Learning Domain-Adaptive Latent Representations of Music Signals Using Variational Autoencoders

In this paper, we tackle the problem of domain-adaptive representation learning for music processing. Domain adaptation is an approach aiming to eliminate the distributional discrepancy of the modeling data, so as to transfer learnable knowledge from one domain to another. With its great success in the fields of computer vision and natural language processing, domain adaptation also shows great potential in music processing, for music is essentially a highly-structured semantic system having domaindependent information. Our proposed model contains a Variational Autoencoder (VAE) that encodes the training data into a latent space, and the resulting latent representations along with its model parameters are then reused to regularize the representation learning of the downstream task where the data are in the other domain. The experiments on cross-domain music alignment, namely an audioto-MIDI alignment, and a monophonic-to-polyphonic music alignment of singing voice show that the learned representations lead to better higher alignment accuracy than that using conventional features. Furthermore, a preliminary experiment on singing voice source separation, by regarding the mixture and the voice as two distinct domains, also demonstrates the capability to solve music processing problems from the perspective of domain-adaptive representation learning