Pith. sign in

REVIEW 4 major objections 5 minor 39 references

MeloBottleneck: Self-Supervised Melody Skeleton Extraction with a Latent Subsequence Bottleneck

T0 review · 4 major / 5 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Learning a melody skeleton as a length-controlled latent subsequence transfers better than imitating keep/delete pseudo-labels.

desk verdict Clean self-supervised subsequence formulation for monophonic skeleton extraction; transfer edge over pseudo-labels is real but rests on thin transfer sets. read the letter →

arxiv 2607.10233 v1 pith:NYB7F24E submitted 2026-07-11 cs.SD cs.LG

Fan Bu , Rongfeng Li , Linfeng Fan This is my paper · ORCID
classification cs.SDcs.LG
keywords melodyskeletonextractionlatentsubsequencebottleneckself-supervisedlearningsymbolicmusicornament-invarianttrainingrhythmicclosurefragmentretrievalinformation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Melody skeleton extraction tries to keep the structural notes of a tune and strip ornaments so that variants can be compared, retrieved, or traced to a theme. Earlier systems either apply hand-crafted reduction rules or train note-by-note classifiers on keep/delete labels that come from the same procedural generators that create the training data; those labels inherit the generator's biases and never force the selected notes to form a coherent shorter melody. MeloBottleneck instead treats the skeleton as an order-preserving latent subsequence of controlled length, turns that subsequence into a self-consistent melody by a deterministic rhythmic-closure step, and trains the extractor by forcing a decoder to re-ornament the skeleton back to the original tune, by matching a frozen melody language model, and by remaining consistent under synthetic ornamentation. On a synthetic test set that matches the generator the pseudo-label classifier still wins, but on two zero-shot transfer sets—classical variations matched to themes and Chinese ornamented melodies matched to gongche skeletons—the latent-subsequence model is stronger and cleaner. The same skeletons also improve sparse fragment retrieval under ornamentation and corruption while shortening queries.

What carries the argument

The MeloBottleneck: a hard top-K subsequence extractor plus a deterministic rhythmic-closure operator that absorbs deleted time into retained notes, trained end-to-end with reconstruction, prior, length, consistency, and ornament-exclusion losses so the selected notes must themselves form a usable reduced melody.

What would settle it

On a held-out style whose real ornaments systematically differ from the paper's procedural ornamenter, measure whether MeloBottleneck's Hard F1 / Insertion Mass advantage over a matched pseudo-label classifier disappears or reverses when both models are trained only with that ornamenter and evaluated against human or alignment-based skeletons.

Watch

Extended reading notes

Core claim

Representing a melody skeleton as a length-controlled, order-preserving latent subsequence—closed into a standalone melody and optimized by re-ornamentation reconstruction, a frozen melody prior, and ornament-invariant consistency—yields more robust zero-shot transfer than training a note-wise classifier on procedural keep/delete pseudo-labels, even when the latter is strongest on a matched synthetic benchmark.

Load-bearing premise

The procedural ornamenter used both to create the strong training views and to mark which notes are synthetic insertions must be a faithful enough model of real ornaments; if it systematically mismatches a target style the invariance and exclusion signals can mis-penalize genuine surface notes.

Editorial extensions

If this is right

  • Skeletonized versions of both queries and documents raise Recall@K and MRR for BM25 fragment retrieval under ornamentation while cutting per-query time.
  • A single trained compressor can be run at externally chosen compression ratios, giving controllable reduction depth without re-training.
  • Because the objective never requires keep/delete labels, the same architecture can be applied to any monophonic corpus that admits a procedural ornamenter, not only to styles for which hand-crafted reduction rules exist.
  • The extracted closed skeletons are already standalone melodies and can therefore serve directly as control signals for harmonization or generation systems that previously needed separate skeleton annotations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the rhythmic-closure operator were generalized to keep or re-allocate rests and metric positions rather than forcing zero inter-event gaps, the same bottleneck could support polyphony and harmony-aware reduction without changing the subsequence-selection core.
  • The same latent-subsequence-plus-reconstruction pattern could be tried on other ordered symbolic sequences (e.g., speech prosody or dance notation) where surface elaboration hides a shorter structural backbone.
  • Because free-length prediction is only indirectly tested, a natural next measurement is whether the model's own predicted ρ already yields usable skeletons for retrieval or theme tracing without oracle length.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes MeloBottleneck, a self-supervised model for monophonic melody skeleton extraction that treats a skeleton as a length-controlled, order-preserving latent subsequence rather than independent note-wise keep/delete labels. A Transformer compressor selects top-K events, a deterministic rhythmic-closure operator absorbs deleted score time into retained events to form a standalone reduced melody, and a re-ornamentation decoder reconstructs the input. Training combines reconstruction, a frozen autoregressive melody prior, ornament-invariant consistency across procedurally ornamented views, and explicit exclusion of synthetic insertions. Evaluation covers synthetic OOD ornament-to-backbone (O2B), zero-shot TAVERN variation-to-theme (V2T), and Jiugong ornamented-to-gongche (O2G), with a matched-capacity pseudo-label classifier baseline, ablations, and a BM25 fragment-retrieval study showing gains from skeletonization.

