Music Genre Classification Using Machine Learning Techniques

Alokit Mishra; Ryyan Akhtar

REVIEW 4 major objections 4 minor 30 references

Music Genre Classification Using Machine Learning Techniques

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper argues that on a small, clean audio benchmark like GTZAN, an SVM with hand-crafted features can outperform an end-to-end CNN because feature engineering regularizes the model, although its own results table reports CNN 85% vs SVM

desk verdict The paper's central claim—SVM beats CNN on GTZAN—is contradicted by its own Table 2, and the methodological gaps are too big to wave away. read the letter →

arxiv 2509.01762 v1 pith:JEOE4YHZ submitted 2025-09-01 cs.SD cs.LG

Alokit Mishra , Ryyan Akhtar This is my paper · ORCID

classification cs.SDcs.LG

keywords musicgenreclassificationGTZANsupportvectormachineconvolutionalneuralnetworkmelspectrogramhand-craftedaudiofeaturesfeatureengineeringnoiserobustness

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

The paper argues that on a small, clean dataset like GTZAN, a Support Vector Machine trained on hand-crafted audio features (MFCCs, chroma, tempo, zero-crossing rate, spectral descriptors) can outperform an end-to-end CNN, because feature engineering imposes a strong inductive bias that regularizes the model when data are scarce. The intended implication is that deep learning should not be applied reflexively: classical classifiers with expert features remain competitive, even superior, in data-constrained settings. A sympathetic reader would care because this runs against the prevailing trend of defaulting to deep architectures and suggests practical, low-compute alternatives for audio tagging. But the paper's own comparative table states the reverse: CNN reaches 85% accuracy and SVM 81% on clean data, and the CNN degrades less under noise. The claimed SVM superiority thus stands on a specific experimental configuration — a small CNN trained for a limited number of epochs — rather than on the results actually reported.

What carries the argument

The load-bearing mechanism is the bias-variance tradeoff in the small-data regime. The hand-crafted feature set (MFCCs, chroma, spectral centroid and bandwidth, tempo, zero-crossing rate, central moments) is condensed expert knowledge normalized to [0,1]; this low-dimensional representation imposes a strong inductive bias on the RBF-kernel SVM, lowering variance. The CNN, by contrast, must learn these features from mel spectrograms; with a two-convolutional-layer architecture trained for 'a limited number of epochs,' it has high capacity and no equivalent regularization, so it overfits the 100 samples per class. The regularization-by-feature-engineering story is the paper's explanatory engin

What would settle it

Run the same GTZAN comparison with a matched, explicitly reported train/test split, training the CNN to convergence (deeper architecture, tuning, augmentation); the paper's own Table 2 already reports CNN 85% vs SVM 81% on clean data, so if the SVM still cannot exceed a properly trained CNN in a controlled replication, the central claim is refuted.

Watch

Extended reading notes

Core claim

The paper's intended discovery is that the bias-variance tradeoff can invert the usual deep-learning advantage on a small benchmark: a Support Vector Machine, fed normalized, expert-engineered features (MFCCs, chroma, spectral centroid/bandwidth, tempo, zero-crossing rate, central moments), finds a well-generalizing low-variance decision boundary, while a high-capacity CNN overfits. The authors attribute the effect to feature engineering acting as a regularizer, and to GTZAN being clean, small, and class-balanced. They additionally report that adding synthetic noise reverses the ranking in favor of the CNN, making noise resilience the deep model's comparative strength. The paper's own result

Load-bearing premise

The paper assumes that its small, briefly trained two-layer CNN fairly represents what deep learning can do on this task; if a properly tuned or deeper CNN scores higher, the claimed SVM advantage collapses.

Editorial extensions

If this is right

On small, clean, well-normalized audio datasets, classical classifiers with hand-crafted features become a serious alternative to deep networks, especially when compute or labeled data are limited.
Feature engineering should be understood as a regularization strategy: when data are scarce, models with strong inductive bias and low variance can generalize better than high-capacity end-to-end networks.
A pragmatic takeaway is to build hybrid models that feed both raw spectrograms and pre-computed features into a network, a dual-stream design the paper proposes for future work.
Noise resilience remains a differentiator for CNNs, so in real-world noisy deployments the deep model may still be the safer choice even on small datasets.

