Towards Unified Music Emotion Recognition across Dimensional and Categorical Models

Dorien Herremans; Jaeyong Kang

REVIEW 1 major objections 4 minor 1 cited by

Towards Unified Music Emotion Recognition across Dimensional and Categorical Models

T0 review · 1 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that music emotion recognition can be unified across categorical tags and continuous valence-arousal ratings, and that training on both kinds of labels at once improves performance on both.

desk verdict The unified multitask setup is genuinely new and the Table III gains look real, but the knowledge distillation claim is mathematically underspecified and untested by ablation, so the paper overreaches as written. read the letter →

arxiv 2502.03979 v2 pith:HCNYNAYM submitted 2025-02-06 cs.SD cs.AIeess.AS

Jaeyong Kang , Dorien Herremans This is my paper · ORCID

classification cs.SDcs.AIeess.AS

keywords musicemotionrecognitionmultitasklearningknowledgedistillationvalence-arousalcategoricallabelsMERTembeddingschordandkeyfeaturesMTG-Jamendo

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 argues that music emotion recognition can be unified across two seemingly incompatible labeling schemes: categorical mood tags such as 'happy' or 'sad', and continuous valence-arousal ratings. The authors build a single multitask network trained on all four datasets at once, fed by audio embeddings from the large self-supervised music model MERT plus chord and key features, and guided by knowledge distillation from per-dataset teachers. The model reaches a PR-AUC of 0.1543 and ROC-AUC of 0.7810 on MTG-Jamendo, beating the best previous system, and training alongside MTG-Jamendo improves valence and arousal $R^2$ on DEAM, PMEmo, and EmoMusic relative to single-dataset training. If correct, the framework would let researchers pool heterogeneous emotion datasets instead of keeping them apart.

What carries the argument

The load-bearing object is a multitask student network with two output branches: one for multi-label mood classification over 56 categories, and one for continuous valence-arousal regression. Both branches consume the same fused representation built from MERT (a large self-supervised music audio model) layer-5 and layer-6 embeddings, averaged over randomly sampled consecutive 30-second segments, the CLS token of a two-layer Transformer encoding the chord progression, and a key embedding. Knowledge distillation couples the branches to per-dataset teachers: the total loss is $\alpha$ times the task loss plus $(1-\alpha)$ times a KL divergence between teacher and student soft outputs, with $\alpha=\beta=0.2$.

What would settle it

Inspect the released code for the dimensional branch of the distillation loss. If the raw valence and arousal scalars are fed directly into the KL-divergence formula in Equation (9), or if a distribution conversion is silently applied, then the objective is either undefined for non-positive values or differs from the paper's description; reproducing the reported $R^2$ improvements would then require fixing the loss or re-running the experiments.

Watch

Extended reading notes

Core claim

The paper's central claim is that categorical and dimensional emotion labels can be handled by one multitask network, and that combining the two label types during training improves performance on both. On the MTG-Jamendo benchmark, the unified model outperforms the winning entry of the 2021 emotion-and-theme competition and two recent transformer-based taggers; on the three dimensional datasets, it reports higher $R^2$ for both valence and arousal than training on each dataset alone. The gains come from three additions working together: a fused input representation (MERT embeddings concatenated with a Transformer-encoded chord progression and a key embedding), multitask training with a selective update so each dataset updates only its own branch, and knowledge distillation that transfers the soft outputs of per-dataset teacher models to the shared student.

Load-bearing premise

The load-bearing premise is that the knowledge-distillation loss for the valence-arousal branch is well defined, which requires converting scalar valence and arousal values into probability distributions; the paper does not specify how this conversion is done, and the formula it gives is undefined for ordinary scalar values.

Editorial extensions

If this is right

A single model can be trained on any mixture of categorical and dimensional emotion datasets, removing the need to discard or manually re-label data that use one scheme.
Adding valence-arousal datasets to the categorical MTG-Jamendo task raises PR-AUC and ROC-AUC, so dimensional labels carry information that helps mood tagging.
Adding the categorical dataset to each dimensional training set raises $R^2$ for valence and arousal, showing the transfer also runs in the opposite direction.
The reported MTG-Jamendo figures (PR-AUC 0.1543, ROC-AUC 0.7810) exceed the best results from the 2021 competition and recent baselines, giving a new reference point for that benchmark.
Because the input representation combines pre-trained audio embeddings with symbolic musical features, the framework permits controlled study of how harmonic and tonal structure contribute to emotion predictions.

