Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Comparison of spectrogram scaling in multi-label Music Genre Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Mel-scaled spectrograms yield statistically higher F1 scores than unscaled spectrograms for multi-label music genre recognition across ResNet depths 34 to 152 on a custom 18,000-song dataset, according to a paired t-test.

desk verdict Plausible descriptive result, but the reported p-value is internally inconsistent and the new multi-label dataset is the main asset. read the letter →

arxiv 2506.02091 v1 pith:6AOWMDKE submitted 2025-06-02 cs.SD cs.LGeess.AS

Bartosz Karpiński , Cyryl Leszczyński This is my paper
classification cs.SDcs.LGeess.AS
keywords multi-labelclassificationmusicgenrerecognitionMelspectrogramscalingtransferlearningResNetimbalanceddataaudio
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 asks whether converting audio spectrograms to the Mel scale—the frequency spacing that matches human hearing—improves multi-label music genre recognition. The authors compare plain spectrograms with Mel spectrograms as inputs to four pretrained ResNet classifiers (ResNet34, 50, 101, 152), each genre handled by a one-vs-all binary model. They report that Mel spectrograms give higher median F1 scores, with variance also lower, and a paired t-test on the differences is significant ($p = 0.0027$, $t = -3.5872$). The result matters because real-world music is eclectic and multi-label, while most benchmark datasets are single-tag and outdated; the paper also contributes a manually labeled 18,019-sample dataset spanning 16 genres from the 1950s to 2024.

What carries the argument

The load-bearing comparison device is the paired t-test applied to per-model F1 differences between the two spectrogram types, with a Shapiro–Wilk normality check and Q-Q plot used to justify the test. The object under test is the Mel scale, a perceptual frequency scale introduced in 1937 that reweights spectrogram energy toward the range humans actually hear, as implemented in the standard Python audio library used for preprocessing. The models are ResNet34, ResNet50, ResNet101, and ResNet152 fine-tuned from ImageNet weights and organized as an ensemble of one-vs-all binary classifiers so the output is multi-label. What carries the argument is the pairing: the same model is trained on both spectrogram types, so the F1 difference removes architecture-level variance and isolates the preprocessing effect.

What would settle it

Recompute the paired t-test with the four ResNet depths as the units: for the reported $t = -3.5872$ this gives a two-tailed $p$ near 0.037, not 0.0027. If the $n = 16$ per-genre differences were actually used, the analysis should account for clustering by model; a reader can rerun the comparison with repeated seeds and cross-validation to see whether the Mel advantage survives.

Watch

Extended reading notes

Core claim

The central claim is that on this highly varied, multi-label dataset, Mel-scaled spectrograms are statistically better than standard spectrograms for genre recognition as measured by F1. Using an ensemble of one-vs-all binary ResNet classifiers fine-tuned from ImageNet weights, the authors computed aggregated F1 scores for each spectrogram type across ResNet34, ResNet50, ResNet101, and ResNet152. The per-model differences passed a Shapiro–Wilk normality check, and a paired dependent-samples t-test gave $t = -3.5872$ with $p = 0.0027$, which the authors read as a statistically significant edge for the Mel scale. They also report that recall is fairly uniform across genres while precision tracks genre subset size, so the Mel advantage is carried mainly by precision.

Load-bearing premise

The claimed significance depends on the paired F1 differences being independent, correctly paired observations; if the test actually pooled one score per genre from the same four models, those observations are not independent and the p-value overstates the evidence.

Editorial extensions

If this is right

  • Practitioners choosing inputs for multi-label genre classifiers should prefer Mel-scaled spectrograms over unscaled ones when using ResNet transfer learning.
  • The Mel advantage appears across all tested depths (34, 50, 101, 152), so the recommendation is not tied to one architecture size.
  • Because recall is nearly flat across genres while precision tracks genre size, the remaining performance bottleneck is class imbalance, not spectrogram scaling.
  • Expanding the smallest genre subsets should improve average F1 more than further preprocessing changes, since F1 is driven by precision on well-represented genres.
  • The 18,019-song multi-label dataset, spanning 16 genres from the 1950s to 2024, offers a more current benchmark than single-label sets like GTZAN for future music genre recognition work.

