Pith. sign in

REVIEW 4 major objections 6 minor 57 references

Learning Normal Patterns in Musical Loops

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper shows that 'normal' patterns in musical loops can be learned without labels, and that a residual-connection Deep SVDD model separates unusual bass and guitar loops from typical ones more cleanly than Isolation Forest or PCA…

desk verdict A clean application of existing components to loop anomaly detection, but the headline claim about improved separation is not actually tested: the evaluation is qualitative, circular, and confounded by overfitting. read the letter →

arxiv 2505.23784 v1 pith:MKYCT5JL submitted 2025-05-22 cs.SD cs.IRcs.LGcs.MMeess.AS

classification cs.SDcs.IRcs.LGcs.MMeess.AS
keywords audioloopdetectionunsupervisedanomalyDeepSVDDmusicalloopsrepresentationlearningHTS-ATfeaturefusionmechanismmusicinformationretrieval
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

This paper tries to establish that the concept of a 'normal' musical loop can be learned from unlabeled audio, so that unusual loops can be flagged automatically. It combines a pre-trained audio transformer (HTS-AT) with a Feature Fusion Mechanism that turns variable-length bass and guitar clips into fixed-size embeddings, then trains a Deep SVDD network to enclose the embeddings of typical loops inside a compact hypersphere. The paper reports that Deep SVDD, and especially an autoencoder with residual connections, separates normal from anomalous loops more cleanly than Isolation Forest or PCA reconstruction error. If this holds, music producers and AI co-creation systems gain a label-free way to spot odd samples, production errors, or serendipitous outliers in their own collections.

What carries the argument

The load-bearing object is the Deep Support Vector Data Description (Deep SVDD) module: a five-layer encoder trained in two phases, first as an autoencoder on the frozen HTS-AT embeddings and then under the one-class objective of pulling normal embeddings toward a hypersphere center $c$. The anomaly score for any loop is its squared Euclidean distance $\|\phi(x) - c\|^2$ from that center. Supporting this, the Feature Fusion Mechanism (FFM) merges global and local temporal views of long clips and pads short clips so that variable-length audio still produces fixed-size inputs for the pre-trained Hierarchical Token-semantic Audio Transformer (HTS-AT), which is what lets the method escape the fixed-length constraint of earlier loop models.

What would settle it

Take a held-out set of clean bass or guitar loops, deliberately corrupt some of them with clicks, phase flips, splice errors, or out-of-genre samples, and check whether corrupt loops systematically receive higher AEwRES anomaly scores than clean ones; if the score distributions overlap, the separation claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that framing loop detection as unsupervised anomaly detection works when the audio is first converted into rich, fixed-dimension embeddings by a frozen pre-trained model. Trained only on unlabeled bass or guitar loops, the Deep SVDD module learns a hypersphere that captures the dataset's normative patterns; any loop whose embedding lands far from the hypersphere center is scored as anomalous. On curated bass and guitar datasets, the authors found that both Deep SVDD variants separate normal from anomalous scores better than Isolation Forest and PCA reconstruction, and that the residual-connection autoencoder (AEwRES) gives the cleanest separation and the most structured latent space, particularly on the more diverse guitar data. The conclusion they draw is that residual connections help the encoder represent pattern diversity without over-compressing the normal manifold, making AEwRES the more promising architecture for loop detection in varied musical contexts.

Load-bearing premise

The evaluation assumes that the 95th percentile of training anomaly scores is a valid boundary between normal and anomalous loops, and that the visible separation between training and validation score histograms reflects genuine detection rather than overfitting to the training set.

Editorial extensions

If this is right

  • Variable-length loops can be analyzed directly, removing the fixed-duration preprocessing required by earlier loop-compatibility models.
  • Loop collections can be audited without labels: distance-to-center scores can flag clicks, phase issues, or unintended noise for quality control.
  • The residual-connection autoencoder becomes the better default Deep SVDD encoder when the normal class is internally diverse, as it is with multi-style guitar loops.
  • The same distance score can drive creative discovery by surfacing loops that deviate from a user's existing library.