Reading between the lines

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

Because each teacher is trained on the same dataset with the same architecture as the student, part of the distillation gain may be self-distillation; training teachers on disjoint data would reveal how much external knowledge is actually transferred.
The random-segment averaging augmentation exposes the model to different temporal contexts each epoch; ablating it against fixed-segment or whole-song pooling would isolate whether the reported gains come from this diversity or from the chord/key features.
The framework's label-unification idea could be carried to other affective domains where datasets use incompatible annotation schemes, such as speech or video emotion recognition.

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

The core idea here is legitimately new: no prior work trains a single MER model on both categorical and dimensional labels at once. The reader’s take is right that the combination is incremental, but it is not trivial—the selective-update strategy and shared representation across heterogeneous label types is a useful engineering contribution. The paper also does something right in the input representation: layering chord/key features on MERT embeddings gives a consistent small boost in Table II, and the code is public. I believe the multitask numbers in Table III, where adding datasets improves held-out metrics across all four datasets, are the real result and the reason this paper should not be ignored.

The soft spots are equally real. The stress-test note correctly nails Eq. (9): the KD loss is written as a KL divergence over distributions, but the paper never says how continuous valence/arousal scalars become distributions. As written, the expression t_i log(t_i/s_i) is undefined for non-positive student outputs, and it is not a KL divergence for two scalars. That is not a tiny notational slip; it means every reported R2 gain that relies on L_KD,VA is built on a loss that has no clear mathematical meaning. The teacher models are also trained on the same data with the same architecture, so any KD benefit would be self-distillation, and there is no ablation that isolates KD. The abstract’s claim that KD “significantly enhances” performance is not supported by the experiments.

The SOTA comparison on MTG-Jamendo is a secondary concern: the margin over lileonardo is about 0.0035 in PR-AUC, and there are no error bars or significance tests anywhere. Still, the improvement is in the right direction and the official split makes the comparison fair. The paper would be stronger if it frankly presented the MTL gains as the contribution and treated KD as a secondary regularizer.

Who is this for? People working on MER or affective computing with heterogeneous labels. It is a workshop-quality idea that, if fixed, could be solid accepted work. The KD loss needs to be either specified properly (e.g., Gaussian KL or a different regression distillation loss) or removed as a claim, and the authors need a KD ablation plus variance reporting. This is fixable.

I would send it to peer review. A competent referee can help the authors either repair the KD math or drop the overclaim, and the multitask result deserves to be in the literature.

Referee Report

1 major / 4 minor

Summary. The manuscript proposes a unified multitask learning framework for Music Emotion Recognition (MER) that trains jointly on categorical labels (MTG-Jamendo) and dimensional valence-arousal labels (DEAM, PMEmo, EmoMusic). The input representation combines MERT embeddings with chord-progression and key features, and knowledge distillation (KD) is used to transfer knowledge from per-dataset teacher models to a multitask student model. The authors report state-of-the-art performance on MTG-Jamendo (PR-AUC 0.1543, ROC-AUC 0.7810) and improved R2 scores on the three dimensional datasets when training on all four datasets.

Significance. If validated, the idea of unifying categorical and dimensional emotion labels in a single multitask architecture is a useful step for MER, which has been hampered by heterogeneous labeling schemes. The combination of MERT embeddings with chord and key features is a sensible and potentially effective representation, and the reported SOTA on the official MTG-Jamendo test split is a concrete empirical contribution. The authors also make their code publicly available, which supports reproducibility. However, the knowledge-distillation component is currently not well specified mathematically and is not isolated by ablations, so the significance of that particular contribution is not yet established.

major comments (1)

[Section III-C, Eq. (9)] Tables II-IV report single point estimates with no error bars, standard deviations, or significance tests. The improvements are small; for example, MTG-Jamendo PR-AUC varies from 0.1521 to 0.1543 across training-set configurations, and the SOTA margin over lileonardo is 0.0035. Without multiple seeds or statistical testing, it is not clear whether the multitask gains or the SOTA claim are robust. The authors should report mean±std over at least several runs and, if possible, confidence intervals or paired significance tests for the key comparisons.

minor comments (4)

[Section II.E / References] The MERT model is cited both as [29] and [54]; these references should be consolidated to avoid confusion.
[Section IV.A, Table I] The 'Length' column in Table I reports 45s for DEAM and EmoMusic but 'full' for MTG-Jamendo and PMEmo. It would help to clarify whether static annotations are computed over the full track or over a 45-s segment, since this affects comparability.
[Section III.B] The 'selective update strategy' is described in words, but the paper does not specify how gradients are blocked for the non-active branch (e.g., stop-gradient, loss masking). A precise statement would improve reproducibility.
[Section III.A.2] There is a small typographical issue in the sentence 'a C minor 7th chord comprises the notes C, E, G, B' — the flat sign is missing; the intended spelling is 'B♭'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central MTG-Jamendo claim is benchmarked on official held-out splits against external baselines; the same-data KD setup is an attribution gap, not a circular derivation.