Significance. If the transfer claim holds, the work offers a useful alternative to heuristic reducers and pseudo-label note classifiers for symbolic MIR tasks that need shorter, ornament-stable melodies (comparison, theme tracing, retrieval). Strengths include a clear task reformulation (selection vs. rhythmic closure), a matched-capacity pseudo-label baseline, multi-regime evaluation, loss ablations with diagnostic lifts, seed-averaged metrics on O2B, and a concrete downstream BM25 experiment with improved Recall@K/MRR and lower query time. Public demo and code materials further support reproducibility. The main scientific stake is whether latent-subsequence learning truly transfers more robustly than pseudo-label imitation outside the synthetic generator regime.

major comments (4)
  1. [Table 1, Table 2, §5.1] Table 1 and Table 2: the central transfer claim (latent subsequences transfer better than matched pseudo-label imitation) rests heavily on V2T and O2G. O2G has only 20 sequences / 3.1K notes. Hard F1 is 0.8942 (±0.0116) for MeloBottleneck vs 0.8893 (±0.0061) for O2B-Learner—within overlapping uncertainty and not a clear win—while IM is stronger. With such a small set, the reported edge is statistically thin and could reverse under different held-out volumes or seeds. The paper should either enlarge O2G, report significance tests / bootstrap CIs, or substantially soften the O2G transfer claim.
  2. [§4.1, §6, Table 1] §4.1 and §6: V2T references are alignment-induced correspondences between variations and themes, not independent human structural reductions. Hard F1/CFA on V2T therefore measure recovery of aligned theme notes under a fixed length budget, not necessarily learning of a musically coherent skeleton. This weakens the interpretation that better V2T scores demonstrate superior skeleton extraction rather than better recovery of the alignment target. The manuscript should either add a human-reduction or expert-judgment evaluation on a subset, or reframe V2T strictly as a theme-recovery / correspondence task and adjust the abstract/conclusion wording accordingly.
  3. [§4.1, §3.2, §6] §4.1 and §6: main selection metrics use oracle compression ratio ρ = K★/L, so free-length prediction is only lightly tested (retrieval uses predicted ρ for queries). The task formulation emphasizes length-controlled skeletons and a ratio predictor (Eqs. 5–6, length regularization), yet the headline comparison isolates selection under known K. At minimum, report Hard F1/IM under predicted ρ on all three benchmarks, and/or error of ρ prediction, so the operational claim is not left mostly to the retrieval setting.
  4. [§3.6, §4.4, §6] §3.6 and §6: ornament-invariant consistency and exclusion (Eqs. 19–21) depend on the same procedural ornamenter O used to build strong views and mark insertions. The paper correctly flags style mismatch risk, but this is load-bearing for the self-supervised story: if O under- or over-represents real ornaments, LC/LE can systematically bias selection. A sensitivity study (e.g., train with a reduced ornament set and evaluate on OOD/real ornaments, or swap O parameters) would better secure that transfer gains are not artifacts of O’s inductive bias.
minor comments (5)
  1. [Title, Abstract, §1] Throughout: several spacing/hyphenation artifacts appear (e.g., “TA VERN”, “formula tion”, “LA TENT”, “MELOBOTTLENECK” line breaks). Clean for camera-ready readability.
  2. [Figure 1, §2.1] Figure 1 is described as contrasting three formulations; ensure the published figure clearly labels (a) prior-driven reduction, (b) note-wise salience, and (c) latent subsequence + rhythmic closure so the task reformulation is self-contained.
  3. [§3.3] Eq. (14) soft-hard path and stop-gradient construction is standard but dense; a short sentence on what is hard at inference vs. soft for gradients would help non-specialist readers.
  4. [Table 3] Table 3 ablations report Δ from Full; adding absolute Full values in a footnote or repeating the Full row metrics for V2T/O2G would ease reading without flipping to Table 1.
  5. [§6] §6 notes that closed-skeleton musical quality is unevaluated; even a small informal listening check or qualitative examples in the supplement would strengthen the “standalone reduced melody” claim beyond selection metrics.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: transfer claim is an empirical comparison on external benchmarks, not a result forced by definition or self-citation.

