Even with modern technology, artificially generating music indistinguishable from human creations proves to be challenging. This paper describes a new approach by learning patterns commonly found in music using autoencoders combined with a novel sampling strategy for creating smooth transitions between learned patterns. The authors train and evaluate their methods on a newly created dataset.
Link to paper: https://www.ijcai.org/Proceedings/2020/0751.pdf
Further reading includes the longer version of the paper called “CONLON A PSEUDO-SONG GENERATOR BASED ON A NEW PIANOROLL, WASSERSTEIN AUTOENCODERS, AND OPTIMAL INTERPOLATIONS”.