full rationale

The paper's primary claims are empirical and externally falsifiable. The MTG-Jamendo result (PR-AUC 0.1543, ROC-AUC 0.7810, Table IV) is compared with MediaEval 2021 baselines on the official train/validation/test splits, so the headline 'state-of-the-art' result does not reduce to a fitted parameter or to a self-citation. The feature ablation (Table II) is a controlled comparison with the same training setup, and the multitask comparison (Table III) evaluates on held-out test sets. The only plausible circularity concern is the knowledge-distillation component: Eq. (9) defines LKD as a KL divergence over distributions while Section III-C states c=2 for valence-arousal regression, and Section IV-B says teachers are trained on the same datasets with the same architecture as the student. If the VA branch has no distributional reparameterization, Eq. (9) is mathematically underspecified for that branch, and the same-data teachers mean any KD benefit would be self-distillation rather than transfer of external knowledge. However, this is a flaw in attributing the improvement to KD (no ablation with α=β=1 is reported), not a circularity in the derivation: the final test-set numbers are not forced by construction, and the categorical MTG-Jamendo result does not depend on the KD loss for its external comparison. The self-citations ([28], [30]) are background/survey references and are not load-bearing for the claimed results. No step in the paper's derivation chain equates a predicted quantity to its own input by definition.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The empirical claims rest on hand-set hyperparameters, a data-dependent MERT layer choice, and domain assumptions about chord/key features and the applicability of KL divergence to regression targets.

free parameters (3)

alpha (KD weight for categorical loss) = 0.2
Controls the balance between BCE and KD loss in Eq. (10). Set without reported sensitivity analysis or validation-based selection.
beta (KD weight for dimensional loss) = 0.2
Controls the balance between MSE and KD loss in Eq. (11). Set without reported sensitivity analysis.
MERT layer pair = 5th and 6th layers
Selected via preliminary experimentation (Section III.A.1). This data-dependent choice can inflate test metrics if the selection used test information.

assumptions (4)

ad hoc to paper KL divergence can be applied to valence-arousal regression outputs as probability distributions (Eq. 9 with c=2).
Continuous VA scalars are not probability distributions; the paper never specifies a transformation to distributions, so this assumption is specific to the proposed method.
domain assumption Chord recognition errors do not materially affect downstream emotion prediction.
Section III.A.2 states chord recognition WCSR scores are 'acceptable' despite confusions such as A minor vs C major, without a sensitivity analysis.
domain assumption MERT embeddings capture sufficient musical semantics for MER.
The framework relies on MERT as the primary audio representation based on prior results [54], without independent verification in this paper.
domain assumption Static valence-arousal annotations and random 70/15/15 splits for DEAM, PMEmo, and EmoMusic are valid for benchmarking.
Section IV.A uses static labels and random splits because official splits do not exist; the paper acknowledges this makes cross-study comparison inconsistent.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Unified Music Emotion Recognition across Dimensional and Categorical Models." pith.science (2026). https://pith.science/paper/HCNYNAYM

@misc{pith2026250203979,
  author       = {Pith},
  title        = {Pith review of: Towards Unified Music Emotion Recognition across Dimensional and Categorical Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HCNYNAYM}},
  note         = {Machine review of arXiv:2502.03979}
}

read the original abstract

One of the most significant challenges in Music Emotion Recognition (MER) comes from the fact that emotion labels can be heterogeneous across datasets with regard to the emotion representation, including categorical (e.g., happy, sad) versus dimensional labels (e.g., valence-arousal). In this paper, we present a unified multitask learning framework that combines these two types of labels and is thus able to be trained on multiple datasets. This framework uses an effective input representation that combines musical features (i.e., key and chords) and MERT embeddings. Moreover, knowledge distillation is employed to transfer the knowledge of teacher models trained on individual datasets to a student model, enhancing its ability to generalize across multiple tasks. To validate our proposed framework, we conducted extensive experiments on a variety of datasets, including MTG-Jamendo, DEAM, PMEmo, and EmoMusic. According to our experimental results, the inclusion of musical features, multitask learning, and knowledge distillation significantly enhances performance. In particular, our model outperforms the state-of-the-art models, including the best-performing model from the MediaEval 2021 competition on the MTG-Jamendo dataset. Our work makes a significant contribution to MER by allowing the combination of categorical and dimensional emotion labels in one unified framework, thus enabling training across datasets.

