Pith. sign in

REVIEW 3 major objections 6 minor 38 references

Generic pretrained audio embeddings, reduced to one vector per bar and segmented by block matching, can locate song-section boundaries without supervised training, matching or beating linear-probing baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 17:12 UTC pith:QZ44ZKFI

load-bearing objection A solid, reproducible MSA benchmark worth refereeing: the embedding/segmenter ranking is plausible and the trimming analysis is genuinely useful, but the headline comparison against linear probing is not a controlled test of supervision. the 3 major comments →

arxiv 2603.27218 v2 pith:QZ44ZKFI submitted 2026-03-28 cs.SD cs.AIcs.LG

Unsupervised Evaluation of Deep Audio Embeddings for Music Structure Analysis

Axel Marmoret This is my paper
classification cs.SD cs.AIcs.LG
keywords Music Structure Analysisboundary detectionself-supervised audio embeddingsbarwise representationunsupervised segmentationCorrelation Block-Matchingself-similarity matrixevaluation trimming
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper asks whether off-the-shelf, generic pretrained audio models already encode enough musical structure to support unsupervised Music Structure Analysis. It extracts one embedding vector per bar from nine such models, builds self-similarity matrices, and segments them with three unsupervised algorithms, evaluating only boundary retrieval. Its central claim is that this barwise unsupervised pipeline generally outperforms recent linear-probing baselines on three popular-music benchmarks, that the Correlation Block-Matching algorithm is the best downstream segmenter (best in 33 of 36 conditions), and that MATPAC++ is the strongest of the nine embeddings. A secondary claim is that standard evaluation metrics are inflated by trivial first/last boundaries, so the paper advocates trimming or double trimming annotations. The significance, if true, is that high-quality structure analysis does not require structure-specific supervised training on annotated data.

Core claim

On its own terms, the paper's finding is that frozen, generic deep audio embeddings contain structure information that unsupervised segmentation can extract at the bar scale. For almost every embedding-dataset pair, replacing spectrogram features with deep embeddings improves boundary F-scores; however, roughly a third of the models still lose to the spectrogram baseline, so the advantage is not automatic. With barwise representations, Correlation Block-Matching outperforms Foote-style novelty detection and spectral clustering, and the authors report that the barwise+CBM pipeline reaches or exceeds linear-probing baselines on shared models. The paper also shows that trimming the first and la

What carries the argument

The barwise embedding is the interface that makes nine heterogeneous models comparable: audio is cut at estimated downbeats, each model's temporal output is averaged to one vector per bar, and barwise self-similarity matrices are built. The Correlation Block-Matching (CBM) algorithm, a dynamic-programming segmentation method that scores block structure around the diagonal, is the engine that the paper finds most effective. Trimming and double trimming—removing first/last segments, and additionally removing silent lead-in/out regions—form the corrective evaluation mechanism behind the metric-inflation claim.

Load-bearing premise

The whole comparison rests on the assumption that cutting audio into bars estimated by an automatic downbeat tracker and averaging each model's embeddings over each bar is a faithful common representation—the paper explicitly notes this conflicts with earlier temporal-preservation findings, and if barwise averaging or downbeat errors damage structure for some models, the model ranking and the comparison to linear probing could change.