Reading between the lines

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

  • The q=0.95 threshold is likely less important than the ranking quality of the distance scores; with labeled anomalies available, the same model could be evaluated by ROC/AUC and the threshold tuned to a user's tolerance.
  • Because the audio encoder is frozen, the method should transfer cheaply to other loop-like units such as drums, synth one-shots, or vocal chops, and the residual-connection benefit would plausibly grow as intra-class diversity grows.
  • A natural next experiment is to fine-tune HTS-AT on the loop domain instead of freezing it, and compare whether loop-specific embeddings improve separation or simply overfit the training collection.
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 / 6 minor

Summary. This paper presents an unsupervised framework for detecting anomalous musical loops by extracting fixed-dimensional embeddings from variable-length audio via a pre-trained HTS-AT and a Feature Fusion Mechanism, then training a Deep SVDD model (standard autoencoder AE or residual autoencoder AEwRES) on the embeddings. The framework is evaluated on curated bass and guitar loop datasets against Isolation Forest and PCA reconstruction-error baselines. The evaluation is entirely qualitative: anomaly thresholds are set at the 95th percentile of training anomaly scores, and the reported 'improved separation' of the Deep SVDD variants, especially AEwRES, is inferred from histograms of anomaly scores, PCA scatter plots, and latent-space heatmaps (Figures 4-6, Appendix C). The authors acknowledge in Section 8 that no ground-truth anomaly labels are available.

Significance. If the claims were substantiated, the proposed pipeline would be a useful contribution to MIR: it addresses the variable-length input limitation with FFM, leverages a strong pre-trained audio encoder, and proposes an unsupervised, user-collection-adaptable anomaly detection tool; the comparison of AE and AEwRES architectures within Deep SVDD is also a reasonable design question. The strengths of the paper include the clear description of the two-stage training procedure (AE pre-training then Deep SVDD fine-tuning), the use of a frozen HTS-AT encoder with FFM to handle variable durations, and the inclusion of baseline methods on the same embeddings. However, the paper's headline claim of improved anomaly separation is not supported by the evidence presented: there is no quantitative evaluation with labeled or synthetic anomalies, no AUROC or precision/recall, no statistical test of distribution separation, and the threshold-based labeling makes 'anomaly' a self-referential construct. The central claim is therefore currently unverified.

major comments (4)
  1. [§5.4, §6.1–6.2] The evaluation is circular: anomalies are defined as samples whose anomaly score exceeds the 95th percentile (q=0.95) of training scores, so any validation sample above this threshold is 'anomalous' by construction, regardless of musical or structural distinctiveness. The paper provides no external or synthetic anomalies and no labeled set, which makes the abstract claim of 'improved anomaly separation' untestable from the reported results. Please add a quantitative evaluation using labeled anomalies (e.g., loops from other instruments or genres, artificially corrupted loops, or non-loop audio) and report AUROC, average precision, or at least a chance-baseline comparison.
  2. [§4.3, Figures 4–5] The observed separation between training and validation score histograms is exactly what overfitting predicts for Deep SVDD: the network is trained to map training embeddings close to the hypersphere center, and with early stopping monitoring validation loss, unseen validation points will tend to receive higher scores even if they are perfectly normal loops. The paper needs a control condition, such as scoring a second set of normal loops from a different source, or comparing the validation score distribution against a permutation baseline, to show that the gap indicates anomaly detection rather than distribution shift.
  3. [§6.2–6.3] The comparison between AE and AEwRES is confounded by the use of raw score thresholds that differ substantially across models (e.g., guitar threshold is about 0.001 for AE and 0.035 for AEwRES, Figure 5). Claiming that AEwRES is 'more promising' because its latent space has higher explained variance in the first two PCs (93.2% vs. 77.7%) conflates representation spread with anomaly detection quality; explained variance does not measure how well anomalies are separated from normal data. Please report threshold-independent metrics, such as ROC curves computed with labeled anomalies, or a standardized overlap statistic (e.g., Kolmogorov-Smirnov) between normal and anomalous score distributions.
  4. [§4.1] No separate test set is reserved; hyperparameters and early stopping are selected using the validation set, and the same validation set is used for all reported analysis. This makes the reported separation figures optimistic and prevents assessment of generalization. A held-out test set (or cross-validation) should be used for the final evaluation.
minor comments (6)
  1. [Abstract] The abstract contains a typo: 'and and principle component analysis' should be 'and principal component analysis'.
  2. [§6.3] The heading 'Compartive Performance Summary' should be 'Comparative Performance Summary'.
  3. [§3.1] The sentence 'The resulting fused map, compromising both broad structure and specific details' should use 'comprising' instead of 'compromising'.
  4. [Appendices B–C] Minor typos appear in the appendices: 'witout' should be 'without' (Figure 7 caption and Figure 8 caption), and 'validatio' should be 'validation' (Figures 9 and 10 captions).
  5. [§3.1, §3.2] References to 'Wu et al.' and 'Dai et al.' appear in the text without citation numbers in several places; please standardize the citation style.
  6. [Figures 4–5] The anomaly-threshold labels in the histograms are small and hard to read; consider annotating with larger text and using comparable axis scales across models to facilitate cross-model comparison.