Figures

Figures reproduced from arXiv: 2502.03979 by the authors.

**Figure 2.** MERT feature extraction and data augmentation workflow. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Content filtering methods for music recommendation: A review
cs.IR 2025-07 conditional

A survey of content-based music recommendation methods, including audio analysis, lyrics analysis, and context awareness, with no new experimental results.

Reference graph

Works this paper leans on

55 extracted references · 49 canonical work pages · cited by 1 Pith paper

[1]

K. R. Agres, R. S. Schaefer, A. V olk, S. van Hooren, A. Holzapfel, S. Dalla Bella, M. M ¨uller, M. De Witte, D. Herremans, R. Ramirez Me- lendez, et al. Music, computing, and health: a roadmap for the current and future roles of music technology for health care and well-being. Music & Science , 4:2059204321997709, 2021

work page 2021
[2]

Aljanaki, F

A. Aljanaki, F. Wiering, and R. C. Veltkamp. Studying emotion induced by music through a crowdsourcing game. Information Processing & Management, 52(1):115–128, 2016

work page 2016
[3]

Aljanaki, Y .-H

A. Aljanaki, Y .-H. Yang, and M. Soleymani. Developing a benchmark for emotional analysis of music. PloS one, 12(3):e0173392, 2017

work page 2017
[4]

Bogdanov, M

D. Bogdanov, M. Won, P. Tovstogan, A. Porter, and X. Serra. The mtg-jamendo dataset for automatic music tagging. ICML, 2019

work page 2019
[5]

V . Bour. Frequency dependent convolutions for music tagging. In MediaEval, 2021

work page 2021
[6]

X. Cai, J. Yuan, R. Zheng, L. Huang, and K. Church. Speech emotion recognition with multi-task learning. In Interspeech, volume 2021, pages 4508–4512. Brno, 2021

work page 2021
[7]

C ¸ ano and M

E. C ¸ ano and M. Morisio. Moodylyrics: A sentiment annotated lyrics dataset. In Proceedings of the 2017 international conference on intelligent systems, metaheuristics & swarm intelligence , pages 118– 124, 2017

work page 2017
[8]

R. Caruana. Multitask learning. Machine learning, 28:41–75, 1997

work page 1997

Show all 55 references

[9]

Chaki, P

S. Chaki, P. Doshi, P. Patnaik, and S. Bhattacharya. Attentive rnns for continuous-time emotion prediction in music clips. In AffCon@ AAAI, pages 36–46, 2020

2020
[10]

Chen and Q

C. Chen and Q. Li. A multimodal music emotion classification method based on multifeature combined network classifier. Mathematical Problems in Engineering , 2020(1):4606027, 2020

2020
[11]

Y .-H. Cho, H. Lim, D.-W. Kim, and I.-K. Lee. Music emotion recognition using chord progressions. In 2016 IEEE Int. Conf. on Systems, Man, and Cybernetics (SMC) , pages 002588–002593. IEEE, 2016

2016
[12]

M. S. Cuthbert and C. Ariza. music21: A toolkit for computer-aided musicology and symbolic music data. 2010

2010
[13]

Davis and M

J. Davis and M. Goadrich. The relationship between precision-recall and roc curves. In Proc. of the 23rd Int. conference on Machine learning , pages 233–240, 2006

2006
[14]

Di Giorgi, M

B. Di Giorgi, M. Zanoni, A. Sarti, and S. Tubaro. Automatic chord recognition based on the probabilistic modeling of diatonic modal harmony. In nDS’13; Proceedings of the 8th International Workshop on Multidimensional Systems , pages 1–6. VDE, 2013

2013
[15]

N. Draper. Applied regression analysis. McGraw-Hill. Inc, 1998

1998
[16]

Elizalde, S

B. Elizalde, S. Deshmukh, M. Al Ismail, and H. Wang. Clap learning audio concepts from natural language supervision. In ICASSP 2023-2023 IEEE Int. Conf. on Acoustics, Speech and Signal Processing (ICASSP) , pages 1–5. IEEE, 2023

2023
[17]

L. N. Ferreira and J. Whitehead. Learning to generate music with sentiment. arXiv:2103.06125, 2021

2021 arXiv
[18]

Greer, X

T. Greer, X. Shi, B. Ma, and S. Narayanan. Creating musical features us- ing multi-faceted, multi-task encoders based on transformers. Scientific Reports, 13(1):10713, 2023

2023
[19]

Hasumi, T

T. Hasumi, T. Komatsu, and Y . Fujita. Music tagging with classifier group chains. arXiv:2501.05050, 2025

2025 arXiv
[20]

Hershey, S

S. Hershey, S. Chaudhuri, D. P. Ellis, J. F. Gemmeke, A. Jansen, R. C. Moore, M. Plakal, D. Platt, R. A. Saurous, B. Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 ieee Int. conference on acoustics, speech and signal processing (icassp) , pages...