Reading between the lines

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

  • A direct follow-up would be to run repeated-seed training and report both per-model and per-genre paired tests, which would show whether the $p = 0.0027$ result is robust to the unit of analysis.
  • The paper's observation that transformer and state-space audio models currently consume plain spectrograms suggests a testable extension: swap in Mel spectrograms for architectures like the Audio Spectrogram Transformer and measure whether the gain transfers beyond CNNs.
  • The strong precision-by-genre-size correlation implies a concrete data-collection rule: double the smallest genre subsets and re-measure F1; if the correlation is causal, average precision should rise more than from any input-scale change.
  • Comparing other perceptual scales (log, Q, F, or a learned scale) against Mel on the same dataset would situate the advantage: is it specific to Mel, or does any hearing-aligned scaling help?
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

3 major / 5 minor

Summary. The manuscript compares standard versus Mel-scaled spectrograms as input representations for multi-label music genre recognition. Using a privately collected, manually labeled dataset of 18,019 songs spanning 16 genres, the authors train one-vs-all ensembles of ResNet34, ResNet50, ResNet101, and ResNet152 with ImageNet transfer learning, and evaluate F1, precision, recall, and accuracy. The central claim is that Mel-scaled spectrograms yield higher F1 scores than standard spectrograms, supported by descriptive figures and a paired t-test reported as t = -3.5872, p = 0.0027. The paper also introduces the custom dataset and argues that it better reflects the multi-label, imbalanced nature of contemporary music than existing benchmarks.

Significance. If the central claim is sound, the paper provides a practically useful comparison for a common preprocessing choice in music information retrieval, and its manually labeled multi-label dataset is a potentially valuable resource. The manuscript also has concrete strengths: the full training code is publicly available, the experimental setup is described in enough detail to be largely reproducible, and the comparison spans four ResNet depths rather than a single architecture. However, the statistical evidence for the headline claim is not currently reliable, so the significance of the contribution depends on repairing the analysis in Section 4.

major comments (3)
  1. [Section 4] The paired t-test has a unit-of-analysis inconsistency that is load-bearing for the paper's main conclusion. The text states that "each model provides an F1 score for both spectrogram types" and that the difference was "calculated for each model," which implies n = 4 paired observations from ResNet34/50/101/152. However, for a paired t-test with t = -3.5872, the reported p = 0.0027 corresponds to approximately n = 16 (df ≈ 15), not n = 4 (df = 3, which would give p ≈ 0.037). The only apparent way to obtain n = 16 is to use per-genre F1 values across the 16 genres, but those observations are not independent: the same four models contribute to every genre, so a plain paired t-test on per-genre differences ignores model-level correlation and is anti-conservative. Because the abstract and conclusions rest on "statistically significant" superiority, this issue must be fixed. Please report exactly what the paired observations are, and either use a correctly specified analysis (e.g., a paired t-test on the four model-level differences, or a mixed-effects model with genre and model as crossed random effects) or explicitly treat the comparison as descriptive.
  2. [Sections 3.3 and 4] The evaluation protocol explicitly states that no cross-validation was performed and that all results come from a single training run. Given that the statistical claim is about generalization of the preprocessing advantage, a single run provides no estimate of training variability, and even a correctly specified t-test on n = 4 model-level differences would have very low power. The manuscript should either add repeated runs with different seeds or confidence intervals, or substantially soften the inferential language so that the conclusion is limited to the observed run.
  3. [Section 3.1 and 4] The second research question posed in Section 3.1 ("Does the difference (or lack thereof) stay the same between different ResNet implementations that vary in depth?") is not answered by any statistical analysis; the paper only provides descriptive figures. If the depth-dependence question is part of the contribution, it needs an explicit analysis (e.g., model-by-model comparisons or an interaction test). Otherwise, it should be removed or reframed as a descriptive observation.
minor comments (5)
  1. [Section 3.1, Table 1] The text says "the smallest one only 23" after listing Table 1, but the smallest genre in the table is Jazz with 278 entries; the sentence is confusing because the table does not include the genre with 23 entries.
  2. [Section 4] The sentence "the variance is also noticeably greater in the standard approach, although it results in having more having higher-scoring models" is ungrammatical and should be rewritten for clarity.
  3. [Throughout] The model names are inconsistent: "ResNet34" and "Resnet50," "ResNet101," "ResNet152" are mixed. Please standardize to a single spelling.
  4. [Section 3.1] The dataset itself is not released despite the code being available; since the dataset is a stated contribution, the reproducibility section should clearly state whether and how the dataset can be accessed.
  5. [References] Several references use inconsistent formatting (e.g., reference [3] has a stray comma before the venue, and some entries mix capitalization styles); a careful proofread of the bibliography is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports an empirical spectrogram comparison with no fitted parameters, self-citation chain, or definitional reductions.

