinely amazing and i wish you posted more >:( but we all have our own lives and schedules. keep posting whenever you can and i hope your very successful in your job outside of youtube

T'would be great to see some more videos like this! Thank you!

Honestly didn't know what you were in to in life until seeing this. Glad I subscribed and would DEFINITELY like more content like this. We appreciate the work you are putting into this project!

Enjoyed the video very much, great editing and acting :) Needed to implement image recognition with ML not long ago for a project and it was a nice recap. Keep it up!

@Qxir

5 жыл бұрын

Thanks :)

I wasn't able to watch this yesterday :( But it's a great video! Your references and self-depricating humor is on point. Also, AImadaus. What a name.

Concert when?

Great timing! Me and my friend are doing a similar project regarding RNN but it’s all about recognizing IQ test patterns rather than patterns in music

Very interesting concept, my good sir. Please continue sharing as things evolve.

Honestly that Song 2 could pass for an indie game ost.

Qxir, I have suspected you worked in computer science! You remind me so much of what I was like 40 years ago. You didn’t ask for advice, but here it is anyway… stay focused on interesting problems, especially ones that can make a difference, and not on wealth. Wealth will take care of itself, if you do the former. And by wealth, I mean you will have enough. Best wishes.

Very interesting. Never thought I would actually learn something on this channel.

Jokes aside (and I liked them a lot) this video has very seriously intelligent points and discoveries with both feet in the future! Keep going.

I commend you're efforts. I was hoping for better results. This makes me want to pull up some old code for a rhythm generator that I worked on in the past. It's not an AI but the results are often pretty pleasant.

Omg when AImadeus started playing I was in tears! Basically from that point on in the video I couldn’t keep it together! You built it up so much and I had such high hopes and expectations for it too Bahahaha!

Rotting Revenant Ruckus had some decent riffs Mate!!!!!

Played around with, and even built a few, learning bots back in the late 1990s and early 2000s. Neural networking music creation is certainly taking things beyond the next level.

“Pop is dead.... and i killed it” Hahhahaha

Dude thats sick! I love your AI creepy tunes

Rotting Revenant Ruckus could actually be pretty heavy with a bit of work. Very interesting!

You should make a colab with Code Bullet , it would be great

I like the music your AI produces. But I'm also into Black Metal so what do I know...

@randoprior4130

5 жыл бұрын

Ain't nothing wrong with a little black metal... Maybe add in a double kick and a nasally pissed off guy with many tattoos.

@craigantihero393

5 жыл бұрын

Saying your into black metal makes you feel so edgy, huh?

@darkerdaemon7794

4 жыл бұрын

@Craig, not as edgy as the razor you lobotimized your sense of humor with.

@AmericanIdiot7659

4 жыл бұрын

@@darkerdaemon7794 Gottem

@endlessnameless8181

3 жыл бұрын

@@darkerdaemon7794 lol

I see potential for some extremely experimental techno. I want to see you train it to do some crazy arpegios.

I love you dad

My dad did some work on the original creation of midi software. To this day I have no idea what exactly his job was, but he was still getting royalty checks from it up until a couple of years ago.

You should make a video on Air Canada Flight 143 Gimli Glider, and a video on Chris Mccandless.

You should team up with purposeless rabbitholes.he has some pretty kool bit crush? Tunes.

I love it and wouldn't mind some more like this

Music is made interesting because of tension and release. If the program could learn which notes create tension when placed before the resolution note like 5-1, it may be able to make something more coherent.

lad your notification button ain't working I didn't get any notifications as a fellow Irishman I'm outraged

Ok now this is epic. Btw I never had any idea you were in computer science that’s awesome. I know I’ve said it a bunch of times but keep these awesome videos up!!!

Did you use scikit-learn and Python for the coding of your network? I use Python for my attempts at networks but my results have been hit and miss. Did you try using the Fourier Transform to decompose waveforms into separate frequencies and use their derivatives as inputs? The outputs could be Time+dt. What did you use as an activation function for your nodes? Did you use gradient descent or a more randomized stochastic function to calculate the learning rate?

I would LOVE to see a follow up on this if one is available

The inputs and outputs and function seem similar to a Markova chain. How does this perform compared to a Markov chain?

