By Gaëtan Hadjeres, Frank Nielsen, and François Pachet.
GLSR-VAE: Geodesic Latent Space Regularization for Variational AutoEncoder Architectures. arXiv:1707.04588, submitted 14 July 2017. DOI: 10.48550/arXiv.1707.04588.
Read the paper (PDF)
arXiv record
DOI
About the paper
How can the hidden representation learned by a generative model become a useful musical control? GLSR-VAE introduces geodesic latent-space regularization for variational autoencoders. It shapes the learned representation so that moving through it corresponds to changes in attributes of the generated material.
Abstract summary
The regularization is added to the VAE training objective to connect latent-space movement with continuous data attributes. The authors demonstrate the method on monophonic music, generating controlled variations of discrete note sequences. The work studies how a learned continuous representation can support intentional exploration of musical variations.
Related work
Flow Machines — the broader research on steerable music generation.
DeepBach: A Steerable Model for Bach Chorales Generation — related work by the same authors on constrained chorale generation.
Deep Learning Techniques for Music Generation — A Survey