Transposition Invariant Symbolic Music Representations

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 233 papers46 conceptsfrozen 2026-10-04

When a human hears "Happy Birthday," they recognize the melody whether it is sung by a bass in the key of C or a soprano in the key of G. To the human ear, the essence of the melody lies not in the absolute frequencies of the notes, but in the distances between them. However, standard computational models often struggle with this concept. If a model treats a middle C (MIDI 60) and a D (MIDI 62) as entirely distinct entities, it must learn the same musical patterns twelve times over—once for every possible key.

To bridge this gap, researchers have developed a suite of techniques to represent music through relative intervals and ratios, creating models that are inherently invariant to changes in transposition and tempo.

The Geometry of Musical Intervals

The most direct way to achieve transposition invariance is to stop recording absolute pitches and start recording the "distance" from the previous note. In MELONET I, Dominik Hörnel demonstrated that interval-based coding captures inherent symmetries in music, such as octave equivalence. By embedding these domain principles directly into the input features, the model's learning burden is significantly reduced.

This approach is particularly powerful for music generation. Moray Allan and Christopher K. I. Williams used pitch intervals relative to a reference melody to automate Bach-style chorale harmonization. Because the model sees intervals rather than absolute notes, it can generalize the "style" of Bach across different keys without needing separate training data for every transposition. Similarly, Eliya Nachmani and colleagues utilized "scale-invariant" representations—encoding notes as intervals relative to the scale's root—to train high-quality generation models on very small datasets.

Invariance in Time and Velocity

The principle of relativity extends beyond pitch. Just as a melody is defined by the ratio of its frequencies, a rhythm is defined by the ratio of its durations. Taizan Suzuki and Takenobu Tokunaga proposed encoding expressive parameters like volume and tempo as ratios relative to parent segments. This allows a model to transfer a "performance style" from one piece of music to another, regardless of the absolute tempo or length.

In the context of music retrieval, such as "Query-by-Humming," this dual invariance is critical. Naziba Mostafa and Martin Rohrmeier have shown that representing melodies as sequences of relative pitch intervals and logarithmic inter-onset interval (IOI) ratios makes systems resilient to the fluctuations of informal human performance. Whether a user hums slightly flat or speeds up mid-tune, the underlying relative representation remains stable.

Embedding Invariance into Neural Architectures

While early work focused on manual feature engineering, recent research integrates invariance directly into neural architectures. Zixun Guo and Dorien Herremans introduced a bias-adjusted sinusoidal encoding for musical notes. In this embedding space, the geometric distance between two tokens is determined solely by the musical interval between them. This allows a Transformer to perform a transposition via a simple linear operation, preserving the relative meaning of the sequence.

For tasks where the relationship between components is complex—like melody and accompaniment—researchers use adversarial techniques to "force" the model to be relative. Jingwei Zhao and Gus Xia employed Domain Adversarial Training (DAT) to disentangle melody from chords. By corrupting the melody through random transpositions and forcing a discriminator to ignore those shifts, the model is compelled to represent harmonic movement in a purely relative sense within its latent space.

Applications in Retrieval and Detection

The robustness of relative representations makes them the standard for identifying musical similarity. In the MOVE framework, Furkan Yesiler and Emilia Gómez applied circular convolution to pitch representations, ensuring that a song and its "cover version" are recognized as the same even if they are performed in different keys.

This is also the foundation of modern plagiarism detection. Tianyao He and Junchi Yan developed systems that treat music as a bipartite graph of relative pitch and duration intervals. By focusing on the internal structure rather than absolute values, these models can detect "fine-grained" theft where a plagiarist has slightly altered the rhythm or transposed the melody to hide their tracks.

By shifting the focus from "what note is this?" to "how does this note relate to the last?", these researchers have moved AI closer to the way humans actually experience and create music.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 35 papers
  2. Version 1 · 2026-10-04 · 33 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_relative_interval_encoding_dc20bec6,
  title = {relative interval encoding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/relative-interval-encoding/rdr_dc20bec692da465fb4c26f7a2afcdbf1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_transposition_invariant_symbolic_music_r_dc20bec6_v1,
  title = {Transposition Invariant Symbolic Music Representations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/transposition-invariant-symbolic-music-representations/rdr_dc20bec692da465fb4c26f7a2afcdbf1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}