The paper proposes a new structure-informed positional encoding framework for music generation with Transformers. The authors note that music generated by deep learning often lacks coherence and long-term organization, which is a distinctive feature of music signals. They design three variants of positional encoding and test them on two symbolic music generation tasks. The methods proposed in this paper improve the melodic and structural consistency of the generated music.
Publication date: 23 Feb 2024
Project Page: bit.ly/structurepeword
Paper: https://arxiv.org/pdf/2402.13301