Circularity Check

2 steps flagged · score 6.0 of 10

Anomaly 'separation' claims rest on a threshold fit to the model's own scores, with no ground-truth labels; the central evaluation is self-definitional.

  1. self definitional [Section 6 (anomaly threshold definition in Section 5.4)]
    "For the subsequent analysis, potential anomalies are defined as samples whose anomaly score exceeds the 95th percentile (q=0.95) threshold derived from the training set scores."

    The evaluation first fits a threshold as the 95th percentile of the model's own training anomaly scores, then colors points and interprets 'normal' vs 'anomaly' using that same threshold. Any conclusion about 'separation' between these two groups is therefore a restatement of the score distribution relative to a quantile of itself; no independent anomaly labels or external benchmark enter. The central claim that AEwRES 'deliver[s] improved anomaly separation' reduces to the geometry of the fitted scores, not to validated detection of musically distinct loops.

  2. fitted input called prediction [Appendix B]
    "Anomaly detection performance indicated successful separation of normal samples and potential anomalies in both architectures; however, AEwRES achieved a lower anomaly threshold (q=0.95 ≈ 0.009) compared to AE (≈ 0.016), potentially indicating enhanced precision."

    Precision is a label-based notion, but no labels exist. The q=0.95 threshold is fit to each model's training scores, so a lower threshold only means the training scores are more concentrated around the learned center. Interpreting that fitted quantile as 'enhanced precision' treats the model's own score concentration as if it were measured detection accuracy, which is exactly a fitted input being renamed as a predictive performance claim.

full rationale

The framework itself is not circular: HTS-AT/FFM and Deep SVDD are adopted from external work (Ruff et al., Chen et al., Wu et al.), the frozen-encoder pipeline is standard, and the self-citations in Sections 1 and 7.1 are motivational rather than load-bearing. The circularity sits in the evaluation. Section 5.4 sets the anomaly threshold at q=0.95 of the training scores; Section 6 then defines 'potential anomalies' as points above that threshold and uses the resulting color-coded histograms and PCA plots to conclude 'improved anomaly separation' (Section 6.3). Since the labels are constructed from the model scores being evaluated, the headline result is self-definitional. Appendix B compounds this by interpreting a lower fitted q95 threshold as 'enhanced precision.' The authors' own Section 8 concedes 'we lack ground-truth labels for definitive anomalous loops,' which confirms that no external validation anchors the claim. The observed train/validation score gap is also what overfitting predicts under the Deep SVDD training objective, but that is a correctness risk rather than a distinct circular step. Overall, there is partial circularity (6/10): the architecture and representations have independent content, but the central 'anomaly separation' claim reduces to a self-fitted threshold.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the suitability of frozen HTS-AT embeddings, the validity of the 95th-percentile threshold, and the representativeness of the curated MusicRadar dataset. No new entities are introduced; the method is a composition of existing components.

free parameters (4)
  • Hypersphere center c = Learned from training embeddings
    Center of the Deep SVDD hypersphere, fit to the training set via the objective in Eq. (1). All anomaly scores are distances to this learned center.
  • Anomaly threshold q = 0.95
    95th percentile of training anomaly scores, chosen by hand as the separation boundary (Section 5.4). The authors note in Section 8 that its optimality is data-dependent.
  • Latent dimension = 32
    Dimensionality of the Deep SVDD latent space, selected during preliminary tuning (Appendix B). This affects the expressiveness of the learned representation.
  • FFM chunk duration d = 10 seconds
    Fixed segment length for the feature fusion mechanism, taken from Wu et al. (2024). It determines the repeat-pad and downsampling behavior for variable-length inputs.
assumptions (4)
  • domain assumption Pre-trained HTS-AT embeddings capture enough musical structure to distinguish normal from anomalous loops
    The entire pipeline rests on the frozen HTS-AT features being discriminative for loop patterns. No fine-tuning or validation of this assumption is provided.
  • domain assumption The 95th percentile training threshold defines a meaningful anomaly boundary
    Figures 4 and 5 use q=0.95 as the threshold, and the authors acknowledge in Section 8 that its optimality is data-dependent and interpretation is relative to the training distribution.
  • domain assumption The curated MusicRadar bass and guitar sets are representative of normal loop patterns
    Section 4.1 manually filters for clear rhythmic structure and typical patterns, but the dataset is not shared and may carry selection biases (e.g., unknown genre and key metadata).
  • standard math Deep SVDD's single hypersphere can model the normal distribution
    The framework from Ruff et al. (2018) is used as is. The paper notes in Section 8 that complex multi-modal distributions may not be perfectly captured by a single hypersphere.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Normal Patterns in Musical Loops." pith.science (2026). https://pith.science/paper/MKYCT5JL