full rationale

The paper's central claim is an empirical comparison of mel-scaled versus standard spectrograms for multi-label music genre recognition. The outcome (F1 score) is measured from trained models, not derived from the input representation by construction, and no parameter is fitted to a subset of data and then relabeled as a prediction. The statistical test in Section 4 is an evaluation of measured differences, not a derivation, and the authors do not invoke any uniqueness theorem or prior work by themselves to force the conclusion. The dataset is self-collected, but training and evaluating on one's own dataset is standard empirical practice and does not make the comparison circular. The skeptically noted issue—that the reported p=0.0027 may imply a different unit of analysis than the stated four models—is a statistical validity concern, not a circularity concern, and therefore does not affect the circularity score. No load-bearing step reduces to its own inputs, so the finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The central comparison rests on standard statistical assumptions and on the transferability of ImageNet representations to spectrogram inputs. No free parameters or invented entities are introduced; the dataset labels are scraped from online databases without formal verification.

assumptions (2)
  • domain assumption ImageNetV1 pretrained weights provide a valid initialization for spectrogram image classification.
    Transfer learning from ImageNet is used for all ResNet models (Section 3.1); if this premise fails, the comparisons could be biased, though both spectrogram types are affected equally.
  • standard math The paired differences in F1 scores are approximately normally distributed and independently sampled.
    The t-test (Section 4) requires these assumptions; Shapiro-Wilk 'failed to reject normality', but the sample is small and the independence of observations is not established.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Comparison of spectrogram scaling in multi-label Music Genre Recognition." pith.science (2026). https://pith.science/paper/6AOWMDKE

@misc{pith2026250602091,
  author       = {Pith},
  title        = {Pith review of: Comparison of spectrogram scaling in multi-label Music Genre Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6AOWMDKE}},
  note         = {Machine review of arXiv:2506.02091}
}
read the original abstract

As the accessibility and ease-of-use of digital audio workstations increases, so does the quantity of music available to the average listener; additionally, differences between genres are not always well defined and can be abstract, with widely varying combinations of genres across individual records. In this article, multiple preprocessing methods and approaches to model training are described and compared, accounting for the eclectic nature of today's albums. A custom, manually labeled dataset of more than 18000 entries has been used to perform the experiments.

Figures

Figures reproduced from arXiv: 2506.02091 by the authors.

Figure 1
Figure 1. A sample default spec￾trogram [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. Average accuracy across models Resnet34 Resnet50 Resnet101 Resnet152 Model 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 F1 Variant Mel Normal [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Average F1 Score across models 7 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Average Recall across models Resnet34 Resnet50 Resnet101 Resnet152 Model 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Precision Variant Mel Normal [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Average precision across models 8 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Average F1 Score across genres for Resnet50 [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Average Recall across genres for Resnet50 [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Average Precision across genres for Resnet50 [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Q–Q Plot for Differences in F1 between mel spectrograms and normal spectrograms [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 33 canonical work pages

  1. [1]

    2017, vol

    Ali, M.A., Siddiqui, Z.A., Automatic music genres classification using machine learning , In- ternational Journal of Advanced Computer Science and Applications. 2017, vol. 8, 8

    work page 2017

  2. [2]

    Bahuleyan, H., Music genre classification using machine learning techniques , arXiv preprint arXiv:1804.01149. 2018

    arXiv 2018

  3. [3]

    , in: Ismir, 9 (2011), page 10

    Bertin-Mahieux, T., Ellis, D.P., Whitman, B., Lamere, P., The million song dataset. , in: Ismir, 9 (2011), page 10

    work page 2011

  4. [4]

