Skip to main content
Have a personal or library account? Click to login
The Potential of Unsupervised Induction of Harmonic Syntax for Jazz Cover

The Potential of Unsupervised Induction of Harmonic Syntax for Jazz

Open Access
|Jun 2025

Figures & Tables

Figure 1

(a) First bars of lead sheet (from the iRealPro forum2) and (b) syntax tree representing the harmony of ‘Take the “A” train' by Billy Strayhorn. Image based on Harasim (2020). Dm7 is a terminal chord symbol, V/V is a preterminal since it is an abstract symbol representing a single terminal chord, and the highest occurring V is a nonterminal since it represents multiple lower‑level symbols. The turnaround has been omitted.

Figure 2

Graphical diagram of the neural PCFG, after Kim et al. (2019). S is the start symbol; A1, A2N are nonterminal; P1, P2, P3P preterminal; and c1, c2, c3V chord symbols. πr are rule probabilities, computed from embeddings E by neural networks f. Full lines between circles are PCFG rule applications, dotted lines indicate probabilistic dependencies, and full lines between rectangles indicate computations by neural networks.

Table 1

Chord embeddings comparison on JHT.

Embeddingsval. F1
Flat.428
Comp2flat.453
Components.431
Table 2

Validation F1 on JHT for progression losses when training on JHT+iRealPro. I/M distinguishes between inside scores and marginals.

SpansI/MF1Δ
No loss.453
RuleI.466+.013
M.484+.03
StatisticalI.438.015
M.465+.012
Figure 3

Validation F1 for (a) the number of nonterminal (NT) and preterminal (P) symbols and (b) the hidden dimension of the MLPs used in the N‑PCFG. Shaded region represents standard deviation.

Table 3

Test F1 and recall (Re) of our models and baselines. Unsupervised models are on top, supervised models are in the middle, and trivial baselines are below. Best results per category in bold.

ModelTrain dataF1Re
N‑PCFGJHT.384.391
N‑PCFG+iRealPro.450.462
+ Prog. loss.461.476
N‑PCFG+ChoCo.474.482
Prog. loss.487.495
SupConJHT.640.643
MuDePJHT.607.606
Random.178.200
Left‑branching.134.142
Right‑branching.176.211
Figure 4

‘Blue Bossa’ by Kenny Dorham: (a) prediction and (b) annotation. Lead sheet in Suppl. Figure 9a.

Figure 5

‘Sunny’ by Bobby Hebb: (a, left) prediction and (b, right) annotation. Lead sheet in Suppl. Figure 9b.

Figure 6

NT precision versus a sample of (a) span ending chords sorted by frequency and (b) spans of chords. ‘o' indicates a diminished chord, ‘%' a half‑diminished chord, ‘m' a minor chord, ‘7’ a major seventh, ‘7’ a minor seventh, and ‘6' a major sixth. Some nonterminals are not used in this plot, which can be explained by the plot not showing all possible spans.

Figure 7

Induced rules with probability πr for N‑PCFG without progression loss. For every symbol on the left of , starting from the start symbol S, the two rules with highest estimated probability are printed, and this is repeated for nonterminal symbols that occur on the right of the in one of the already printed rules (but the set of printed nonterminal symbols is only expanded 3×). A are nonterminal and P preterminal symbols. This is only a fraction of all the rules.

DOI: https://doi.org/10.5334/tismir.217 | Journal eISSN: 2514-3298
Language: English
Submitted on: Sep 2, 2024
Accepted on: May 12, 2025
Published on: Jun 20, 2025
Published by: Ubiquity Press
In partnership with: Paradigm Publishing Services
Publication frequency: 1 issue per year

© 2025 Ruben Cartuyvels, John Koslovsky, Marie-Francine Moens, published by Ubiquity Press
This work is licensed under the Creative Commons Attribution 4.0 License.