full rationale

MeloBottleneck’s central claim—that learning skeletons as length-controlled order-preserving latent subsequences transfers more robustly than matched note-wise pseudo-label imitation—is an empirical head-to-head on held-out regimes (synthetic OOD O2B, TAVERN V2T, Jiugong O2G), not a first-principles derivation. Training objectives (re-ornamentation reconstruction LR, frozen AR melody prior LP, length/timeline regularizers, ornament-invariant consistency LC, ornament exclusion LE) do not embed the evaluation keep/delete labels; the matched O2B-Learner baseline is an external comparator trained on procedural pseudo-labels, not a teacher whose outputs are re-labeled as MeloBottleneck predictions. Rhythmic closure is a deterministic post-selection operator, not a fitted quantity renamed as a result. Self-citations ([21], [22]) supply training/evaluation corpora (Jiugong/gongche resources) and do not load-bear uniqueness theorems or force the transfer ranking. Architectural citations (SEQ3, BART, MusicBERT, FiLM, DCTTS) are standard prior art, not smuggled ansätze that define the claimed transfer superiority. Evaluation under oracle ratio isolates selection quality against independently constructed references (alignment-induced V2T themes; held-out gongche skeletons). No step reduces Eq. X to Eq. Y by construction or renames a fit as a prediction. Score 0 is therefore appropriate.

Assumptions & free parameters 4 free parameters · 4 assumptions · 3 invented entities

The central transfer claim rests on standard Transformer/seq2seq machinery, a domain model of monophonic events, a hand-designed rhythmic-closure rule, a procedural ornamenter, and a collection of loss weights and annealing schedules chosen by validation. No new physical entities are postulated; the invented pieces are architectural operators whose value is measured by the reported benchmarks.

free parameters (4)
  • loss weights (λ_R, λ_P, λ_L, λ_C, λ_E, λ_T) = (1.8,0.6,10,4,2) + anneal
    Set to (1.8, 0.6, 10.0, 4.0, 2.0) with λ_T annealed 0.1→0; chosen by validation and directly control the trade-off among reconstruction, prior, length, consistency and exclusion.
  • ρ_min / ρ_max and length-regularization (μ_L, σ_L) = ρ∈[1/3,1], μ=2/3, σ=0.2
    ρ range [1/3,1], μ_L=2/3, σ_L=0.2 control the distribution of bottleneck lengths the model is exposed to; hand-chosen to improve adaptability.
  • soft-gather temperature τ and leaky penalty κ = τ 1.5→0.5, κ=0.7
    τ annealed 1.5→0.5, κ=0.7; control gradient flow through the hard top-K selection.
  • timeline-alignment σ_T and early-training schedule = σ_T=0.075
    σ_T=0.075 and λ_T annealing prevent early collapse to local regions; ad-hoc regularizer.
assumptions (4)
  • domain assumption A monophonic melody can be represented as a sequence of (pitch, duration, delta-time) events after fixed quantization (R=12).
    SimpleMono representation (§3.1); standard in symbolic MIR but not forced by physics.
  • ad hoc to paper Deleting interior notes can be repaired by absorbing deleted score time into the preceding retained event (zero inter-event gaps).
    Rhythmic-closure operator Close(z) in §3.3; a modeling choice that makes the subsequence a standalone melody.
  • domain assumption Procedural ornament operations (grace notes, trills, inserts, rearticulation, and OOD variants) generate realistic strong views and correctly mark non-skeleton insertions.
    Used for LC and LE (§3.6) and for the O2B benchmark; acknowledged as a limitation in §6.
  • domain assumption A frozen autoregressive melody LM trained on ordinary (augmented) melodies supplies a useful melody-likeness prior for reduced skeletons.
    LP in §3.5; assumes ordinary-melody statistics remain informative for skeletons.
invented entities (3)
  • MeloBottleneck hard-subsequence compressor + soft ordered gather
    purpose: Select a length-controlled, order-preserving latent note subsequence with differentiable top-K approximation.
    Core architectural invention; no independent existence outside the model.
  • Deterministic rhythmic-closure operator Close(z)
    purpose: Turn a selected subsequence into a self-consistent standalone melody by reassigning durations.
    Defined ad hoc in §3.3; necessary for sequence-level objectives.
  • SimpleMono multi-attribute event representation
    purpose: Compact monophonic encoding of pitch, duration and delta-time with tied embedding/output heads.
    Paper-specific reduction of compound-token schemes; enables the multi-head losses.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MeloBottleneck: Self-Supervised Melody Skeleton Extraction with a Latent Subsequence Bottleneck." pith.science (2026). https://pith.science/paper/NYB7F24E

