NonupleCLMR: Contrastive Learning of Musical Representation with NonupleMIDI

Abstract

An excellent music representation means that the model can better understand music with symbolic data, which is conducive to better complete various downstream tasks. The current mainstream research for music representation learning is to use mature pre-training techniques in natural language processing. However, there are some obvious differences between symbolic music and natural language: (1) The basic elements in symbolic music have multiple features. (2) Symbolic music is composed of multiple tracks, which means that not only the sequence association within the track must be considered, but the relationship between different tracks must also be considered. (3) There is a lack of training data in symbolic music. In response to the existing problems, we provide a comprehensive solution by proposing a novel contrastive learning framework of Musical Representation with NonupleMIDI (NonupleCLMR). We design several strategies to deal with the above problems, including NonupleMIDI encoding and simCLMR mechanism to enhance better representation with symbolic music data. In the experimental part, we not only indicate the advantages of NonupleCLMR in music performance in downstream tasks at the phrase level and song level but also design a series of ablation experiments to reveal the effect of the proposed strategy.