What would settle it

Run the same boundary-retrieval comparison with embeddings computed at the model's original temporal resolution (or at beat level) instead of bar-averaged vectors; if MATPAC++ and CBM no longer dominate, or if the gap over linear probing collapses, the barwise interface—not the embeddings—carried the result.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Deep embeddings can substitute for hand-crafted spectrogram features in unsupervised MSA, improving boundary retrieval on standard pop corpora.
  • CBM should be a default segmenter for barwise embeddings: it wins in 33 of 36 embedding/dataset conditions over checkerboard novelty and spectral clustering.
  • Unsupervised barwise segmentation extracts structure information from generic SSL models at least as well as linear probing, removing the need for a supervised head.
  • The ranking is actionable: masked-prediction continuous embeddings such as MATPAC++ lead, while discrete codec embeddings and music-only SSL models lag.
  • Adopting trimming/double trimming would lower absolute scores on existing benchmarks and make MSA evaluation more rigorous.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the barwise premise holds, the natural next step—hinted at in the conclusion—is to build self-supervised models trained directly at the bar scale and pair them with CBM; the paper suggests this could beat dedicated supervised and SSL baselines.
  • The metric-inflation result implies that previously published state-of-the-art MSA numbers may not be comparable across papers; re-scoring past methods under double trimming could reshuffle leaderboards.
  • Because CBM's advantage is interpreted as homogeneity over novelty, one testable extension is to apply the same barwise+CBM recipe to other sequentially organized media (video scenes, documents, code) using generic deep embeddings.
  • The paper leaves open the temporal-averaging question: re-running the comparison without barwise averaging (or with a model's native resolution) is the direct experiment that would show whether the ranking is a property of the embeddings or of the barwise interface.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents an unsupervised evaluation of nine open-source, generic pre-trained audio embeddings for music structure analysis, focusing on boundary retrieval. Audio is segmented into bars using Beat This!, embeddings are averaged per bar, and three unsupervised segmentation algorithms (Foote, LSD, CBM) are applied to barwise self-similarity matrices. Experiments on RWC-Pop, SALAMI, and Harmonix report hit-rate F-measures at 0.5s and 3s tolerances, with and without annotation trimming. The authors report that deep embeddings generally outperform a barwise time-frequency baseline, that CBM is the best downstream segmentation method in 33 of 36 embedding-dataset conditions, that MATPAC++ is the strongest embedding overall, and that their unsupervised pipeline compares favorably with recent linear-probing baselines from Toyama et al. [15]. They also advocate trimming/double-trimming to avoid artificial metric inflation.

Significance. If the findings are robust, the paper makes a useful contribution: it provides a systematic, reproducible comparison of modern general-purpose audio models on a classic MSA task, releases code, uses standard datasets and metrics, reports standard deviations, and includes robustness appendices. The barwise CBM extension of [7] to deep embeddings and the explicit attention to annotation trimming are both valuable. However, the headline comparisons against linear probing and the model ranking rest on experimental-design choices — confounded pipeline comparison, test-set hyperparameter selection, and a lossy barwise aggregation — that presently limit the strength of the claims. These issues are fixable with additional experiments and a more careful framing, but they are load-bearing for the central conclusions.

major comments (3)
  1. [§4.3, Table 1] The comparison against linear-probing baselines from [15] is confounded. The proposed pipeline differs from the baselines in two variables at once: the input representation (bar-averaged embeddings vs. whatever representation was fed to the linear probe in [15]) and the decoder (CBM vs. linear head). The text itself acknowledges this when it attributes gains to 'the CBM algorithm and barwise processing' rather than to the absence of supervision. Consequently, the abstract's claim that 'our unsupervised boundary estimation methodology generally yields stronger performance than recent linear probing baselines' is not established as a statement about the role of supervision. A matched linear-probe baseline on the same barwise embeddings (or, conversely, CBM on the [15] features) is required to isolate the supervision factor. In addition, the shared-model comparison is only on Harmonix, so '
  2. [§3.1, Figures 1-2, Table 1] Hyperparameters are selected on the evaluation data. Figure 1's caption explicitly describes the 'optimistic condition' where the best-performing hyperparameters per model and dataset are chosen; Table 1 uses hyperparameters selected as 'best on average across all datasets' per model. Both procedures use test labels for model selection, which inflates absolute scores and can change rankings. For example, in Table 1 (fixed CBM hyperparameters) the Barwise TF-CBM baseline reaches 40.02% F0.5 on SALAMI, whereas the best deep embedding (MATPAC++) reaches 38.90%; this contradicts the unqualified §4.1 claim that 'Deep embeddings achieve the highest scores across all datasets and algorithms.' Please report results under a genuinely held-out or label-free hyperparameter-selection rule, or at least show the full distribution over configurations for the central comparisons.
  3. [§3.1] The barwise averaging choice is a possible source of bias for the embedding ranking. The paper explicitly notes that this contradicts the temporal-preservation findings of [15] and declares investigating its impact out of scope. Since the central claims are 'MATPAC++ is best' and 'CBM + barwise processing beats linear probing,' this is not a minor detail: if per-bar averaging discards structure information that is encoded temporally by some models (e.g., MERT), the ranking could change. I recommend a sensitivity experiment with a less lossy temporal aggregation (e.g., beat-level embeddings, or concatenated frame-level statistics) on a subset of models. If that is infeasible, the ranking claims should be explicitly scoped to 'under barwise averaging.'
minor comments (6)
  1. [§4.3] The phrase 'significantly stronger performance' is not supported by significance tests; with standard deviations around 15-20 points, 'numerically stronger' would be more accurate.
  2. [Table 2] For Harmonix the paper says trimming is 'de facto double trimming,' but the 'Trimming' row is empty for Harmonix and only 'Double trimming' is reported. Please clarify the relationship between the two conditions for that dataset.
  3. [Figure 2] The caption says rows are ordered by decreasing average of F0.5s and F3s; it is unclear whether the average is across the two metrics, across datasets, or both. Please state the ordering rule precisely.
  4. [Table 1] The footnote for ⋆ and † is not used in the main body of the table; the literature rows are the relevant ones. Please place the markers consistently.
  5. [§3.1] The hyperparameter grid for Foote is limited to kernel size and median filter in {8,12,16}; the appendix shows robustness for these, but the grid is coarse. A brief justification of the range would be helpful.
  6. [§4.1] The statement 'Deep embeddings achieve the highest scores across all datasets and algorithms' is only valid under the per-dataset optimistic hyperparameter selection of Figure 1; please make this dependence explicit in the main text.