    2392–2396

    Choi, K., Fazekas, G., Sandler, M., Cho, K., Convolutional recurrent neural networks for music classification , in: 2017 IEEE International conference on acoustics, speech and signal processing (ICASSP) (IEEE, 2017), pp. 2392–2396

    work page 2017

  5. [5]

    2017, vol

    Costa, Y.M., Oliveira, L.S., Silla Jr, C.N., An evaluation of convolutional neural networks for music classification using spectrograms, Applied soft computing. 2017, vol. 52, pp. 28–38

    work page 2017

  6. [6]

    Dhall, A., Srinivasa Murthy, Y., Koolagudi, S.G., Music genre classification with convolutional neural networks and comparison with f, q, and mel spectrogram-based images , in: Advances in Speech and Music Technology: Proceedings of FRSM 2020 (Springer, 2021), pp. 235–248

    work page 2020

  7. [7]

    Dieleman, S., Brakel, P., Schrauwen, B., Audio-based music classification with a pretrained con- volutional network, in: 12th International Society for Music Information Retrieval Conference (ISMIR-2011) (University of Miami, 2011), pp. 669–674

    work page 2011

  8. [8]

    Dong, M., Convolutional neural network achieves human-level accuracy in music genre classi- fication, arXiv preprint arXiv:1802.09697. 2018

    work page Pith review arXiv 2018

Show all 39 references
  1. [9]

    Erol, M.H., Senocak, A., Feng, J., Chung, J.S., Audio mamba: Bidirectional state space model for audio representation learning , IEEE Signal Processing Letters. 2024

    2024

  2. [10]

    French, M., Handy, R., Spectrograms: turning signals into pictures , Journal of Engineering Technology, vol. 24, pp. 32-35. 2007

    2007

  3. [11]

    Gong, Y., Chung, Y.A., Glass, J., Ast: Audio spectrogram transformer , arXiv preprint arXiv:2104.01778. 2021

    2021 arXiv

  4. [12]

    36 (2022), pp

    Gong, Y., Lai, C.I., Chung, Y.A., Glass, J., Ssast: Self-supervised audio spectrogram trans- former, in: Proceedings of the AAAI Conference on Artificial Intelligence , vol. 36 (2022), pp. 10699–10709

    2022

  5. [13]

    KM, A., i in., Deep learning based music genre classification using spectrogram , in: Proceed- ings of the International Conference on IoT Based Control Networks & Intelligent Systems- ICICNIS (2021)

    2021

  6. [14]

    Lin, J., Hu, H., Audio mamba: Pretrained audio state space model for audio tagging , arXiv preprint arXiv:2405.13636. 2024

    2024 arXiv

  7. [15]

    , in: ISMIR (2009), pp

    Lukashevich, H.M., Abeßer, J., Dittmar, C., Grossmann, H., From multi-labeling to multi- domain-labeling: A novel two-dimensional approach to music genre classification. , in: ISMIR (2009), pp. 459–464. 12

    2009

  8. [16]

    Matocha, M., Zieli´ nski, S., Music genre recognition using convolutional neural networks , Ad- vances in Computer Science Research. 2018

    2018

  9. [17]

    McFee, B., Raffel, C., Liang, D., Ellis, D.P., McVicar, M., Battenberg, E., Nieto, O., librosa: Audio and music signal analysis in python. , SciPy. 2015, vol. 2015, pp. 18–24

    2015

  10. [18]

    St´ efan van der Walt, Jarrod Millman (2010), pp

    Wes McKinney, Data Structures for Statistical Computing in Python , in: Proceedings of the 9th Python in Science Conference , pod red. St´ efan van der Walt, Jarrod Millman (2010), pp. 56 – 61

    2010

  11. [19]

    1101–1107

    Mehta, J., Gandhi, D., Thakur, G., Kanani, P., Music genre classification using transfer learning on log-based mel spectrogram , in: 2021 5th International Conference on Computing Methodologies and Communication (ICCMC) (IEEE, 2021), pp. 1101–1107

    2021

  12. [20]