Reading between the lines

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

The paper's reported numbers (CNN 85% vs SVM 81% clean; CNN 79% vs SVM 66% noisy) do not support the abstract's claim that SVM has 'superior classification accuracy'; the CNN wins in both settings, so the headline conclusion appears to be a misstatement of the paper's own evidence.
A meaningful test the paper leaves unexplored: the 3-second segmented CSV, which multiplies training instances tenfold and could improve the SVM's accuracy, is never used in the reported results; running it might genuinely flip the ranking.
Because the paper does not report the train/test split, the known GTZAN artist-leakage problem (same artist in train and test) is unaddressed; a proper split could change both models' scores and is the minimum control needed before the comparison can be trusted.
The conclusion that 'CNNs could not leverage their full potential' is a direct consequence of the CNN baseline being small and briefly trained; a stronger deep baseline (deeper, tuned, with augmentation) is the obvious next experiment and would likely widen the gap in the CNN's favor.

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.

Desk Editor's Note

read the letter

You should know the stress-test note is right, and it lands hard. The abstract, Section 6, and Section 7 all say the SVM achieves superior accuracy over the CNN, but Table 2 reports CNN at 85% clean and 79% noisy, versus SVM at 81% and 66%. Section 4.6 says the CNN 'achieved an accuracy of 85' and Section 5.4 says CNN 'consistently outperformed across all settings.' No reported condition supports the headline. Section 7.1's 'under certain controlled conditions' is never specified. This is not a wording issue; the paper's main conclusion contradicts its own quantitative results. The reader's verdict of REJECT is the right one.

What the paper does reasonably well: the feature definitions (MFCCs, chroma, spectral centroid) are competently explained, the pre-emphasis and librosa pipeline is clear, and the PCA visualization plus confusion-matrix discussion give a decent sense of which genres are confusable. It also states in places that CNNs are more robust, which is at least consistent with the table.

The soft spots beyond the contradiction are substantial. There is no reported train/test split or cross-validation protocol, so GTZAN's known artist leakage is never addressed. No hyperparameters for the SVM or CNN are given, and the CNN is admitted to be trained for 'a limited number of epochs'—so even the non-contradicted numbers cannot be interpreted as a fair comparison. The 'novel' noise robustness experiment lacks protocol details, and its motivation is undercut because the abstract doesn't mention noise at all. There are also internal numerical slips: Section 4.5 says seven binary classifiers for a ten-class problem, and Section 7.1 says '1000 samples per genre' when it is 100 per genre. Unresolved placeholders ('Figure Y', 'Figure W') add to the impression that the manuscript was not finished. No code or data are provided.

This paper is not a serious research contribution in its current form. It reads like an undergraduate project that assembles standard tools and standard features on a standard benchmark, then draws the opposite conclusion from its own table. It could be a starting point for a proper study, but as submitted it should not consume referee time. Desk reject is appropriate. I would not cite it, and I would not bring it to a reading group except as a warning about internal consistency.

Referee Report

4 major / 4 minor

Summary. The paper compares classical machine-learning classifiers (logistic regression, random forest, gradient boosting, SVM) trained on hand-crafted audio features with a CNN trained on mel-spectrograms, using the GTZAN dataset. The abstract and conclusions claim that the SVM achieves superior classification accuracy compared to the CNN, attributing this to the regularizing effect of engineered features on a small dataset. Results are reported for clean and noisy conditions, with accuracy and F1 scores. The manuscript also includes PCA-based visualization, confusion-matrix discussion, and a high-level analysis of why classical models might be competitive.