Circularity Check

0 steps flagged

No circularity: empirical benchmark with external baselines; self-citation and the confounded linear-probing comparison are not definitional reductions.

full rationale

The paper makes no formal derivation claims; its contributions are empirical benchmark comparisons. Central results are: deep embeddings vs. barwise TF baselines, CBM vs. Foote vs. LSD, CBM/barwise vs. linear probing from [15], and the effect of trimming. Each is computed from independently fixed inputs (frozen pretrained embeddings, published segmenters, public datasets, fixed metrics) and compared against external literature numbers. The author's own CBM algorithm and barwise framework from [7] are used and cited, and the finding that CBM is best is thus not fully independent in provenance; however, CBM is run as one of three candidate segmenters under a shared grid search, and the ranking is empirically observed (33 of 36 conditions), not imposed by construction. The comparison to linear probing baselines is not fully controlled because the two systems differ in both input representation and decoder, as the paper itself partly acknowledges when it attributes the gain to 'the CBM algorithm and barwise processing'; this is a validity threat, not circularity. Barwise processing is justified by a musical rationale and a prior published result, and no equation is defined in terms of the outcome it predicts. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is imported via self-citation. Hence no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claims rest on the barwise interface, the choice of pretrained models and segmenters, and the evaluation protocol. No new physical entities or mathematical axioms are introduced; the main free parameters are segmentation hyperparameters tuned on the evaluation data.

free parameters (3)
  • Foote kernel size / median filter = searched {8,12,16}
    Selected per model/dataset as best or best-average; controls novelty curve smoothing and affects all Foote scores.
  • LSD number of clusters k = searched {4,6,8,9,10,11,12,13,14,16}
    k controls segmentation granularity; median filtering is set equal to k; the paper notes high variability with this parameter.
  • CBM kernel type and SSM similarity = Full vs 7-band kernel; RBF vs cosine similarity
    Chosen as best per model/dataset (or best average); the CBM penalty enforcing segment sizes is disabled by design, which is an additional modeling choice.
axioms (4)
  • domain assumption Musical structure boundaries align with bar boundaries.
    §2.2: embeddings are computed per bar, justified by musical intuition and prior work [7]; if false, barwise representation loses or shifts boundaries.
  • domain assumption Mean temporal pooling of embeddings preserves structure information.
    §3.1: embeddings are averaged across the temporal dimension to yield one vector per bar; the paper acknowledges this contradicts temporal-preservation findings in [15] and sets the question aside.
  • domain assumption Downbeat estimation by Beat This! is accurate enough for barwise segmentation.
    §3.1: Beat This! is used for all datasets; downbeat errors propagate into bar boundaries and hence into embeddings and predicted segment boundaries.
  • domain assumption Boundary retrieval alone can assess structural information.
    §1 and §3.2: the paper focuses exclusively on boundary retrieval and does not label sections, so conclusions about structure are limited to boundary positions.

pith-pipeline@v1.3.0-alltime-deepseek · 15544 in / 11302 out tokens · 112949 ms · 2026-08-02T17:12:49.982497+00:00 · methodology

