SteelyGAN: Semantic Unsupervised Symbolic Music Genre Transfer

Abstract

AbstractRecent progress in music genre transfer is greatly influenced by unpaired image-to-image transfer models. However, state-of-the-art unpaired music genre transfer models sometimes cannot keep the basic structure of the original song after genre transfer. In this paper, we propose SteelyGAN, a music genre transfer model that performs style transfer on both pixel-level (2D piano rolls) and latent-level (latent variables), by combining latent space classification loss and semantic consistency loss with cycle-connected generative adversarial networks. We also focus on music generation in individual bars of music with the novel Bar-Unit structure, in order to reduce coupling of music data within a 4-bar segment. We propose a new MIDI dataset, the Free MIDI Library, which features less data duplication and more comprehensive meta-data than other music genre transfer datasets. According to experiments and evaluations we perform separately on three pairs of music genres, namely Metal\(\leftrightarrow \)Country, Punk\(\leftrightarrow \)Classical and Rock\(\leftrightarrow \)Jazz, transferred and cycle-transferred music data generated by SteelyGAN have achieved higher classification accuracy, as well as better objective and subjective evaluation results than those generated by other state-of-the-art models.KeywordsSymbolic music generationMusic genre transferUnpaired transfer