We need more videos of this

tales from the bottle PLEASE

I love this. Thank you.

ONE IDEA: A digital audio is made of samples where each sample is 1/44100 seconds (assuming a samplerate of 44100, you need at least [maxfrequency of X * 2] samplerate to be able to produce a max frequency of X, max frequency we can hear is 20000). Unless the song is mono there is one sample for left channel and other right channel (so 44100 *2 per second). Each frequency is a integer value between +32767 and -32768, how distant from 0 the value is, means the volume with more distant from 0 meaning louder and 0 means true silence. You could make some AI, that start by generating random list of numbers, and then producing the audio result of this list of numbers. You would listen to this result and then give an score between 1 and 100, this score would be the fitness of this result. Based at how you rate stuff (that is the fitness) the song would produce better results. The program goldwave can read texts with that data and play the audio (export some mp3 audio you already have to this txt format [goldwave allow you to do that too] to check the layout of this audio text) SECOND IDEA (IF FIRST DON'T WORK): There are some guys/texts talking about how using audio data instead of midi is too hard for machine learning/learning ai... because of the amount of data (44100 numbers per second , times 2 if the song is stereo), I am not an ai coder and know almost nothing about it, if that's the case then my first idea wont work. If my idea wont work there is an second idea that may work. Each instrument has a timbre and most of musicians learn just one instrument and will make all their songs with this instrument for their entire life, so you can see a timbre is am very important thing. If the first idea don't work, you could instead of trying to generate an entire song with previous method, you could try to generate a timbre (imagine sinewave, squarewave, piano sound for some examples). A timbre would have smaller lenght around 1 second and so less data you need to analyse. The program would start by generating random data and then playing this timbre and play the frequencies of the middle 12 octaves of the piano (check here for frequencies en.wikipedia.org/wiki/Piano_key_frequencies if you dont know them) and you would give an score between 0 and 100 and this would be the fitness and based at the fitness the program would generate timbres to you rate and over time it would generate good timbres (based at your opinion)

@puffmangofrom08

3 жыл бұрын

Nerd

@noahhastings6145

3 жыл бұрын

NEEEEERRRRRRDDDD

@exedeath

3 жыл бұрын

@@noahhastings6145 If I was a nerd I would do it myself.

love the blur reference

AImadeus sure could make a career in making background music for horror games and such. "Song 2" for example sure would fit a Doom game

Amazing I love your work please keep posting and remember me when your famous cause I know you will be famous

Well done, you are a multi talented being who we don't deserve to have wasting his time to just entertain us with your Wonderous Stories. Have a look at The Nashville Numbering System, that might help you get the AI to write songs of a more song-like nature. I'll Email you a brief guide to the basics of music theory, that might help you to understand the Nashville system.

Did you pre-process the songs to make them all in the same key? It's a step that I think would likely help AIs grasp the relationship between the notes, but I haven't seen any videos about composing AI that normalize by key. I wonder how that could change the results (thinking the notes would be represented in half-note steps from the base note of the scale, and the octave part remains the same). Or the AI could predict the next interval between notes instead of the actual note itself... The possibilities! They're endless!

Qxir, I love your voice!

A few years ago I got myself a nasty dose of brain damage. I've never really understood why my brain behaves like it does. That simple and well explained description. Has helped understand it. I'm sure not what you intended but cheers boss!

Rotting Revenant Ruckus was actually alright. But "Funeral Dirges" was hilarious.

What tool did you use for the neural networks or did you code the back propagation yourself?

Have you looked at any other music making AIs? There's one called AIVA that has an interactive website and lets you choose the genre, tempo, key, instrumentation and other things, then the AI (very quickly) writes a song that's usually good

what if you made a network based on duets where the primary duet player is the input and the second player is the output. the input and out data will clearly be related to each other, giving the network a boost. then once the network is made you could plug in a solo song and the network could try and make a duet track.

Because you have a community you should create a test to see if people can discern work of an IA and work of human. After you use Kappa to graph distance between people and then you color your graph with correctness of people and see if some people are able to know if they are correct.

Very interesting what field of study in computer is this I hate reading books so I would love to watch some KZread videos

Great new video Qxir. Keep it up. You're quite good at it.

@Qxir

5 жыл бұрын

Ty!

So did you let this keep running ? hw did it turn out did it learn to make a song?

