Multi-Level Representation of Long MIDI Sequences: Integrating Bar-Level Encoding with Music-Level Context

Symbolic music, represented as MIDI files, encapsulates intricate performance details and complex temporal and structural dependencies. Effectively modelling ultra-long MIDI sequences is essential for understanding sophisticated compositions and advancing tasks like music generation, classification, and performance analysis. However, these long MIDI sequences pose significant challenges due to their complexity, multi-track concurrency, and extensive temporal relationships. This paper introduces a novel model named LongMIDI-Net that enhances the capability of bar-level pretrained large-scale MIDI sequence understanding models, extending their effectiveness to handle complete and long MIDI sequences. The proposed approach integrates structure-sensitive models for processing bar-level segments with temporal-sensitive models to capture global relationships across entire sequences. This hierarchical design significantly reduces sequence length while maintaining high model performance. Comprehensive experiments on classification tasks across multiple datasets demonstrate the superior effectiveness of the proposed model, achieving consistently strong results. Furthermore, ablation studies highlight the advantages of bar-level segmentation over random slicing, showcasing its ability to provide a more effective and structurally coherent representation of MIDI sequences. These findings underline the importance of combining local and global information for advancing symbolic music understanding.