0 comments
read the original abstract

Music Structure Analysis (MSA) aims to uncover the high-level organization of musical pieces. State-of-the-art methods are often based on supervised deep learning, but these methods are bottlenecked by the need for heavily annotated data and inherent structural ambiguities. In this paper, we propose an unsupervised evaluation of nine open-source, generic pre-trained deep audio models, on MSA. For each model, we extract barwise embeddings and segment them using three unsupervised segmentation algorithms (Foote's checkerboard kernels, spectral clustering, and Correlation Block-Matching (CBM)), focusing exclusively on boundary retrieval. Our results demonstrate that modern, generic deep embeddings generally outperform traditional spectrogram-based baselines, but not systematically. Furthermore, our unsupervised boundary estimation methodology generally yields stronger performance than recent linear probing baselines. Among the evaluated techniques, the CBM algorithm consistently emerges as the most effective downstream segmentation method. Finally, we highlight the artificial inflation of standard evaluation metrics and advocate for the systematic adoption of ``trimming'', or even ``double trimming'' annotations to establish more rigorous MSA evaluation standards.

Figures

Figures reproduced from arXiv: 2603.27218 by Axel Marmoret.

Figure 1
Figure 1. Figure 1: Comparison of the best results obtained with deep models and the Barwise TF features (non-deep learning [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Best results obtained with all deep learning models, and their best downstream segmentation algorithm. Rows [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Best results obtained with all deep learning models, using the CBM segmentation algorithm. Rows are ordered [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Best results obtained with all deep learning models, using the Foote segmentation algorithm. Rows are ordered [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Best results obtained with all deep learning models, using the LSD segmentation algorithm. Rows are ordered by [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of segmentation scores across downstream segmentation algorithms. Results are restricted to the [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Distribution of segmentation scores across all embeddings for the CBM algorithm, according to various parameter [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Distribution of segmentation scores across all embeddings for the Foote algorithm, according to various parameter [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distribution of segmentation scores across all embeddings for the LSD algorithm, according to various parameter [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 4 linked inside Pith

  1. [1]

    ground truth

    INTRODUCTION Beyond just sound, music is an intricate organization of scales and rhythms. At its core, musical composition re- lies on repetition, contrast, and variation to organize sound into coherent forms. These elements create an internal logic that dictates a song’s shape, establishing recognizable sections separated by distinct transitions. Music S...

    Pith/arXiv arXiv 2026

  2. [2]

    affinity matrices

    METHODOLOGY 2.1 Deep Audio Models To evaluate the structural information captured by modern representation learning, we extract embeddings from nine distinct open-source deep audio models [16–24]. These deep audio models differ significantly in their architec- tures and training objectives. Three of these nine mod- els (MERT [17], MusicFM [18], and MuQ [1...

  3. [3]

    double trimming

    EXPERIMENTAL SETTINGS 3.1 Implementation Details We operate entirely at the bar scale, estimating downbeats using the Beat This! [25] model. To ensure compatibil- ity, raw audio signals are segmented into bar-length chunks before computing embeddings from each model’s final la- tent layer. Because the evaluated models operate at differ- ent native tempora...

  4. [4]

    Deep em- beddings achieve the highest scores across all datasets and algorithms, demonstrating clear improvements over spectrogram-based features

    RESULTS AND DISCUSSION 4.1 Deep Modelsvs.Barwise TF features Figure 1 compares peak deep audio models’ perfor- mance against the Barwise TF baseline. Deep em- beddings achieve the highest scores across all datasets and algorithms, demonstrating clear improvements over spectrogram-based features. Table 1: Deep model segmentation performance across datasets...

    1959

  5. [5]

    CONCLUSION In this work, we investigated the training-free capabili- ties of nine generic deep audio models for Music Struc- ture Analysis, specifically focusing on boundary retrieval. By leveraging unsupervised downstream segmentation al- gorithms at the bar scale, we demonstrated that deep au- dio embeddings generally provide a distinct advantage over t...

  6. [6]

    Audio-based music structure analy- sis: Current trends, open challenges, and applications,

    O. Nietoet al., “Audio-based music structure analy- sis: Current trends, open challenges, and applications,” Trans. Int. Soc. for Music Information Retrieval, vol. 3, no. 1, 2020

    2020

  7. [7]

    Reductive, exclusionary, normalising: the limits of generative ai music,

    F. Morreale, M. A. Martinez-Ramirez, R. Masu, W. Liao, and Y . Mitsufuji, “Reductive, exclusionary, normalising: the limits of generative ai music,”Trans. Int. Soc. Music Information Retrieval, vol. 8, no. 1, 2025

    2025

  8. [8]

