PIANOTREE VAE: Structured Representation Learning for Polyphonic Music
This work addresses representation learning for polyphonic music, a domain-specific problem for music AI, and is incremental as it extends existing VAE methods.
The authors tackled the problem of representation learning for polyphonic music, which had been largely limited to monophonic music, by proposing PianoTree VAE, a tree-structured extension of VAE, and demonstrated its validity through semantically meaningful latent codes, improved reconstruction, and benefits for downstream music generation.
The dominant approach for music representation learning involves the deep unsupervised model family variational autoencoder (VAE). However, most, if not all, viable attempts on this problem have largely been limited to monophonic music. Normally composed of richer modality and more complex musical structures, the polyphonic counterpart has yet to be addressed in the context of music representation learning. In this work, we propose the PianoTree VAE, a novel tree-structure extension upon VAE aiming to fit the polyphonic music learning. The experiments prove the validity of the PianoTree VAE via (i)-semantically meaningful latent code for polyphonic segments; (ii)-more satisfiable reconstruction aside of decent geometry learned in the latent space; (iii)-this model's benefits to the variety of the downstream music generation.