Do we need more complex representations for structure? A comparison of note duration representation for Music Transformers

TL;DR

This paper investigates whether simple, unannotated MIDI representations can match the structural performance of more complex, annotated representations in Music Transformers, emphasizing the efficiency of unannotated data.

Contribution

It demonstrates that minor modifications to basic MIDI representations can improve Music Transformer performance without relying on annotated structural data.

Findings

Unannotated MIDI representations can achieve comparable structural similarity metrics.

Small tweaks to MIDI representations lead to significant performance gains.

Searching for better unannotated representations is more cost-effective than annotating data.

Abstract

In recent years, deep learning has achieved formidable results in creative computing. When it comes to music, one viable model for music generation are Transformer based models. However, while transformers models are popular for music generation, they often rely on annotated structural information. In this work, we inquire if the off-the-shelf Music Transformer models perform just as well on structural similarity metrics using only unannotated MIDI information. We show that a slight tweak to the most common representation yields small but significant improvements. We also advocate that searching for better unannotated musical representations is more cost-effective than producing large amounts of curated and annotated data.