@misc{pith2026260710233,
  author       = {Pith},
  title        = {Pith review of: MeloBottleneck: Self-Supervised Melody Skeleton Extraction with a Latent Subsequence Bottleneck},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NYB7F24E}},
  note         = {Machine review of arXiv:2607.10233}
}
read the original abstract

Melody skeleton extraction aims to derive a shorter melody that preserves structural notes while removing ornaments. Prior methods rely on hand-crafted reduction rules or note-wise salience classifiers trained with heuristically or procedurally generated pseudo-labels. Such supervision can inherit generator bias and does not explicitly optimize a coherent reduced melody. We introduce MeloBottleneck, a self-supervised framework that represents a skeleton as a length-controlled, order-preserving latent subsequence. A hard-bottleneck extractor selects note events, a rhythmic-closure operator produces a self-consistent skeleton, and a re-ornamentation decoder reconstructs the input melody. Training combines reconstruction, a frozen autoregressive melody prior, ornament-invariant consistency across procedurally ornamented views, and ornament exclusion. We evaluate three regimes: synthetic out-of-distribution ornament-to-skeleton, TAVERN variation-to-theme, and Jiugong ornamented-to-gongche. A matched pseudo-label classifier excels on the synthetic benchmark, while MeloBottleneck transfers better, achieving competitive selection quality on TAVERN and Jiugong. Skeletonized melodies also improve BM25-based fragment retrieval, boosting Recall@K and MRR while reducing query time. Overall, the results suggest that learning skeletons as latent subsequences yields more robust transfer than pseudo-label imitation.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 4 linked inside Pith

  1. [1]

    Such skeletons can support symbolic comparison, retrieval, and theme tracing across variants, all tasks where surface elaboration can ob- scure shared melodic material

    INTRODUCTION In monophonic symbolic music, melody skeleton extrac- tion aims to derive a shorter melody that preserves struc- tural notes while removing ornaments. Such skeletons can support symbolic comparison, retrieval, and theme tracing across variants, all tasks where surface elaboration can ob- scure shared melodic material. A practical extractor sh...

  2. [2]

    , eL), melody skeleton extraction aims to derive a shorter melody that preserves structural note events while removing orna- ments

    MELODY SKELETON EXTRACTION: TASK FORMULA TION AND PRIOR WORK 2.1 Task Formulation Given a monophonic symbolic melodyx= (e 1, . . . , eL), melody skeleton extraction aims to derive a shorter melody that preserves structural note events while removing orna- ments. We formulate the skeleton not as a set of indepen- dent note labels, but as an order-preservin...

    arXiv 2026

  3. [3]

    Early computational work by Frankel et al

    offer hierarchical views of structural importance. Early computational work by Frankel et al. [6, 7] encoded Schenkerian elaboration procedures in LISP. Later systems implemented GTTM-style analyzers [2] and probabilistic time-span trees [8]. Previous works [9, 10] further used PCFGs for melodic reduction, Marsden [11] applied auto- matic Schenkerian redu...