    Nakamura, H., Huang, H.H., Kawagoe, K., Detecting musical genre borders for multi-label genre classification , in: 2013 IEEE International Symposium on Multimedia (IEEE, 2013), pp. 532–533

    2013

  13. [21]

    Ndou, N., Ajoodha, R., Jadhav, A., Music genre classification: A review of deep-learning and traditional machine-learning approaches , in: 2021 IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS) (IEEE, 2021), pp. 1–6

    2021

  14. [22]

    Nirmal, M., Mohan, S., Music genre classification using spectrograms , in: 2020 International conference on power, instrumentation, control and computing (PICC) (IEEE, 2020), pp. 1–5

    2020

  15. [23]

    2017, vol

    Oramas, S., Nieto, O., Barbieri, F., Serra, X., Multi-label music genre classification from audio, Text, and Images Using Deep Features. 2017, vol. 21

    2017

  16. [24]

    Paszke, A., Pytorch: An imperative style, high-performance deep learning library , arXiv preprint arXiv:1912.01703. 2019

    1912 arXiv

  17. [25]

    2011, vol

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., i in., Scikit-learn: Machine learning in python , the Journal of machine Learning research. 2011, vol. 12, pp. 2825–2830

    2011

  18. [26]

    2020, vol

    Pelchat, N., Gelowitz, C.M., Neural network music genre classification , Canadian Journal of Electrical and Computer Engineering. 2020, vol. 43, 3, pp. 170–173

    2020

  19. [27]

    Sanden, C., Zhang, J.Z., Enhancing multi-label music genre classification through ensemble techniques, in: Proceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval (2011), pp. 705–714

    2011

  20. [28]

    Schindler, A., Knees, P., Multi-task music representation learning from multi-label embeddings, in: 2019 International Conference on Content-Based Multimedia Indexing (CBMI) (IEEE, 2019), pp. 1–6

    2019

  21. [29]

    1053–1059

    Shams, S., Dindar, S.S., Jiang, X., Mesgarani, N., Ssamba: Self-supervised audio representation learning with mamba state space model, in: 2024 IEEE Spoken Language Technology Workshop (SLT) (IEEE, 2024), pp. 1053–1059

    2024

  22. [30]

    2008, vol

    Silla, C.N., Koerich, A.L., Kaestner, C.A., A machine learning approach to automatic music genre classification, Journal of the Brazilian Computer Society. 2008, vol. 14, pp. 7–18. 13

    2008

  23. [31]

    1937, vol

    Stevens, S.S., Volkmann, J., Newman, E.B., A scale for the measurement of the psychological magnitude pitch, The journal of the acoustical society of america. 1937, vol. 8, 3, pp. 185–190

    1937

  24. [32]

    Sturm, B.L., The gtzan dataset: Its contents, its faults, their effects on evaluation, and its future use, arXiv preprint arXiv:1306.1461. 2013

    2013 arXiv

  25. [33]

    2002, vol

    Tzanetakis, G., Cook, P., Musical genre classification of audio signals , IEEE Transactions on speech and audio processing. 2002, vol. 10, 5, pp. 293–302

    2002

  26. [34]

    2021, vol

    Waskom, M.L., seaborn: statistical data visualization , Journal of Open Source Software. 2021, vol. 6, 60, page 3021

    2021

  27. [35]

    Yadav, S., Tan, Z.H., Audio mamba: Selective state spaces for self-supervised audio represen- tations, arXiv preprint arXiv:2406.02178. 2024

    2024 arXiv

  28. [36]

    3382–3386

    Yang, H., Zhang, W.Q., Music genre classification using duplicated convolutional layers in neural networks., in: Interspeech (2019), pp. 3382–3386

    2019

  29. [37]

    Zhao, H., Zhang, C., Zhu, B., Ma, Z., Zhang, K., S3t: Self-supervised pre-training with swin transformer for music classification , in: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (IEEE, 2022), pp. 606–610

    2022

  30. [38]

    Zhu, W., Omar, M., Multiscale audio spectrogram transformer for efficient audio classifica- tion, in: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (IEEE, 2023), pp. 1–5

    2023

  31. [39]

    Zink Media Inc., Discogs search page, https://www.discogs.com/search/. 14

Pith tools

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