Significance. If the central claim were supported, the paper would offer a useful counterpoint to the prevailing deep-learning advantage in music genre classification, especially for moderately sized datasets. Credit is due for experimenting with several classical baselines, using both track-level and segment-level features, and for including a noise-robustness comparison. However, the paper's stated contribution is directly contradicted by its own quantitative results: Table 2 reports CNN 85% clean / 79% noisy / F1 0.83 versus SVM 81% / 66% / F1 0.78. Moreover, no train/test split, cross-validation, error bars, or hyperparameter search is reported, so the numerical comparisons cannot be interpreted as statistically meaningful. The paper ships no code or reproducible evaluation protocol, so the claimed inversion of the usual result is not supportable.

major comments (4)

[Abstract, §5.4 (Table 2), §6, §7] The central claim that 'the SVM ... achieves superior classification accuracy compared to the end-to-end CNN model' is contradicted by every reported condition. Table 2 shows CNN 85% clean / 79% noisy with F1 0.83, versus SVM 81% / 66% with F1 0.78. §4.6 also states the CNN 'achieved an accuracy of 85' and that CNNs are 'better suited' for the task; §5.3 and §5.4 state CNN 'consistently outperformed across all settings.' Section 7.1 says SVM outperformed CNN 'under certain controlled conditions,' but those conditions are never specified and no reported experiment supports that statement. This is a load-bearing internal contradiction, not a wording issue, because the abstract and conclusions rest on the claimed SVM superiority.
[§5 (Evaluation)] The evaluation protocol is severely under-specified. There is no stated train/test split ratio, no cross-validation, no repeated runs, and no error bars. The GTZAN dataset is known to contain artist-leakage risks, and this manuscript does not mention how tracks were partitioned to avoid correlated examples across train and test. Without this information, the reported 81% vs 85% difference is not interpretable. Additionally, no hyperparameter ranges or selection procedures are given for the SVM (C, gamma) or the CNN (epochs, batch size, learning rate, dropout rate), so the comparison is not a controlled experiment.
[§4.6, §7.1] The CNN baseline is too weak to support the paper's intended comparison. §4.6 says the CNN had only two convolutional layers and was 'trained for a limited number of epochs,' with no architecture details, tuning, data augmentation, or training curves. Section 7.1 partly explains the result by 'Limited Feature Extraction Scope,' stating that 'CNNs could not leverage their full potential.' This means the comparison is between a well-engineered classical pipeline and an untuned shallow CNN, not a fair test of 'end-to-end CNN model' as claimed in the abstract. If the CNN had been properly tuned or deepened, the headline claim would likely collapse further.
[§3, §4.5, §7.1] There are factual errors in dataset and model descriptions. Section 3 correctly states that each genre contains 100 audio files, but §7.1 says 'GTZAN’s modest dataset size (1000 samples per genre)'—the dataset has 1000 total samples, not 1000 per genre. Section 4.5 says logistic regression 'involves training seven separate binary classifiers' for a 10-genre problem. These are not merely cosmetic; the dataset-size argument in §7.1 is built on the incorrect '1000 samples per genre' premise.

minor comments (4)

[§5.2, §5.3] The manuscript refers to 'Figure Y' and 'Figure W' as placeholders. These figures must be numbered and referenced properly.
[§3 (Table 1)] Table 1 appears to be empty in the provided text. It should contain the genre distribution, or the reference should be removed.
[§4.6] 'the CNN achieved an accuracy of 85' is missing a percent sign and a period. Also, the manuscript uses inconsistent capitalization for 'Librosa' vs 'librosa.'
[General] The title duplicates that of an earlier paper by Bahuleyan (2018), which is cited as [2]. The authors should at least discuss the relationship and distinguish their contribution, or choose a more specific title.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified; the study is an empirical comparison, and the abstract's SVM-superiority claim conflicts with Table 2 but is not derived from its own inputs.