What scores do actual songs get on your formula for judging songs?

You deserve way more views than you get

@Qxir

5 жыл бұрын

I get a lot though!

@akishot6735

5 жыл бұрын

Qxir because it is extremely well deserved, should have it listen to only to Falco next

@lan_daddyttv4710

3 жыл бұрын

@@Qxir No sir.

You should give it a rating system out of 100, have normal songs as 60-70's, classical music as 80's & then use random sound giberish as 0-20 so the ai can learn what is "desirable". This will force the system to produce songs using techniques that will achieve high ratings. Number 1's could also be used as 100 rating refference but i feel like the woulds matter equally in them so would use OST/Dubstep or something staple techno thats popular. Anyways I hope to see some more experiments with this as I really liked your methods and ideas!

Great video! 👏 👏

Rotting revenant ruckus. Audio of the year.

This was super interesting.

I think it’s interesting that it learned the high level structure of pop music.

I think it just needs more then 2 notes as a input for a better output since classical music has at times up to 3-4 notes being played

This is the best video you’ve ever done.

Bruh when you played the song I was dead, and the titles were low key hilarious

Has the recording contract been signed? When does the album get released? I rather quit liked the third song.

Holy sh*t that music works really well for parts of a horror game!! It's bloody creepy! :D Also awesome work mate!

@Qxir

4 жыл бұрын

Thanks :)

And so the uprising began.

Im thinking about getting into coding, any tips?

UnIronically really loved Rotting Revenant Ruckus

Good video keep it up.

"Pop id Dead And I Killed It" sounds like a nice tune for a funeral.

What county u from ? I’m from Galway

If you assign a number to each key on a key board that would be the input. The output would be the combination of keys to play to song. I don't know, but It makes sense to me

That was the best pop music I ever heard!

Your thoughts on AIVA?

I'd really like to hear it try to make some 1950's piano rock'n'roll

The last classic one could had been a boss theme for a nes game. ;)

love ur videos

Try contact/collab with Adam Neely about music theory.

Can we get an update on what AImadeus has been up to lately?

Honestly, "Pop is Dead... And I Killed It" would make for GREAT background music in a 90's video game.

Can you do a video of Shawn Nelson who went in a tank rampage in 1995

Music is like prime numbers. It's the space between the patterns. I think what you need is to just use your best programs to determine the silence instead of the sound.

Umdude, that's fucking cool. Maybe design multiple iterations of the AI that learn off of eachother, you could train one on just midi drums, and use that as a main input for the next ai that would control a baseline or some higher notes and sounds. If you kept all outputs at the same bpm and had a way to set the key the ai is playing In, it probably wouldnt sound to terrible. I'm not a coder, but if you did it right you could become an EDM artist without writing a single song

OoooOooh Qxir done be smaaart. He dun retaineded all thems book learnins. I uh... mayonnaise is an instrument!

Qxir’s actually pretty damn smart! Who knew? Just kidding man, this was really interesting. Well done.👏🏽👏🏽👏🏽

I say just like writing music is a trial and error process so should writing the code for an AI to create music. All the aspects of music could be broken down into modules that work together for the desired result and each module tweaked individually. There can be modules for the rules of counterpoint, melody, harmony, rhythm, style, etc. Maybe some riff loops that keep repeating while you make input changes to the AI parameters... but then you're inserting yourself in the process and might as well be writing the music yourself... but maybe that's a temporary presence on your part and the true test when you step away and watch the AI work on its own. The key, I think, is complexity, which I always tell myself is nothing more than a bunch of simple things put together. Mapping out the many facets of what makes certain music appeal the the ears and minds of the majority of listeners would certainly help. In this sense you have to be more of a musician than computer scientist to make it work. Really you have to be both. Sorry for having rambled; I just found this very interesting.

Honestly, this was super interesting even if the results were... I'll be generous, let's say, surprising. Maybe, you could add an input which is a relative critical rating of the song. That way, AImadeus could know what kinds of songs sound good or bad. I don't know anything about the topic of programming or AI, but it's still very fascinating.

@TrentFalkenrath

2 жыл бұрын

Wait, no. A repetition factor. If the same note occurs more than one in a short time, make it so it has to change to a different one. That will introduce variety.