| component-id | type | name | description | image | work-package | pilot | project | release-date | release-number | release-link | doi | licence | demo | changelog | copyright | contributors | related-components | bibliography | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Harmory |
KnowledgeGraph |
Harmory: the Harmonic Memory |
A Knowledge Graph of interconnected harmonic patterns aimed to support computationally creative applications. |
assets/harmory_wide.png |
|
|
polifonia-project |
01-02-2023 |
v1.0 |
10.5281/zenodo.8021211 |
|
Copyright (c) 2023 Harmory Contributors |
|
|
|
Harmory is a Knowledge Graph of interconnected harmonic patterns aimed to support creative applications in a fully transparent, accountable, and musically plausible way.
We leverage the Tonal Pitch Space - a cognitive model of Western tonal harmony to project chord progressions into a musically meaningful space. Then, we use novelty-based methods for structural analysis to segment chord sequences into meaningful harmonic structures. The latter are then compared with each other, across all progressions and via harmonic similarity, to reveal common/recurring harmonic patterns.
A KG is created to semantically establish relationships between patterns, based on: (i) temporal links, connecting two patterns if they are observed consecutively in the same progression; and (ii) similarity links among highly-similar patterns. By traversing the KG, and moving across patterns via temporal and similarity links, new progressions can be created in a combinational settings; but also, unexpected and surprising relationships can be found among pieces and composers of different genre, style, and historical period. This is also enabled by the scale and diversity of Harmory, which is built from ChoCo, the largest existing collection of harmonic annotations.
Currently, Harmory contains ~26K harmonic segments from 1800 harmonic (~10% of ChoCo, corresponding to all the audio partitions). Out of all segments: 13667 (16%) correspond to the same pattern families, 66175 (53%) are pattern-friendly (they share non-trivial similarities with other segments), whereas 8176 (32%) are inherently unique (they are found in other songs). More statistics are available at this link.