full rationale

The paper's argument chain is empirical: extract hand-crafted features, train classical classifiers, train a CNN on mel-spectrograms, and compare accuracy/F1 on clean and noisy GTZAN data. No equation in Sections 4.1-4.5 defines a feature or model in terms of the target comparison. No parameter is fit to a subset and then reported as a prediction of a closely related quantity; the confusion matrix and robustness tests are separate evaluations. The reference list contains no self-citations by Mishra and Akhtar; the similarly-titled Bahuleyan reference is a different author. The main issue is an internal contradiction, not circularity: the abstract and Sections 6-7 assert SVM outperforms CNN, while Table 2 and Sections 4.6 and 5.3 report CNN (85%, 79% noisy, F1 0.83) outperforming SVM (81%, 66%, F1 0.78) in every condition. Section 7.1's 'certain controlled conditions' are never specified, making the SVM-superiority claim post hoc and unsupported, but explaining a result after the fact is not the same as reducing the result to an input assumption. No derivation chain here is equivalent to its own premise, so no circularity is found.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper does not introduce new particles, forces, or theoretical constructs. Its load-bearing assumptions are all experimental protocol assumptions: a valid GTZAN split, a fair CNN baseline, no leakage, and a well-defined noise test. The free parameters are mostly unstated hyperparameters and fixed feature extraction settings that the central comparison depends on.

free parameters (5)

SVM regularization C and RBF gamma = not reported
SVM performance depends on these hyperparameters, but no values or tuning procedure are given. The reported 81% could depend on this unspecified choice.
CNN architecture and training hyperparameters = not reported (two convolutional layers, 'limited number of epochs')
Number of filters, dropout rate, learning rate, batch size, and epochs are not given, making the CNN an uncontrolled baseline.
Pre-emphasis coefficient alpha = 0.97
Chosen by common practice and not justified; it affects all downstream features.
MFCC settings (n_mels, nfft, hop size) = 20, 2048, 512
These feature extraction parameters are fixed by choice and not varied or justified.
Feature normalization range = 0 to 1
Section 7.1 credits normalization for SVM's success; the scaling is itself a modeling choice that affects kernel distances.

assumptions (5)

domain assumption GTZAN is used as a single benchmark with no artist leakage and a standard fixed split
No train/test split or cross-validation protocol is described; the paper treats reported accuracies as directly comparable.
domain assumption The CNN baseline fairly represents end-to-end deep learning
Section 4.6 admits limited epochs and no tuning, yet the comparison is used to draw conclusions about deep learning generally.
domain assumption No information leakage occurs between feature extraction and training
The pipeline description does not state whether normalization or aggregation statistics are computed per split or on the full dataset.
standard math STFT, MFCC, PCA, and SVM kernel mathematics are correct standard background
The paper invokes standard signal processing and kernel methods without proving them; this is acceptable background but is relied on throughout.
domain assumption The synthetic noise experiment is well defined and equivalent across models
Section 5.3 says Gaussian and pink noise were added at varying SNR, but no SNR values, noise generation code, or protocol details are given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Music Genre Classification Using Machine Learning Techniques." pith.science (2026). https://pith.science/paper/JEOE4YHZ

@misc{pith2026250901762,
  author       = {Pith},
  title        = {Pith review of: Music Genre Classification Using Machine Learning Techniques},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEOE4YHZ}},
  note         = {Machine review of arXiv:2509.01762}
}

read the original abstract

This paper presents a comparative analysis of machine learning methodologies for automatic music genre classification. We evaluate the performance of classical classifiers, including Support Vector Machines (SVM) and ensemble methods, trained on a comprehensive set of hand-crafted audio features, against a Convolutional Neural Network (CNN) operating on Mel spectrograms. The study is conducted on the widely-used GTZAN dataset. Our findings demonstrate a noteworthy result: the SVM, leveraging domain-specific feature engineering, achieves superior classification accuracy compared to the end-to-end CNN model. We attribute this outcome to the data-constrained nature of the benchmark dataset, where the strong inductive bias of engineered features provides a regularization effect that mitigates the risk of overfitting inherent in high-capacity deep learning models. This work underscores the enduring relevance of traditional feature extraction in practical audio processing tasks and provides a critical perspective on the universal applicability of deep learning, especially for moderately sized datasets.

Figures

Figures reproduced from arXiv: 2509.01762 by the authors.