@misc{pith2026250523784,
  author       = {Pith},
  title        = {Pith review of: Learning Normal Patterns in Musical Loops},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MKYCT5JL}},
  note         = {Machine review of arXiv:2505.23784}
}
read the original abstract

This paper introduces an unsupervised framework for detecting audio patterns in musical samples (loops) through anomaly detection techniques, addressing challenges in music information retrieval (MIR). Existing methods are often constrained by reliance on handcrafted features, domain-specific limitations, or dependence on iterative user interaction. We address these limitations through an architecture combining deep feature extraction with unsupervised anomaly detection. Our approach leverages a pre-trained Hierarchical Token-semantic Audio Transformer (HTS-AT), paired with a Feature Fusion Mechanism (FFM), to generate representations from variable-length audio loops. These embeddings are processed using one-class Deep Support Vector Data Description (Deep SVDD), which learns normative audio patterns by mapping them to a compact latent hypersphere. Evaluations on curated bass and guitar datasets compare standard and residual autoencoder variants against baselines like Isolation Forest (IF) and and principle component analysis (PCA) methods. Results show our Deep SVDD models, especially the residual autoencoder variant, deliver improved anomaly separation, particularly for larger variations. This research contributes a flexible, fully unsupervised solution for processing diverse audio samples, overcoming previous structural and input limitations while enabling effective pattern identification through distance-based latent space scoring.

Figures

Figures reproduced from arXiv: 2505.23784 by the authors.

Figure 1
Figure 1. Architectural overview of the proposed loop detection model. The model comprises two [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Distribution of curated guitar and bass samples across (a) musical genres and (b) musical [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. The architecture of the autoencoder (AE) network with residual connections (dotted arrows) (AEwRES). The model follows a symmetric design (inspired by U-net architecture [Ronneberger et al., 2015]) with an encoder (left) and decoder (right). The input dimension of 1024 is progressively reduced through five encoder layers and then reconstructed through five decoder layers. Each layer comprises a linear transformation… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The figure displays the anomaly score distribution histograms for models applied to the [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: The figure displays anomaly score distribution histograms for models applied to the Guitar [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: The figure displays box plots comparing anomaly score distributions across models (A [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: The figure displays anomaly score distribution histograms for [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: The figure presents three visualizations of latent space representations from [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: Comparative analysis of (a,b) AE and (c,d) AEwRES models using bass dataset, as described in Section 4.1. The first two rows (a,b) show the results obtained by AE model during (a) training and (b) validatio phases. The second two rows (c,d) displays the results obtaine…
Figure 10
Figure 10. Figure 10: Comparative analysis of (a,b) AE and (c,d) AEwRES models using guitar dataset, as described in Section 4.1. The first two rows (a,b) show the results obtained by AE model during (a) training and (b) validatio phases. The second two rows (c,d) displays the results obta…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 33 canonical work pages

  1. [1]

    Electronic music

    Nick Collins, Margaret Schedel, and Scott Wilson. Electronic music. Cambridge introductions to music. University Press, Cambridge, 1. publ edition, 2014. ISBN 978-1-107-64817-3

  2. [2]

    Mark J. Butler. Unlocking the groove: rhythm, meter, and musical design in electronic dance music. Profiles in popular music. Indiana University Press, Bloomington, 2006. ISBN 978-0-253-34662-9 978-0-253-21804-9. OCLC: ocm61162077

  3. [3]

    Reese, Lynne S

    David E. Reese, Lynne S. Gross, and Brian Gross. Audio production worktext: concepts, techniques, and equipment. Elsevier Focal Press, Burlington, MA, 6th ed edition, 2009. ISBN 978-0-240-81098-0. OCLC: 255903242