  4. [4]

    METHOD: MELOBOTTLENECK Following Section 2, MeloBottleneck treats a melody skeleton as a shorter, order-preserving latent subsequence rather than a set of independent note labels. Given an aug- mented melodyx ′ =A(x), the model selects a length- controlled subsequencez, converts it into a rhythmically closed skeleton¯z, and trains the bottleneck by recons...

    1928

  5. [5]

    File-level splitting (train:valid:test= 18:1:1) and sliding-window segmentation are adopted

    EXPERIMENTAL SETUP 4.1 Data and benchmarks We train on a monophonic symbolic corpus aggre- gated from seven folk-song collections spanning Chinese, British, Irish, Swedish, Dutch, and worldwide repertoires [21–27]. File-level splitting (train:valid:test= 18:1:1) and sliding-window segmentation are adopted. We evaluate on three benchmarks: (a) Main O2B, a ...

  6. [6]

    RESULTS 5.1 Main comparison Table 1 reports Hard F1, CFA, and IM under the oracle ra- tio. On synthetic O2B, the pseudo-label classifier achieves O2B (OOD synth) V2T (cross-domain) O2G (in-domain) Diagnostic lifts on O2B Hard F1 Hard F1 Hard F1 Strong-beat Duration Full 0.8809(±0.0066) 0.6677(±0.0078) 0.8942(±0.0116) 1.8342(±0.0744)1.7878(±0.1122) w/oL R ...

  7. [7]

    DISCUSSION AND LIMITA TIONS Notion of skeleton differs.Melody skeletons are not uniquely defined under different analytical objectives and downstream tasks. MeloBottleneck learns anoperational skeleton induced by our objectives—an order-preserving latent subsequence that (i) retains information for re- ornamentation reconstruction, (ii) forms a standalone...

  8. [8]

    A deterministic rhyth- mic closure operator turns the selected subsequence into a self-consistent reduced melody

    CONCLUSION We introduced MeloBottleneck, a self-supervised frame- work that models a melody skeleton as a length-controlled, order-preserving latent subsequence rather than a set of independent note-wise labels. A deterministic rhyth- mic closure operator turns the selected subsequence into a self-consistent reduced melody. Training combines re-ornamentat...

Show all 39 references
  1. [9]

    All adopted methods were retained only after experimental validation and man- ual inspection

    AI USAGE STA TEMENT AI tools were used in this study to support technical ex- ploration, including surveying candidate methods under given specific design requirements. All adopted methods were retained only after experimental validation and man- ual inspection. AI tools were ...

  2. [10]

    Lerdahl and R

    F. Lerdahl and R. Jackendoff,A Generative Theory of Tonal Music. Cambridge, MA, USA: The MIT Press, 1983

    1983

  3. [11]

    Implementing “a generative theory of tonal music

    M. Hamanaka, K. Hirata, and S. Tojo, “Implementing “a generative theory of tonal music”,”Journal of New Music Research, vol. 35, no. 4, pp. 249–277, 2006

    2006

  4. [12]

    Auto- matic melody reduction via shortest path finding,

    Z. Wang, Y . Wu, R. Dannenberg, and G. Xia, “Auto- matic melody reduction via shortest path finding,” in Proc. of the 26th Int. Society for Music Information Retrieval Conf., Daejeon, South Korea, 2025, pp. 346– 353

    2025

  5. [13]

    A statistical model for melody reduction,

    T. Hu and C. Arthur, “A statistical model for melody reduction,” inProc. of the Future Directions of Music Cognition Int. Conf., Virtual, 2021, pp. 49–53

    2021

  6. [14]

    Schenker,Free Composition (Der freie Satz), ser

    H. Schenker,Free Composition (Der freie Satz), ser. New Musical Theories and Fantasies, E. Oster, Ed. New York: Longman, 1979, vol. 3, original work pub- lished 1935

    1979

  7. [15]

    A lisp- based system for the study of schenkerian analysis,

    R. Frankel, S. Rosenschein, and S. Smoliar, “A lisp- based system for the study of schenkerian analysis,” Computers and the Humanities, vol. 10, pp. 21–32, February 1976

    1976

  8. [16]

    Schenker’s theory of tonal music—its explica- tion through computational processes,