**Figure 6.** Waveform of Jazz [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗

**Figure 7.** Waveform of Hip-Hop [PITH_FULL_IMAGE:figures/full_fig_p004_7.png] view at source ↗

Figures from the paper (12 more)

**Figure 8.** Figure 8: Waveform of Disco 4 [PITH_FULL_IMAGE:figures/full_fig_p004_8.png]

**Figure 9.** Figure 9: Waveform of Country [PITH_FULL_IMAGE:figures/full_fig_p005_9.png]

**Figure 10.** Figure 10: Waveform of Blues 4.2 Convolutional Neural Networks Convolutional Neural Networks (CNNs) are a sophisticated class of deep learning models meticulously engineered for processing structured grid data, with a primary focus on images while also extending their capabili…

**Figure 11.** Figure 11: Magnitude of Short-Time Fourier Transform (STFT) 4.3.1 Time Domain Features These features were derived directly from the raw audio signal. • Central Moments: This set of features includes statistical metrics like the mean, standard deviation, skewness, and kurtos…

**Figure 12.** Figure 12: Average Standard Deviation • Tempo: Tempo is a critical attribute in music, representing the speed or pace at which a piece of music is played, typically measured in Beats Per Minute (BPM). Different music genres and styles often have distinctive tempo ranges, with so…

**Figure 14.** Figure 14: Spectral Bandwidth 4.4 Exploratory Data Visualization with PCA To better understand the separability of different genres in feature space, we applied Principal Component Analysis (PCA) to the extracted MFCC features. PCA reduces the high-dimensional data (57 features…

**Figure 13.** Figure 13: Magnitude-Weighted Frequency • Spectral Bandwidth: Spectral bandwidth is a crucial feature that describes the width of the frequency range in which most of the signal’s energy is concentrated. It is calculated using the p-th order moment around the spectral centroid,…

**Figure 15.** Figure 15: PCA visualization of music genres in fea [PITH_FULL_IMAGE:figures/full_fig_p009_15.png]

**Figure 16.** Figure 16: Example mel-spectrogram used for CNN input. 5 Evaluation The models’ performance was assessed using accuracy and F1-score. 5.1 Metrics • Accuracy: The proportion of correctly classified samples. While intuitive, it can be misleading for imbalanced datasets. 10 [PI…

**Figure 17.** Figure 17: Example mel-spectrogram used for CNN input [PITH_FULL_IMAGE:figures/full_fig_p011_17.png]

**Figure 18.** Figure 18: Amplitude (Loudness of the corresponding [PITH_FULL_IMAGE:figures/full_fig_p011_18.png]

**Figure 20.** Figure 20: Accuracy degradation of SVM and CNN under varying noise levels. 5.4 Comparative Results Table [PITH_FULL_IMAGE:figures/full_fig_p012_20.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 29 canonical work pages

[1]

and Yu, D., 2014

Abdel-Hamid, O., Mohamed, A., Jiang, H., Deng, L., Penn, G. and Yu, D., 2014. Convo- lutional neural networks for speech recognition. IEEE/ACM Transactions on audio, speech, and language processing, 22(10), pp.1533-1545

work page 2014
[2]

Music Genre Classification using Machine Learning Techniques

Bahuleyan, H., 2018. Music genre classifica- tion using machine learning techniques. arXiv preprint arXiv:1804.01149

work page Pith review arXiv 2018
[3]

and Kaestner, C.A., 2008, October

Silla Jr, C.N., Koerich, A.L. and Kaestner, C.A., 2008, October. A machine learning approach to automatic music genre classification. In Journal of the Brazilian Computer Society

work page 2008
[4]

and Kumar, V.V., 2019, March

Chillara, S., Kavitha, K.R., Narayana, K.S., Varma, G.P.S. and Kumar, V.V., 2019, March. Music genre classification using machine learn- ing algorithms. In 2019 3rd International Con- ference on Trends in Electronics and Informatics (ICOEI) (pp. 515-519). IEEE

work page 2019
[5]

and Aydin, N., 2020, September

Elbir, A. and Aydin, N., 2020, September. Music genre classification and recommendation by us- ing artificial intelligence. In 2020 International Conference on Data Analytics for Business and Industry: Way Towards a Sustainable Economy (ICDABI) (pp. 1-6). IEEE

