A suggestion if I may:
It might be interesting to "feed" your neural network with some songs from known artists like the Beatles and see what comes out of it.

 

Thank you very much Futurewine.
PM me if you want to send me midi to see what it generates.

 

At the moment nothing is going to happen LOL. I did a huge mistake in the beginning of the project and I now have to retrain it if I want to add influences from bands/genres that it wasn't trained on. But eventually it could generate music not just by genre but by artists too.

 

Again, it sounds great, except for some weird key changes.

If you manage to give it a set of rules to improvise over the right key, even if you have to input the key manually, it could be really usable, even commercially.

I'm not a programmer, but as a generally creative person, I know that the hardest part usually is when you're 15% close to "perfection", and the closer you get, the harder it gets as well...

Find a way to teach it some rules, like Rapid Composer does, and you're set.

 

Do you have job vacancies in the neural network? I am looking for a job.

 

RemindMe! 8 Months "Check this out!"

Oh, wait, it's not Reddit...

This option would come in handy, though. I really wish to see what this thing will be able to do after another 8 months.

 

Don't worry, V2 is going to be awesome.

 

This is insane, very nice work, any possibility of us testing it out?

 

Thanks
Yes when v2 is ready i'm probably going to let people from this forum try it out. For now you can send me a midi file and I will put it though the network and send you the result. Just PM me

 

I am interested to see the source code if you can share it and it is open source. Also, a paper if anything is published. Did you use available ANN packages like Tensorflow/Keras/MATLAB or you implemented the network manually?

 

@SceleJohn Firstly, I want to thank you for sharing what you are doing with us, and wish you success. I also want to offer you my help in the area of chord recognition with Convolusional Neural Networks. I'm very familiar with Dynamic Bayesian Networks, the Takuya Fujishima's pitch class profiles system, and the use of hidden Markov models trained by Expectation-Maximization algorithm to recognize chords.

 

I'm not going to make the code open source and I haven't published anything yet. The only info that I can give at the moment is that most of the stuff I use is made by me.

 

Thanks Pirate.
This network doesn't recognize chords or notes what so ever. The midi gets converted by a separate program to a system that i call "4D coordinates" which basically creates "maps" that represent the original midi and the neural net is designed to make continuations of those "maps". So the network itself does not even know what a note is, it just creates patterns based on the original and the data from it's training. The reason I went with my "4D coordinates" system is because it allows for the program to "remember" parts of a song more accurately and it's more modular.
This is part of a bigger project.I'm trying to make a DAW with different tools like this.

 

@SceleJohn Have you thought about resulting "maps" containing chords,and key identifiers in order to have the NN make random continuations of those maps but in proper key and chord progressions?

 

I think giving it key identifiers or preset chords would limit it in some ways. For example if it got an identifier for c major it would probably never go out of that. Unless if it was programmed to use identifiers and chords as strong suggestions, but that still wouldn't fix the problem. I think with enough training it will "learn" to stay in key and perform correct modulations.

 

That is a fact my friend. It will learn.
As I said earlier, if you need some help just send me a PM.

 

If you think that,why are you contributing to the development of these softwares?!
I don't understand there.

 

The tools that I'm trying to create are supposed to assist with music composition not replace musicians, think of it as a very advanced fl studio riffer. But there are companies out there that are developing AI that makes music entirely on it's own. That's what I'm scared of.

 

Oh well, at least when Skynet decides to obliterate us, we'll prolly have some nice music playing! =D

 

Or they could imprison us and then generate unlimited Justin Bieber music just to torture us.