What I would really like to see is an attempt to write multipart contrapuntal works like Bach with some kind of AI. The rules are fairly well understood, but Bach knew how to adapt and even violate them all but still wind up with amazing pleasant music.
As for deep learning models which create good contrapuntal music, see e.g. 'Biaxial RNN' https://github.com/danieldjohnson/biaxial-rnn-music-composit... by Daniel D. Johnson, who is now at Google Brain but wrote this as an independent(!) researcher. (Note that the existing code requires Python 2.x It would be interesting to forward-port it so it can work with Python 3.x and a maintained version of Theano. Replicating the model using Tensorflow would also be quite worthwhile.)
If you're interested in Bach's work specifically, the "BachBot" and "DeepBach" projects are also interesting but less accessible.
Example output for all of these models can be found on the Internet, just look around for it. The proprietary system AIVA is also worth mentioning because even though it's so proprietary and secretive, the compelling and "serendipitous" music it manages to come up with is a tell-tale sign that it's actually doing well-founded deep learning stuff behind the scenes, much like the aforementioned open systems. Note that much of the released output has been orchestrated (AFAICT) manually by humans, but at some point I was able to find some piano-format reductions that are most likely very close to what the AI actually created, somewhere on the official site.
Any prior information about the domain can be encoded in a way you prepare the input samples, and this would be equally useful and relevant regardless of the DL architecture you decide to use (RNNs, autoencoders, GANs, transformers).
Using raw audio samples has two advantages: First, you provide all available training data to a model, and trust that it will learn to extract what it needs from it. Second, unlike any other music representation formats, you have huge amounts of training data. So essentially all you need is a huge transformer and many thousands of GPUs to train it on, using all mp3s you can get your hands on, and you can reasonably expect a similar level of quality as the quality of text generated by GPT-3. What do you think?
Adding features to provide further information about the input data can be helpful, but not nearly as much as ensuring that strong priors are reflected in the actual model architecture, whenever relevant to the domain. Johnson's paper about biaxial RNN provides one example of how to do this in the context of music.
Forked it and spent the day upgrade to Aesara (the maintained Theano fork) and python3.
Haven't fully trained/tested the model, but feel free to take a look: https://github.com/kpister/biaxial-rnn-music-composition
Also, the old version could use a special incantation:
theano.config.experimental.unpickle_gpu_on_cpu = True
This project is quite similar to something I've been working on, but I realised early on that you can't split up features like this and get credible results - credible meaning "appropriate for the style grammar."
It's a bit - only a bit, but let's go with it - like trying to generate sentences by swapping out nouns and adverbs. You end up with something that is grammatically correct in theory but makes no sense in practice.
Classical music particularly is fundamentally integrated in a way that textbook analysis doesn't fully explore.
Here is another very good one: https://openai.com/blog/musenet/
I'm a trained musician myself and interested in automatic music composition following the progress for the last thirty years, but only recent work (like the ones referenced) produce convincing results (besides Cope's work of course, but which required manual selection and editing).
You might also be interested in this survey paper: https://arxiv.org/abs/1709.01620
If so, the most important reason that it sounds terrible is the music is generated with MuseScore, without adjusting dynamics, tempos, etc. Actually, the Beethoven's original sonata in this blog is also generated with MuseScore, and it sounds not so good even with dynamics added.
However, this is why I agree that deep leaning is more promising than this manual approach, since too many variables you need to adjust to make music sound good rather than syntactically correct.
I think it would be cool to combine the two. Instead of generating raw midi, your GAN or reinforcement learning agent or whatever could try to generate sequences of transformations to melodic fragments. Neural program synthesis type stuff.
Or maybe one could build an automatic music analysis tool that can start from the score and try to infer the program that generated them. (Is that a thing already?)
What's cool about this library isn't the quality of the (midi, retch) output -- what's cool is the actual Python library behind it, and the writeup, which are both super easy to read and follow.
And I think that is not a very nice thing to say about other people’s work.