work page 2020
[6]

and Sharma, G., 2020

Ghildiyal, S. and Sharma, G., 2020. Music genre classification: A review. Vision, 4(3), p.42

work page 2020
[7]

and Dogo, E.M., 2021

Ndou, M.S. and Dogo, E.M., 2021. Music Genre Classification Using Machine Learning

work page 2021
[8]

and Mermelstein, P., 1990

Davis, S. and Mermelstein, P., 1990. Compari- son of parametric representations for monosyl- labic word recognition in continuously spoken sentences. In Readings in speech recognition(pp. 65-74). Morgan Kaufmann. 14

work page 1990

Show all 30 references

[9]

Greedy function approxi- mation: a gradient boosting machine

Friedman, J.H., 2001. Greedy function approxi- mation: a gradient boosting machine. Annals of statistics, pp.1189-1232

2001
[10]

and Ritter, M., 2017, March

Gemmeke, J.F., Ellis, D.P., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M. and Ritter, M., 2017, March. Audio set: An ontology and human-labeled dataset for au- dio events. In 2017 IEEE international confer- ence on acoustics, speech and signal processing (ICA...

2017
[11]

and Delerue, O., 2000

Gouyon, F., Pachet, F. and Delerue, O., 2000. On the use of zero-crossing rate for an appli- cation of classification of percussive sounds. In Proceedings of the COST G6 Conference on Dig- ital Audio Effects (DAFX-00) , Verona, Italy

2000
[12]

and Kurth, F., 2010, January

Grosche, P., M¨ uller, M. and Kurth, F., 2010, January. Cyclic tempogram-a mid-level repre- sentation for music signals. In 2010 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (pp. 5522-5525). IEEE

2010
[13]

and Friedman, J.,

Hastie, T., Tibshirani, R. and Friedman, J.,
[14]

and Woraratpa- nya, T., 2012, March

Ittichaichareon, C., Suksri, S. and Woraratpa- nya, T., 2012, March. Speech recognition us- ing MFCC. In 2012 international conference on computer, information and telecommunication systems (CITS) (pp. 1-4). IEEE

2012
[15]

and Cai, L.H., 2002, July

Jiang, D.N., Lu, L., Zhang, H.J., Tao, J.H. and Cai, L.H., 2002, July. Music type classifi- cation by spectral contrast feature. In Proceed- ings of the 2002 IEEE International Conference on Multimedia and Expo (Vol. 1, pp. 113-116). IEEE

2002
[16]

and Schoner, B., 2012

Kim, Y.E., Whitman, B. and Schoner, B., 2012. Power-normalized cepstral coefficients (PNCC) for robust speech recognition. In Acoustics, Speech and Signal Processing (ICASSP), 2012 IEEE International Conference on (pp. 4101- 4104). IEEE

2012
[17]

and Hinton, G.E.,

Krizhevsky, A., Sutsever, I. and Hinton, G.E.,
[18]

and Chun, A., 2010, Oc- tober

Li, T., Chan, A.B. and Chun, A., 2010, Oc- tober. Automatic musical pattern feature ex- traction using convolutional neural networks. In 11th Int. Society for Music Information Re- trieval Conf.(ISMIR 2010) (pp. 61-66)

2010
[19]

and Rauber, A., 2005, September

Lidy, T. and Rauber, A., 2005, September. Eval- uation of feature extractors and psycho-acoustic transformations for music genre classification. In Ismir (pp. 34-41)

2005
[20]

and Schindler, A., 2016, August

Lidy, T. and Schindler, A., 2016, August. Paral- lel convolutional neural networks for music genre and mood classification. In 2016 international conference on signal processing and communica- tions (spcom) (pp. 1-5). IEEE

2016
[21]

and Ellis, D.P., 2005

Mandel, M.I. and Ellis, D.P., 2005. Song-level features and support vector machines for mu- sic classification. In 6th international conference on music information retrieval (ismir 2005) (pp. 594-599)

