Style-VT: Style Conditioned Chord Generation by Variational Transformer With Chord Substitution

Recent advancements in music modeling have shown the potential of Transformer in tasks such as chord progression generation. Despite these successes, controlling chord progressions to reflect genre-specific characteristics remains underexplored. In this work, we present Style-VT, a style-conditioned chord generation model that combines the Transformer architecture with a variational autoencoder. Our model leverages genre embeddings to generate chord progressions that align with the stylistic norms of specific genres. We also introduce a newly defined chord substitution into the objective function, allowing for greater flexibility in generated chord progressions. By extensive experiments, we demonstrate that Style-VT generates a more flexible chord progressions and achieves higher harmonic similarities with human-composed chords than the existing methods.