LSTM Neural Network that tries to write piano melodies similar to Bach's (2016)

30 LSTM layers of 30 neurons each? Two notes as input? Those are some unusual choices...

For a dataset this small, a single layer LSTM with 300 neurons, accepting four bars of input at a time, would most likely do much better.

Very cool, you might be interested in an open source chord-melody dataset I’m hand curating for generating pop melodies:

https://github.com/shiehn/chord-melody-dataset. I’m feeding into LSTM’s and and its working quite well in my system :)

http://treblemaker.ai/

Btw I just noticed that he's the same guy whose work on GAN was used by an art collective, and the resulting image sold for $14k. https://twitter.com/DrBeef_/status/1055285640420483073

I’m not sure why this has been compared to Bach. It’s not very similar at all, and quite jarringly disorganised. I can’t comment on the method used to make it (though would love to understand more, as this is an area that professionally interests me) but you may want to go back to the drawing board :)

Cool! Music generated by neural networks is such an interesting field.

Me and some friends also had a go at it. We used two different datasets, one from Bach and one from Final Fantasy. In contrast to many other approaches, we allowed chords as well as playing multiple notes at the same time (not necessarily the same as playing a chord). You can find some results here: https://github.com/HalfVoxel/deep-piano (they are definitely cherry-picked, but it worked out quite well overall imo).

Very different to Bach.

Glad to see people posting links to music modeling projects - it's a huge space, and a ton of fun to actively work on or just mess with if you like music and code. Even simple markov models can get someplace with a bit of tweaking [0][1].

If you like this, you might like some more recent work from Anna huang et. al. at Magenta, using a modified Transformer network [2][3].

Some other relevant reading includes DeepBach [4][5], and outside the domain of Bach alone, the venerable folkRNN [6]. There's a huge chain of work I posted a while back here [7], and Kyle McDonald wrote a really great review piece here [8].

Modeling music in latent spaces has also seen a number of recent publications [9][10][11], and having smooth latent spaces for generation seems (to me) to be a very natural interface for manipulating things, which has a natural analogue in "real" composition, cool video at [12].

There have also been some really fun online demos for interacting with these types of systems [13]

There's a lot of interesting work going on in music modeling, hopefully leading to new and interesting musical tools and toys. Interactivity (often via clever conditioning) can really change the way you approach these types of problems [14].

[0] https://github.com/kastnerkyle/magenta/tree/my_cleanup_branc...

[1] https://www.youtube.com/watch?v=Ay8mHW0-NW8&list=PLRMa_gJ8vx...

[2] https://twitter.com/kastnerkyle/status/1018617315196309504

[3] https://arxiv.org/abs/1809.04281

[4] https://www.flow-machines.com/archives/deepbach-polyphonic-m...

[5] https://github.com/Ghadjeres/DeepBach

[6] https://github.com/IraKorshunova/folk-rnn

[7] https://news.ycombinator.com/item?id=14822281

[8] https://medium.com/artists-and-machine-intelligence/neural-n...

[9] https://storage.googleapis.com/magentadata/papers/multitrack...

[10] https://magenta.tensorflow.org/music-vae

[11] https://www.youtube.com/watch?v=BbyvbO2F7ug

[12] https://www.youtube.com/watch?v=A8d-PUuUJnA

[13] https://codepen.io/teropa/

[14] https://twitter.com/chrisdonahuey/status/1051868153700016128

A Markov chain would have done better. Something is seriously wrong here - not even the 2-grams are accurate. If your neural network can't outperform a simple control like a Markov chain, it's hardly worth publishing.

A file with space in its name? We had an entire article about not to do this here.

Nice idea.

I suspect Bach made use of a lot more neurons to write this:

https://m.youtube.com/watch?v=aEkXet4WX_c&t=7s

From 2016. Add that year to the title.