Hyperbolic Music Transformer for Structured Music Generation

In the field of music generation, generating structured music is a highly challenging research topic. Music generation methods are currently learned in Euclidean space and usually modeled as a time series without structural properties, but due to the limitations of the time series representation in Euclidean space, the hierarchical structure of music is difficult to learn, and the generated music is poorly structured. Therefore, based on hyperbolic theory, this paper proposes a Hyperbolic Music Transformer model, which considers the hierarchy in music and models the structured components of music in hyperbolic space. Meanwhile, in order for the network to have sufficient capacity to learn music data with hierarchical and power regular structure, a hyperbolic attention mechanism is proposed, which is an extension of the attention mechanism in hyperbolic space based on the definition of hyperboloid and Klein model. Subjective and objective experiments show that the model proposed in this paper is able to generate high-quality music with structure.