Published September 8, 2026 | Version v1

Where Does Instrumentality Reside in Symbolic Music? A Structural Analysis of a Symbolic-Domain Instrument Classifier

Authors/Creators

  • 1. ROR icon Nippon Institute of Technology

Description

Instrument recognition has traditionally relied on timbral cues in audio signals. However, in symbolic music representations such as MIDI, timbre is absent, raising the question of how instrument identity is inferred from note sequences alone. In this work, we investigate how instrumentality is approximated in a learned symbolic-domain instrument classifier by analyzing its behavior under controlled structural perturbations. Rather than focusing on classification accuracy, we examine how perturbations such as pitch normalization, voice extraction, and temporal reversal affect predicted instrument labels and internal representations. Our results show that instrument identity is strongly influenced by structural features such as pitch register and simultaneous note configuration, while temporal ordering also has non-negligible effects. Furthermore, changes in predicted labels are not directly explained by the magnitude of representation shifts, suggesting structured reassignment patterns rather than uniform degradation. These findings suggest that, when timbral information is unavailable, the analyzed classifier relies on structural regularities in symbolic patterns to infer instrument identity. This provides an empirical perspective on how instrumentality can be approximated in learned symbolic representations through structural cues rather than timbral information itself.

Files

Where Does Instrumentality Reside in Symbolic Music - A Structural Analysis of a Symbolic-Domain Instrument Classifier.pdf