  4. [4]

    Bill Gibson. The S . M . A . R . T . guide to producing music with samples, loops, and MIDI . Thomson Course Technology PTR : Artistpro Pub., Boston, MA, 2005. ISBN 978-1-59200-697-7. OCLC: 61519237

  5. [5]

    A music loop explorer system

    Sebastian Streich and Bee Suan Ong. A music loop explorer system. In ICMC, 2008

  6. [6]

    A loop sequencer that selects music loops based on the degree of excitement

    Tetsuro Kitahara, Kosuke Iijima, Misaki Okada, Yuji Yamashita, and Ayaka Tsuruoka. A loop sequencer that selects music loops based on the degree of excitement. In Proceedings of the 12th Sound and Music Computing Conference (SMC 2015), pages 435--438, 2015

  7. [7]

    Bo-Yu Chen, Jordan B. L. Smith, and Yi-Hsuan Yang. Neural Loop Combiner : Neural Network Models for Assessing the Compatibility of Loops , 2020. URL https://arxiv.org/abs/2008.02011. Version Number: 2

  8. [8]

    On the Development and Practice of AI Technology for Contemporary Popular Music Production

    Emmanuel Deruty, Maarten Grachten, Stefan Lattner, Javier Nistal, and Cyran Aouameur. On the Development and Practice of AI Technology for Contemporary Popular Music Production . 5 0 (1): 0 35--49, 2022. ISSN 2514-3298. doi:10.5334/tismir.100. URL https://transactions.ismir.net/articles/10.5334/tismir.100?utm_source=TrendMD&utm_medium=cpc&utm_campaign=Tra...

Show all 57 references
  1. [9]

    Cheng-Zhi Anna Huang, Hendrik Vincent Koops, Ed Newton-Rex, Monica Dinculescu, and Carrie J. Cai. AI Song Contest : Human-AI Co-Creation in Songwriting , 2020. URL http://arxiv.org/abs/2010.05388

  2. [10]

    Toward Interactive Music Generation : A Position Paper

    Shayan Dadman, Bernt Arild Bremdal, Børre Bang, and Rune Dalmo. Toward Interactive Music Generation : A Position Paper . 10: 0 125679--125695, 2022. ISSN 2169-3536. doi:10.1109/ACCESS.2022.3225689. URL https://ieeexplore.ieee.org/document/9966445/

  3. [11]

    Crafting Creative Melodies : A User-Centric Approach for Symbolic Music Generation

    Shayan Dadman and Bernt Arild Bremdal. Crafting Creative Melodies : A User-Centric Approach for Symbolic Music Generation . 13 0 (6): 0 1116, 2024. ISSN 2079-9292. doi:10.3390/electronics13061116. URL https://www.mdpi.com/2079-9292/13/6/1116

  4. [12]

    Multi-agent Reinforcement Learning for Structured Symbolic Music Generation

    Shayan Dadman and Bernt Arild Bremdal. Multi-agent Reinforcement Learning for Structured Symbolic Music Generation . In Philippe Mathieu, Frank Dignum, Paulo Novais, and Fernando De La Prieta, editors, Advances in Practical Applications of Agents , Multi - Agent Systems , and ...

  5. [13]

    Escalona

    Miguel Civit, Javier Civit-Masot, Francisco Cuadrado, and Maria J. Escalona. A systematic review of artificial intelligence-based music generation: Scope , applications, and future trends. 209: 0 118190, 2022. ISSN 09574174. doi:10.1016/j.eswa.2022.118190. URL https://linkingh...

  6. [14]

    A Survey of Audio-Based Music Classification and Annotation

    Zhouyu Fu, Guojun Lu, Kai Ming Ting, and Dengsheng Zhang. A Survey of Audio-Based Music Classification and Annotation . 13 0 (2): 0 303--319, 2011. ISSN 1520-9210, 1941-0077. doi:10.1109/TMM.2010.2098858. URL http://ieeexplore.ieee.org/document/5664796/

  7. [15]

    Ras and Alicja A

    Zbigniew W. Ras and Alicja A. Wieczorkowska, editors. Advances in Music Information Retrieval , volume 274 of Studies in Computational Intelligence . Springer Berlin Heidelberg, 2010. ISBN 978-3-642-11673-5 978-3-642-11674-2. doi:10.1007/978-3-642-11674-2. URL https://link.spr...