2005
[22]

and Lumini, A., 2016

Nanni, L., Brahnam, S. and Lumini, A., 2016. Combining visual and acoustic features for mu- sic genre classification. Expert Systems with Ap- plications, 45, pp.108-117

2016
[23]

Feature selection for music genre classification

Ng, K.O., 2004. Feature selection for music genre classification. Stanford University

2004
[24]

and Le, Q.V., 2019

Park, D.S., Chan, W., Zhang, Y., Chiu, C.C., Zoph, B., Cubuk, E.D. and Le, Q.V., 2019. Specaugment: a simple data augmentation method for automatic speech recognition. In Proc. Interspeech 2019 (pp. 2613-2617)

2019
[25]

and Zoia, G., 2005

Scaringella, N. and Zoia, G., 2005. Modeling mu- sic genre evolution with an hmm-based system. In Proceedings of the 13th annual ACM interna- tional conference on Multimedia (pp. 143-146)

2005
[26]

and Batra, D., 2017

Selvaraju, R.R., Cogswell, M., Das, A., Vedan- tam, R., Parikh, D. and Batra, D., 2017. Grad- cam: Visual explanations from deep networks 15 via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (pp. 618-626)

2017
[27]

and Zisserman, A., 2014

Simonyan, K. and Zisserman, A., 2014. Very deep convolutional networks for large-scale im- age recognition. arXiv preprint arXiv:1409.1556

2014 arXiv
[28]

and Waibel, A., 1998

Soltau, H., Schultz, T., Westphal, M. and Waibel, A., 1998. Recognition of music types. In 1998 IEEE International Conference on Acous- tics, Speech and Signal Processing, ICASSP’98 (Cat. No. 98CH36181) (Vol. 2, pp. 1137-1140). IEEE. 16

1998
[2001]

Springer Series in Statistics

The elements of statistical learning. Springer Series in Statistics
[2012]

Advances in neural in- formation processing systems, 25

Imagenet classification with deep convolu- tional neural networks. Advances in neural in- formation processing systems, 25

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

and Yu, D., 2014

Abdel-Hamid, O., Mohamed, A., Jiang, H., Deng, L., Penn, G. and Yu, D., 2014. Convo- lutional neural networks for speech recognition. IEEE/ACM Transactions on audio, speech, and language processing, 22(10), pp.1533-1545

work page 2014

[2] [2]

Music Genre Classification using Machine Learning Techniques

Bahuleyan, H., 2018. Music genre classifica- tion using machine learning techniques. arXiv preprint arXiv:1804.01149

work page Pith review arXiv 2018

[3] [3]

and Kaestner, C.A., 2008, October

Silla Jr, C.N., Koerich, A.L. and Kaestner, C.A., 2008, October. A machine learning approach to automatic music genre classification. In Journal of the Brazilian Computer Society

work page 2008

[4] [4]

and Kumar, V.V., 2019, March

Chillara, S., Kavitha, K.R., Narayana, K.S., Varma, G.P.S. and Kumar, V.V., 2019, March. Music genre classification using machine learn- ing algorithms. In 2019 3rd International Con- ference on Trends in Electronics and Informatics (ICOEI) (pp. 515-519). IEEE

work page 2019

[5] [5]

and Aydin, N., 2020, September

Elbir, A. and Aydin, N., 2020, September. Music genre classification and recommendation by us- ing artificial intelligence. In 2020 International Conference on Data Analytics for Business and Industry: Way Towards a Sustainable Economy (ICDABI) (pp. 1-6). IEEE

work page 2020

[6] [6]

and Sharma, G., 2020

Ghildiyal, S. and Sharma, G., 2020. Music genre classification: A review. Vision, 4(3), p.42

work page 2020

[7] [7]

and Dogo, E.M., 2021

Ndou, M.S. and Dogo, E.M., 2021. Music Genre Classification Using Machine Learning

work page 2021

[8] [8]

and Mermelstein, P., 1990

Davis, S. and Mermelstein, P., 1990. Compari- son of parametric representations for monosyl- labic word recognition in continuously spoken sentences. In Readings in speech recognition(pp. 65-74). Morgan Kaufmann. 14

work page 1990