    Under- standing the potentials and limitations of prompt-based music generative ai,

    Y . Choi, J. Moon, J. Yoo, and J.-H. Hong, “Under- standing the potentials and limitations of prompt-based music generative ai,” inProceedings of the 2025 CHI Conference on Human Factors in Computing Systems, 2025

    2025

  9. [9]

    Automatic audio segmentation using a mea- sure of audio novelty,

    J. Foote, “Automatic audio segmentation using a mea- sure of audio novelty,” in2000 IEEE Int. Conf. Multi- media and Expo. ICME2000.IEEE, 2000, pp. 452– 455

    2000

  10. [10]

    Un- supervised music structure annotation by time series structure features and segment similarity,

    J. Serr `a, M. M ¨uller, P. Grosche, and J. L. Arcos, “Un- supervised music structure annotation by time series structure features and segment similarity,”IEEE Trans. Multimedia, vol. 16, no. 5, pp. 1229–1240, 2014

    2014

  11. [11]

    Analyzing song structure with spectral clustering,

    B. McFee and D. Ellis, “Analyzing song structure with spectral clustering,” inInt. Soc. Music Information Re- trieval Conf. (ISMIR), 2014, pp. 405–410

    2014

  12. [12]

    Barwise music structure analysis with the correlation block- matching segmentation algorithm,

    A. Marmoret, J. E. Cohen, and F. Bimbot, “Barwise music structure analysis with the correlation block- matching segmentation algorithm,”Trans. Int. Soc. for Music Information Retrieval, vol. 6, no. 1, pp. 167– 185, 2023

    2023

  13. [13]

    Unsupervised learning of deep fea- tures for music segmentation,

    M. C. McCallum, “Unsupervised learning of deep fea- tures for music segmentation,” in2019 IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 346–350

    2019

  14. [14]

    Deep embed- dings and section fusion improve music segmentation,

    J. Salamon, O. Nieto, and N. J. Bryan, “Deep embed- dings and section fusion improve music segmentation,” inInt. Soc. Music Information Retrieval Conf. (ISMIR), 2021

    2021

  15. [15]

    Su- pervised metric learning for music structure feature,

    J.-C. Wang, J. B. Smith, W.-T. Lu, and X. Song, “Su- pervised metric learning for music structure feature,” inInt. Soc. Music Information Retrieval Conf. (ISMIR), 2021, pp. 730–737

    2021

  16. [16]

    Self-supervised learning of multi-level audio repre- sentations for music segmentation,

    M. Buisson, B. McFee, S. Essid, and H. C. Crayencour, “Self-supervised learning of multi-level audio repre- sentations for music segmentation,”IEEE/ACM Trans. on Audio, Speech, and Language Processing, vol. 32, pp. 2141–2152, 2024

    2024

  17. [17]

    Music boundary detection us- ing neural networks on combined features and two- level annotations,

    T. Grill and J. Schl ¨uter, “Music boundary detection us- ing neural networks on combined features and two- level annotations,” inInt. Soc. Music Information Re- trieval Conf. (ISMIR), 2015, pp. 531–537

    2015

  18. [18]

    All-in-one metrical and func- tional structure analysis with neighborhood attentions on demixed audio,

    T. Kim and J. Nam, “All-in-one metrical and func- tional structure analysis with neighborhood attentions on demixed audio,” in2023 IEEE Workshop on Appli- cations of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2023

    2023

  19. [19]

    Using pairwise link prediction and graph attention networks for mu- sic structure analysis,

    M. Buisson, B. Mcfee, and S. Essid, “Using pairwise link prediction and graph attention networks for mu- sic structure analysis,” inInt. Soc. Music Information Retrieval Conf. (ISMIR), 2024

    2024

  20. [20]

    Do foundational audio encoders un- derstand music structure?