  8. [16]

    Plumbley

    Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D. Plumbley. PANNs : Large - Scale Pretrained Audio Neural Networks for Audio Pattern Recognition , August 2020. URL http://arxiv.org/abs/1912.10211. arXiv:1912.10211 [cs, eess]

  9. [17]

    Zero- Shot Audio Classification via Semantic Embeddings , 2020

    Huang Xie and Tuomas Virtanen. Zero- Shot Audio Classification via Semantic Embeddings , 2020. URL https://arxiv.org/abs/2011.12133. Version Number: 2

  10. [18]

    Gemmeke, Daniel P

    Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio Set : An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference on Acoustics , Speech and Signal Proce...

  11. [19]

    Automatic tagging using deep convolutional neural networks, 2016

    Keunwoo Choi, George Fazekas, and Mark Sandler. Automatic tagging using deep convolutional neural networks, 2016. URL https://arxiv.org/abs/1606.00298. Version Number: 1

  12. [20]

    A Comprehensive Review on Audio based Musical Instrument Recognition : Human-Machine Interaction Towards Industry 4.0

    Dash, Sukanta Kumar , Solanki, S S , and Chakraborty, Soubhik . A Comprehensive Review on Audio based Musical Instrument Recognition : Human-Machine Interaction Towards Industry 4.0. 82 0 (01), 2023. ISSN 00224456, 09751084. doi:10.56042/jsir.v82i1.70251. URL http://op.niscair...

  13. [21]

    Large-scale Contrastive Language - Audio Pretraining with Feature Fusion and Keyword -to- Caption Augmentation , March 2024

    Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Marianna Nezhurina, Taylor Berg-Kirkpatrick, and Shlomo Dubnov. Large-scale Contrastive Language - Audio Pretraining with Feature Fusion and Keyword -to- Caption Augmentation , March 2024. URL http://arxiv.org/abs/2211.06687. arXiv...

  14. [22]

    Repeating pattern discovery and structure analysis from acoustic music data

    Lie Lu, Muyuan Wang, and Hong-Jiang Zhang. Repeating pattern discovery and structure analysis from acoustic music data. In Proceedings of the 6th ACM SIGMM international workshop on Multimedia information retrieval , pages 275--282, New York NY USA, October 2004. ACM. ISBN 978...

  15. [23]

    Structural analysis and segmentation of music signals

    Bee Suan Ong et al. Structural analysis and segmentation of music signals. Citeseer, 2006

  16. [24]

    State of the art report: Audio-based music structure analysis

    Jouni Paulus, Meinard M \"u ller, and Anssi Klapuri. State of the art report: Audio-based music structure analysis. 2010

  17. [25]

    Mysore, Cheng-i Wang, Jordan B

    Oriol Nieto, Gautham J. Mysore, Cheng-i Wang, Jordan B. L. Smith, Jan Schlüter, Thomas Grill, and Brian McFee. Audio- Based Music Structure Analysis : Current Trends , Open Challenges , and Applications . Transactions of the International Society for Music Information Retrieva...

  18. [26]

    Symbolic Music Loop Generation with Neural Discrete Representations , October 2022

    Sangjun Han, Hyeongrae Ihm, Moontae Lee, and Woohyung Lim. Symbolic Music Loop Generation with Neural Discrete Representations , October 2022. URL http://arxiv.org/abs/2208.05605. arXiv:2208.05605 [cs, eess]

  19. [27]

    Searching For Loops And Sound Samples With Feature Learning

    Jan Jakubik. Searching For Loops And Sound Samples With Feature Learning . pages 13--18, September 2022. doi:10.15439/2022F279. URL https://annals-csis.org/Volume_31/drp/279.html

  20. [28]

    HTS - AT : A Hierarchical Token - Semantic Audio Transformer for Sound Classification and Detection , 2022

    Ke Chen, Xingjian Du, Bilei Zhu, Zejun Ma, Taylor Berg-Kirkpatrick, and Shlomo Dubnov. HTS - AT : A Hierarchical Token - Semantic Audio Transformer for Sound Classification and Detection , 2022. URL https://arxiv.org/abs/2202.00874. Version Number: 1

  21. [29]

    Deep one-class classification

    Lukas Ruff, Robert Vandermeulen, Nico Goernitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Alexander Binder, Emmanuel Muller, and Marius Kloft. Deep one-class classification. PMLR, pages 4393--4402, 2018

  22. [30]

    Semi- Supervised Sound Event Detection with Pre-Trained Model

    Liang Xu, Lizhong Wang, Sijun Bi, Hanyue Liu, and Jing Wang. Semi- Supervised Sound Event Detection with Pre-Trained Model . In ICASSP 2023 - 2023 IEEE International Conference on Acoustics , Speech and Signal Processing ( ICASSP ) , pages 1--5. IEEE, 2023. ISBN 978-1-7281-632...