    ——, “Schenker’s theory of tonal music—its explica- tion through computational processes,”International Journal of Man-Machine Studies, vol. 10, no. 2, pp. 121–138, March 1978

    1978

  9. [17]

    σGTTM III: Learning-Based Time-Span Tree Generator Based on PCFG,

    M. Hamanaka, K. Hirata, and S. Tojo, “σGTTM III: Learning-Based Time-Span Tree Generator Based on PCFG,” inProc. of the 11th Int. Symposium on Com- puter Music Multidisciplinary Research, Plymouth, UK, 2015, pp. 303–317

    2015

  10. [18]

    A probabilistic context- free grammar for melodic reduction,

    É. Gilbert and D. Conklin, “A probabilistic context- free grammar for melodic reduction,” inProc. of the Int. Workshop on Artificial Intelligence and Music, in conjunction with the 20th Int. Joint Conf. on Artificial Intelligence (IJCAI), Hyderabad, India, 2007, pp. 83– 94

    2007

  11. [19]

    Automatic melodic reduction using a su- pervised probabilistic context-free grammar,

    R. Groves, “Automatic melodic reduction using a su- pervised probabilistic context-free grammar,” inProc. of the 17th Int. Society for Music Information Retrieval Conf., New York City, USA, 2016, pp. 775–781

    2016

  12. [20]

    Recognition of variations using auto- matic schenkerian reduction,

    A. Marsden, “Recognition of variations using auto- matic schenkerian reduction,” inProc. of the 11th Int. Society for Music Information Retrieval Conf., Utrecht, Netherlands, 2010, pp. 501–506

    2010

  13. [21]

    Melodic skeleton: A mu- sical feature for automatic melody harmonization,

    W. Sun, J. Wu, and S. Yuan, “Melodic skeleton: A mu- sical feature for automatic melody harmonization,” in Proc. of the IEEE Int. Conf. on Multimedia and Expo Workshops, 2022, pp. 1–6

    2022

  14. [22]

    WuYun: Exploring hierarchical skeleton-guided melody generation using knowledge-enhanced deep learning,

    K. Zhang, X. Wu, T. Zhang, Z. Huang, X. Tan, Q. Liang, S. Wu, and L. Sun, “WuYun: Exploring hierarchical skeleton-guided melody generation using knowledge-enhanced deep learning,”arXiv preprint arXiv:2301.04488, 2023

    2023 arXiv

  15. [23]

    Extractive summary as discrete la- tent variables,

    A. Komatsuzaki, “Extractive summary as discrete la- tent variables,”arXiv preprint arXiv:abs/1811.05542, 2018

    2018 arXiv

  16. [24]

    Seq 3: Differentiable sequence-to- sequence-to-sequence autoencoder for unsupervised abstractive sentence compression,

    C. Baziotis, I. Androutsopoulos, I. Konstas, and A. Potamianos, “Seq 3: Differentiable sequence-to- sequence-to-sequence autoencoder for unsupervised abstractive sentence compression,” inProc. of the 2019 Conf. of the North American Chapter of the Associa- tion for Computation...

    2019

  17. [25]

    BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and com- prehension,

    M. Lewis, Y . Liu, N. Goyal, M. Ghazvininejad, A. Mo- hamed, O. Levy, V . Stoyanov, and L. Zettlemoyer, “BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and com- prehension,” inProc. of the 58th Annual Meeting of the Association ...

    2020

  18. [26]

    MusicBERT: Symbolic music understand- ing with large-scale pre-training,

    M. Zeng, X. Tan, R. Wang, Z. Ju, T. Qin, and T.- Y . Liu, “MusicBERT: Symbolic music understand- ing with large-scale pre-training,”arXiv preprint arXiv:2106.05630, 2021

    2021 arXiv

  19. [27]

    PianoBART: Symbolic piano music genera- tion and understanding with large-scale pre-training,

    X. Liang, Z. Zhao, W. Zeng, Y . He, F. He, Y . Wang, and C. Gao, “PianoBART: Symbolic piano music genera- tion and understanding with large-scale pre-training,” inProc. of the IEEE Int. Conf. on Multimedia and Expo (ICME), 2024

    2024

  20. [28]

