Jeongganbo Arrangement: Computational Revival

• Jeongganbo Arrangement is a computational reconstruction of 15th-century Korean court music using a formalized grid notation system.
• It employs transformer-based architectures, including BERT-like MLM and encoder–decoder models, to generate style-faithful multi-instrument arrangements.
• The method integrates precise beat-counter embeddings with JG-like encoding to ensure rigorous rhythmic coherence and accurate ensemble orchestration.

Jeongganbo Arrangement refers to the computational and musical reconstruction of 15th-century Korean court music using the Jeongganbo grid notation system, as implemented by Han et al. for ensemble revival and performance. The project advances a formalized token-based encoding of Jeongganbo and applies modern transformer-based architectures to achieve style-faithful arrangements for six traditional instruments, grounded entirely in a restricted, strictly metrical encoding. All claims and workflow details are drawn from Han et al. (2024) (Han et al., 2024).

1. Formal Jeongganbo Encoding Scheme

The Jeongganbo notation consists of a two-dimensional grid where each “jeonggan” square represents a beat (gak), and note onsets are specified by vertical positions within the box. The arrangement project proposes a “Jeonggan-like” (JG-like) encoding, whose elementary vocabulary includes:

• Bar boundary token (“|”) marking jeonggan endings,
• Line break (“\n”) for measure boundaries,
• 16 sub-beat position tokens $s \in \{0, \dots, 15\}$ indicating subdivisions within the jeonggan,
• Pitch tokens (e.g., A4, $B^\flat$4, C5) following position tokens,
• Ornamentation (sigimsae) tokens suffixing the pitch.

A note event is formally encoded as the tuple $(g, k, s, p)$ where:

• $g$: jeonggan index (beat number),
• $s$: sub-position within the jeonggan,
• $p$: pitch symbol.

The onset time in quarter-note units is computed as: $\tau = (g-1) \times 1 + \frac{s}{16}$ with duration determined by the difference $\tau' - \tau$ to the next event or by sustaining through empty jeonggans.

This system circumvents the need for a wide duration vocabulary: only 16 position tokens are required for rhythmic resolution, and duration is deduced contextually. The positional logic of Jeongganbo ensures only valid rhythmic groupings, as $s=15$ cannot precede $s=0$ within one jeonggan.

2. Transformer-Based Arrangement Architectures

Two transformer variants underpin the arrangement system:

A. BERT-like Masked LLM (MLM) for Melody Infill

• Input: Each jeonggan unrolled into six time-frames per rhythmic minimum; one-hot representations for symbol and ornamentation, with explicit beat counter embeddings.
• Architecture: 12 layers, 128 hidden size, 4 heads, 0.2 dropout.
• Objective: Span masking (15% jeonggans), frame-level (5–20% frames), ornamentation masking; cross-entropy recovery.

B. Encoder–Decoder Transformer for Ensemble Orchestration

• Input: JG-like token streams for up to five instruments plus instrument ID; no conventional position embedding, instead using three one-hot beat counter embeddings (measure index $B^\flat$0, beat-within-measure $B^\flat$1, sub-beat $B^\flat$2).
• Encoder/decoder: 6 layers each, 128 hidden dim, 4 heads, 0.2 dropout, with standard multi-head attention and cross-attention.
• Objective: Negative log-likelihood (cross-entropy) with teacher forcing for the target instrument.

The BERT-MLM enables style-preserving melodic infill, while the encoder-decoder transformer aligns and generates multi-part arrangements, effectively modeling heterophony. Beat counter embeddings enable precise metrical awareness, which is critical in low-resource (85-piece) scenarios.

3. Positional Encoding of Note Durations

The encoding does not represent explicit duration tokens (e.g., quarter, eighth notes). Instead, each jeonggan is subdivided into 16 positions: $B^\flat$3 As an example, position 4 corresponds to $B^\flat$4 quarter-note into the jeonggan; a subsequent note at position 8 has duration $B^\flat$5 quarter-note.

When a run of jeonggans is empty, the preceding note sustains through them by convention, with bar tokens signaling note extension if no onset occurs. This implicit duration model eliminates the combinatorial overhead of legacy symbolic durations.

4. End-to-End Arrangement Pipeline

The full workflow is structured as follows:

A. Data Preprocessing

• Optical Music Recognition (OMR): CNN+transformer OMR is applied to all 85 extant Jeongganbo ensemble scores, yielding 141,820 monophonic lines tokenized in JG-like format.
• Tokenization: Insertion of “|” and “\n” at jeonggan and measure boundaries, respectively.

B. Melody Transformation (Monophonic to Piri)

• Rhythm normalization: Adaptation from 8/8-grouping to a 10-jeonggan span congruent with Yeominlak tradition, with empty jeonggans inserted for stylistic conformity.
• Masking: Sliding 4-measure window with 2 measures teacher-forced; next 4 measures masked at 15%.
• Inference: MLM infills masked slots, generating a piri-line retaining full JG token structure.

C. Orchestration (Sequential Six-Part Generation)

• Sequential passes: Each part generated conditionally (Piri→Geomungo, Piri+Geomungo→Gayageum, ... , final part is Daegeum).
• Refinement loop: After initial generation, each part is regenerated conditioned on the other five to enhance ensemble coherence.

D. Post-processing

• Inverse tokenization yields MusicXML scores; minor manual adjustments include pitch transpositions and corrections for non-scalar notes.

5. Evaluation: Metrics, Expert Review, and Performance

Objective Metrics

• Length-Match Rate: Fraction of generated lines matching input jeonggan count; evaluates structural fidelity.
• F₁-score: Exact onset+pitch alignment with ground truth.

Sample metrics (Piri→Geomungo / All→Daegeum):

Encoding	Len-Match	F₁-Score
JG-like	0.942/1.00	0.679/0.614
REMI-like	0.923/1.00	0.567/0.532
ABC-like	1.00/0.903	0.704/0.542
No beat-counter	0.135–0.403	0.043–0.090

JG-like encoding outperforms others for rhythmically complex parts (daegeum), while ABC-like performs best for simpler instruments (geomungo). Removal of beat-counter embeddings substantially degrades both metrics, underlining their importance.

Subjective Expert Review

Seven musicians from the National Gugak Center rated ensemble outputs pre- and post-refinement on a 1–5 scale for range, ornamentation, rhythm, and harmony. Mean ± std scores improved following refinement, especially in ornamentation and ensemble coherence.

Live Performance Feedback

The arrangement was performed by the Court Music Orchestra on King Sejong’s birth anniversary; expert performers noted genre-appropriate rhythm and idiomatic ornamentations. Minor pitch adjustments were necessary for daegeum and gayageum, but overall outputs required minimal intervention.

6. Design Rationale and Implications

Each major choice in the Jeongganbo arrangement pipeline is justified with reference to quantitative and qualitative outcomes. JG-like encoding yields a compact vocabulary and enforces rhythmic grammar. Beat-counter embeddings are essential in regimes with limited training data to maintain metrical positioning. The BERT-MLM enables effective infilling of incomplete historical melodies, while the encoder–decoder model, through cross-attention and refinement looping, achieves stylistically coherent ensemble textures. The comprehensive workflow demonstrates that transformer-based generative models, when paired with encoding schemes faithful to traditional notation and explicit metrical embeddings, can successfully revive historical works with limited but highly structured data (Han et al., 2024).