  23. [31]

    Deep Residual Learning for Image Recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition . In 2016 IEEE Conference on Computer Vision and Pattern Recognition ( CVPR ) , pages 770--778, Las Vegas, NV, USA, June 2016. IEEE. ISBN 978-1-4673-8851-1. doi:10.1109/CVPR.201...

  24. [32]

    Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. MobileNets : Efficient Convolutional Neural Networks for Mobile Vision Applications . URL http://arxiv.org/abs/1704.04861

  25. [33]

    Zhengshan Shi and Gautham J. Mysore. LoopMaker : Automatic creation of music loops from pre-recorded music. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems , Chi '18, pages 1--6. Association for Computing Machinery, 2018. ISBN 978-1-4503-5620-6....

  26. [34]

    Music loop extraction from digital audio signals

    Bee Suan Ong and Sebastian Streich. Music loop extraction from digital audio signals. In 2008 IEEE International Conference on Multimedia and Expo , pages 681--684. IEEE, 2008. ISBN 978-1-4244-2570-9. doi:10.1109/ICME.2008.4607526. URL http://ieeexplore.ieee.org/document/4607526/

  27. [35]

    Matthew E. P. Davies, Philippe Hamel, Kazuyoshi Yoshii, and Masataka Goto. AutoMashUpper : Automatic Creation of Multi-Song Music Mashups . 22 0 (12): 0 1726--1737, 2014. ISSN 2329-9290, 2329-9304. doi:10.1109/TASLP.2014.2347135. URL http://ieeexplore.ieee.org/document/6876193/

  28. [36]

    Psychoacoustic Approaches for Harmonic Music Mixing

    Roman Gebhardt, Matthew Davies, and Bernhard Seeber. Psychoacoustic Approaches for Harmonic Music Mixing . 6 0 (5): 0 123, 2016. ISSN 2076-3417. doi:10.3390/app6050123. URL https://www.mdpi.com/2076-3417/6/5/123

  29. [37]

    A perceptually-motivated harmonic compatibility method for music mixing

    Gilberto Bernardes, Matthew Davies, and Carlos Guedes. A perceptually-motivated harmonic compatibility method for music mixing. 2017

  30. [38]

    Jordan B. L. Smith and Masataka Goto. Nonnegative Tensor Factorization for Source Separation of Loops in Audio . In 2018 IEEE International Conference on Acoustics , Speech and Signal Processing ( ICASSP ) , pages 171--175. IEEE, 2018. ISBN 978-1-5386-4658-8. doi:10.1109/ICASS...

  31. [39]

    FMA : A Dataset For Music Analysis , 2016

    Michaël Defferrard, Kirell Benzi, Pierre Vandergheynst, and Xavier Bresson. FMA : A Dataset For Music Analysis , 2016. URL https://arxiv.org/abs/1612.01840. Version Number: 3

  32. [40]

    Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko

    Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent:...

  33. [41]

    Attentional Feature Fusion , 2020

    Yimian Dai, Fabian Gieseke, Stefan Oehmcke, Yiquan Wu, and Kobus Barnard. Attentional Feature Fusion , 2020. URL https://arxiv.org/abs/2009.14082. Version Number: 2

  34. [42]

    Audio-text Retrieval with Transformer-based Hierarchical Alignment and Disentangled Cross-modal Representation , 2024

    Yifei Xin, Zhihong Zhu, Xuxin Cheng, Xusheng Yang, and Yuexian Zou. Audio-text Retrieval with Transformer-based Hierarchical Alignment and Disentangled Cross-modal Representation , 2024. URL http://arxiv.org/abs/2409.09256

  35. [43]

    Background-aware Modeling for Weakly Supervised Sound Event Detection

    Yifei Xin, Dongchao Yang, and Yuexian Zou. Background-aware Modeling for Weakly Supervised Sound Event Detection . In INTERSPEECH 2023 , pages 1199--1203. ISCA, 2023 a . doi:10.21437/Interspeech.2023-330. URL https://www.isca-archive.org/interspeech_2023/xin23_interspeech.html