    K. Toyamaet al., “Do foundational audio encoders un- derstand music structure?” in2026 IEEE Int. Conf. Acoustics, Speech and Signal Processing (ICASSP), 2026

    2026

  21. [21]

    Masked autoencoders that listen,

    P.-Y . Huanget al., “Masked autoencoders that listen,” vol. 35, 2022, pp. 28 708–28 720

    2022

  22. [22]

    Mert: Acoustic music understanding model with large-scale self-supervised training,

    Y . Liet al., “Mert: Acoustic music understanding model with large-scale self-supervised training,”arXiv preprint arXiv:2306.00107, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    A foundation model for music informatics,

    M. Won, Y .-N. Hung, and D. Le, “A foundation model for music informatics,” in2024 IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 1226–1230

    2024

  24. [24]

    Muq: Self-supervised music represen- tation learning with mel residual vector quantization,

    H. Zhuet al., “Muq: Self-supervised music represen- tation learning with mel residual vector quantization,” IEEE Trans. Audio, Speech and Language Processing, 2025

    2025

  25. [25]

    M2d-clap: Masked modeling duo meets clap for learning general-purpose audio- language representation

    D. Niizumiet al., “M2d-clap: Masked modeling duo meets clap for learning general-purpose audio- language representation.” ISCA, 2024, pp. 57–61

    2024

  26. [26]

    Matpac++: Enhanced masked latent prediction for self-supervised audio representation learning,

    A. Quelennec, P. Chouteau, G. Peeters, and S. Essid, “Matpac++: Enhanced masked latent prediction for self-supervised audio representation learning,”arXiv preprint arXiv:2508.12709, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    High-fidelity audio compression with im- proved rvqgan,

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, “High-fidelity audio compression with im- proved rvqgan,”Advances in Neural Information Pro- cessing Systems, vol. 36, pp. 27 980–27 993, 2023

    2023

  28. [28]

    Codicodec: Unifying continuous and discrete compressed repre- sentations of audio,

    M. Pasini, S. Lattner, and G. Fazekas, “Codicodec: Unifying continuous and discrete compressed repre- sentations of audio,”arXiv preprint arXiv:2509.09836, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,

    Y . Wuet al., “Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,” in2023 IEEE Int. Conf. Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023

    2023

  30. [30]

    Beat this! accurate beat tracking without dbn postprocessing,

    F. Foscarin, J. Schl ¨uter, and G. Widmer, “Beat this! accurate beat tracking without dbn postprocessing,” in Int. Soc. Music Information Retrieval Conf. (ISMIR), 2024

    2024

  31. [31]

    Systematic exploration of computational music structure research,

    O. Nieto and J. P. Bello, “Systematic exploration of computational music structure research,” inInt. Soc. Music Information Retrieval Conf. (ISMIR), 2016, pp. 547–553

    2016

  32. [32]

    Pytorch: An imperative style, high- performance deep learning library,

    A. Paszkeet al., “Pytorch: An imperative style, high- performance deep learning library,”Advances in neural information processing systems, vol. 32, 2019

    2019

  33. [33]

    mir eval: A transparent implementa- tion of common MIR metrics,

    C. Raffelet al., “mir eval: A transparent implementa- tion of common MIR metrics,” inInt. Soc. Music Infor- mation Retrieval Conf. (ISMIR), 2014, pp. 367–372

    2014

  34. [34]

    mirdata: Software for repro- ducible usage of datasets

    R. M. Bittneret al., “mirdata: Software for repro- ducible usage of datasets.” inInt. Soc. Music Informa- tion Retrieval (ISMIR), 2019, pp. 99–106

    2019

  35. [35]

    RWC Music Database: Popular, Classical and Jazz Music Databases,

    M. Goto, H. Hashiguchi, T. Nishimura, and R. Oka, “RWC Music Database: Popular, Classical and Jazz Music Databases,” inInt. Soc. Music Information Re- trieval Conf. (ISMIR), 2002, pp. 287–288

    2002

  36. [36]

    Rwc revisited: Towards a community- driven mir corpus,

    S. Balkeet al., “Rwc revisited: Towards a community- driven mir corpus,”Trans. Int. Soc. for Music Informa- tion Retrieval, vol. 9, no. 1, 2026

    2026

  37. [37]

    Design and creation of a large-scale database of structural annotations,

    J. B. Smithet al., “Design and creation of a large-scale database of structural annotations,” inInt. Soc. Music Information Retrieval Conf. (ISMIR), 2011, pp. 555– 560

    2011

  38. [38]

    The harmonix set: Beats, downbeats, and functional segment annotations of western popular music,

    O. Nietoet al., “The harmonix set: Beats, downbeats, and functional segment annotations of western popular music,” inInt. Soc. Music Information Retrieval Conf. (ISMIR), 2019, pp. 565–572. A. SCORES ACCORDING TO THE DOWNSTREAM SEGMENTA TION ALGORITHM In this section, we present three figures: one for each downstream segmentation algorithm (Figure 3 for CB...

    2019