    Effi- ciently trainable text-to-speech system based on deep convolutional networks with guided attention,

    H. Tachibana, K. Uenoyama, and S. Aihara, “Effi- ciently trainable text-to-speech system based on deep convolutional networks with guided attention,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 4784–4788

    2018

  21. [29]

    FiLM: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. Courville, “FiLM: Visual reasoning with a general conditioning layer,” inProc. of the 32nd AAAI Conf. on Artificial Intelligence, New Orleans, Louisiana, USA, 2018, pp. 3942–3951

    2018

  22. [30]

    The renaissance of expert systems: Optical recog- nition of printed chinese jianpu musical scores with lyrics,

    F. Bu, R. Li, Z. Li, Y . Li, L. Fan, and P. Huang, “The renaissance of expert systems: Optical recog- nition of printed chinese jianpu musical scores with lyrics,”arXiv preprint arXiv:2512.14758, December 2025

    2025

  23. [31]

    Construction and applications of theGongchenotation dataset inJiugong Dacheng Nanbei Ci Gongpu(《九宫大成南北词宫谱》工尺 谱数据集构建及其若干应用),

    R. Li, F. Bu, and Q. Su, “Construction and applications of theGongchenotation dataset inJiugong Dacheng Nanbei Ci Gongpu(《九宫大成南北词宫谱》工尺 谱数据集构建及其若干应用),”Chinese Music (中国 音乐), vol. 04, pp. 129–137, August 2024, in Chinese

    2024

  24. [32]

    BFDB: A dataset of british folk melodies in ABC format,

    A. Deedman and M. Barthet, “BFDB: A dataset of british folk melodies in ABC format,” Dataset, January 2025, version v1

    2025

  25. [33]

    The essen folksong collection in Kern format,

    H. Schaffrath, “The essen folksong collection in Kern format,” [Computer database], Menlo Park, CA, 1995

    1995

  26. [34]

    Henrik norbeck’s ABC tunes,

    H. Norbeck, “Henrik norbeck’s ABC tunes,” https: //www.norbeck.nu/abc/, 2026, last updated 2026-01- 27; accessed 2026-04-24

    2026

  27. [35]

    Tunesformer: Form- ing irish tunes with control codes by bar patching,

    S. Wu, X. Li, F. Yu, and M. Sun, “Tunesformer: Form- ing irish tunes with control codes by bar patching,” in Proceedings of the 2nd Workshop on Human-Centric Music Information Retrieval 2023 co-located with the 24th International Society for Music Information Re- trieval Confe...

    2023

  28. [36]

    The meertens tune collections: MTC-FS-INST 2.0,

    P. van Kranenburg and M. de Bruin, “The meertens tune collections: MTC-FS-INST 2.0,” Meertens In- stitute, Amsterdam, Meertens Online Reports 2019-1, 2019

    2019

  29. [37]

    Theme and variation encodings with roman numerals (TA VERN): A new data set for symbolic mu- sic analysis,

    J. Devaney, C. Arthur, N. Condit-Schultz, and K. Nisula, “Theme and variation encodings with roman numerals (TA VERN): A new data set for symbolic mu- sic analysis,” inProc. of the 16th Int. Society for Music Information Retrieval Conf., Málaga, Spain, 2015, pp. 728–734

    2015

  30. [38]

    Wang,An Anthology of Notated Northern and Southern Ci Music in Nine Modes: A Critical Edition with Commentary (新定九宮大成南北詞宮譜譯註), S.-S

    Z.-L. Wang,An Anthology of Notated Northern and Southern Ci Music in Nine Modes: A Critical Edition with Commentary (新定九宮大成南北詞宮譜譯註), S.-S. Koo and L.-C. Cheung, Eds. Hong Kong: The Chinese University Press and The Chinese University of Hong Kong, 2009, in Chinese; 9 vols.; ISB...

    2009

  31. [39]

    Nine modes manual online,

    Pao Yue-kong Library, The Hong Kong Polytechnic University, “Nine modes manual online,” https://dc.lib. polyu.edu.hk/ninemodes/intro.html, 2021, digital col- lection / online database based onJiugong Dacheng Nanbei Ci Gongpu(九宮大成南北詞宮譜). Accessed: 2026-04-24

    2021

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.