  36. [44]

    Improving Weakly Supervised Sound Event Detection with Causal Intervention , 2023 b

    Yifei Xin, Dongchao Yang, Fan Cui, Yujun Wang, and Yuexian Zou. Improving Weakly Supervised Sound Event Detection with Causal Intervention , 2023 b . URL http://arxiv.org/abs/2303.05678

  37. [45]

    Tax and Robert P.W

    David M.J. Tax and Robert P.W. Duin. Support Vector Data Description . Machine Learning, 54 0 (1): 0 45--66, January 2004. ISSN 0885-6125. doi:10.1023/B:MACH.0000008084.60811.49. URL http://link.springer.com/10.1023/B:MACH.0000008084.60811.49

  38. [46]

    Free music samples: royalty-free loops, hits, and multis to download - sampleradar, n.d

    MusicRadar . Free music samples: royalty-free loops, hits, and multis to download - sampleradar, n.d. URL https://www.musicradar.com/news/tech/free-music-samples-royalty-free-loops-hits-and-multis-to-download-sampleradar. Accessed: 2024-10-04

  39. [47]

    Deep- Rhythm for Global Tempo Estimation in Music

    Hadrien Foroughmand and Geoffroy Peeters. Deep- Rhythm for Global Tempo Estimation in Music . November 2019. doi:10.5281/ZENODO.3527890. URL https://zenodo.org/record/3527890. Publisher: Zenodo

  40. [48]

    A Review of Activation Function for Artificial Neural Network

    Andrinandrasana David Rasamoelina, Fouzia Adjailia, and Peter Sincak. A Review of Activation Function for Artificial Neural Network . In 2020 IEEE 18th World Symposium on Applied Machine Intelligence and Informatics ( SAMI ) , pages 281--286, Herlany, Slovakia, January 2020. I...

  41. [49]

    U- Net : Convolutional Networks for Biomedical Image Segmentation , 2015

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- Net : Convolutional Networks for Biomedical Image Segmentation , 2015. URL http://arxiv.org/abs/1505.04597

  42. [50]

    Decoupled Weight Decay Regularization , 2019

    Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization , 2019. URL http://arxiv.org/abs/1711.05101

  43. [51]

    SGDR : Stochastic Gradient Descent with Warm Restarts , 2017

    Ilya Loshchilov and Frank Hutter. SGDR : Stochastic Gradient Descent with Warm Restarts , 2017. URL http://arxiv.org/abs/1608.03983

  44. [52]

    Isolation Forest

    Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation Forest . In 2008 Eighth IEEE International Conference on Data Mining , pages 413--422. IEEE. ISBN 978-0-7695-3502-9. doi:10.1109/ICDM.2008.17. URL http://ieeexplore.ieee.org/document/4781136/

  45. [53]

    Jolliffe

    Ian T. Jolliffe. Principal Component Analysis. Springer Series in Statistics. Springer, 2nd ed edition. ISBN 978-0-387-95442-4. doi:10.1007/b98835. URL http://link.springer.com/10.1007/b98835

  46. [54]

    A model for serendipitous music retrieval

    Markus Schedl, David Hauger, and Dominik Schnitzer. A model for serendipitous music retrieval. In Proceedings of the 2nd Workshop on Context-awareness in Retrieval and Recommendation , pages 10--13. ACM, 2012. ISBN 978-1-4503-1192-2. doi:10.1145/2162102.2162105. URL https://dl...

  47. [55]

    Serendipitous information retrieval

    Elaine Toms. Serendipitous information retrieval. 2001

  48. [56]

    Anomaly detection: A survey

    Varun Chandola, Arindam Banerjee, and Vipin Kumar. Anomaly detection: A survey. 41 0 (3): 0 1--58, 2009. ISSN 0360-0300, 1557-7341. doi:10.1145/1541880.1541882. URL https://dl.acm.org/doi/10.1145/1541880.1541882

  49. [57]

    A novel anomaly detection scheme based on principal component classifier

    Mei-Ling Shyu, Shu-Ching Chen, Kanoksri Sarinnapakorn, and Liwu Chang. A novel anomaly detection scheme based on principal component classifier. Proceedings of International Conference on Data Mining , 2003

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.