2017
[21]

G. Hinton. Distilling the knowledge in a neural network. arXiv:1503.02531, 2015

2015 arXiv
[22]

Hizlisoy, S

S. Hizlisoy, S. Yildirim, and Z. Tufekci. Music emotion recognition using convolutional long short term memory deep neural networks. Eng. Sci. Technol. Int J. , 24(3):760–767, 2021

2021
[23]

Huang, S

Z. Huang, S. Ji, Z. Hu, C. Cai, J. Luo, and X. Yang. Adff: Attention based deep feature fusion approach for music emotion recognition. arXiv:2204.05649, 2022

2022 arXiv
[24]

H.-T. Hung, J. Ching, S. Doh, N. Kim, J. Nam, and Y .-H. Yang. Emopia: A multi-modal pop piano dataset for emotion recognition and emotion- based music generation. arXiv:2108.01374, 2021

2021 arXiv
[25]

Jeong, G

E. Jeong, G. Oh, and S. Lim. Multitask emotion recognition model with knowledge distillation and task discriminator. arXiv:2203.13072, 2022

2022 arXiv
[26]

X. Jia. A music emotion classification model based on the improved convolutional neural network. Computational Intelligence and Neuro- science, 2022(1):6749622, 2022

2022
[27]

Jonggwon, C

P. Jonggwon, C. Kyoyun, J. Sungwook, K. Dokyun, and P. Jonghun. A bi-directional transformer for musical chord recognition. In 20th Inter- national Society for Music Information Retrieval Conference (ISMIR), Delft, The Netherlands , 2019

2019
[28]

Kang and D

J. Kang and D. Herremans. Are we there yet? a brief survey of music emotion prediction datasets, models and outstanding challenges. arXiv:2406.08809, 2024

2024 arXiv
[29]

Y . Li, R. Yuan, G. Zhang, Y . Ma, X. Chen, H. Yin, C. Xiao, C. Lin, A. Ragni, E. Benetos, et al. Mert: Acoustic music understanding model with large-scale self-supervised training. arXiv:2306.00107, 2023

2023 arXiv
[30]

R. Liu, A. Roy, and D. Herremans. Leveraging llm embeddings for cross dataset label alignment and zero shot music emotion prediction. arXiv:2410.11522, 2024

2024 arXiv
[31]

X. Liu, Q. Chen, X. Wu, Y . Liu, and Y . Liu. Cnn based music emotion classification. arXiv:1704.05665, 2017

2017 arXiv
[32]

Makris, K

D. Makris, K. R. Agres, and D. Herremans. Generating lead sheets with affect: A novel conditional seq2seq framework. In 2021 Int. Joint Conf. on Neural Networks (IJCNN) , pages 1–8. IEEE, 2021

2021
[33]

Malheiro, R

R. Malheiro, R. Panda, P. J. Gomes, and R. P. Paiva. Bi-modal music emotion recognition: Novel lyrical features and dataset. In 9th Int. Workshop on Music and Machine Learning–MML , 2016

2016
[34]

Mayerl, M

M. Mayerl, M. V ¨otter, A. Peintner, G. Specht, and E. Zangerle. Recognizing song mood and theme: Clustering-based ensembles. In MediaEval, 2021

2021
[35]

Mazzetta, A

G. Mazzetta, A. Greco, M. Tagliasacchi, and A. Pescap `e. A multi- source deep learning model for music emotion recognition. In Proc. of the AIxHMI Workshop at CEUR-WS , volume 3903, pages 1–8. CEUR Workshop Proc., 2024

2024
[36]

L. B. Meyer. Emotion and meaning in music. PhD thesis, The University of Chicago, 1954

1954
[37]

Ong and D

J. Ong and D. Herremans. Constructing time-series momentum portfo- lios with deep multi-task learning. Expert Systems with Applications , 230:120587, 2023

2023
[38]

Panda, R

R. Panda, R. Malheiro, and R. P. Paiva. Musical texture and expressivity features for music emotion recognition. In 19th Int. Society for Music Information Retrieval Conf. (ISMIR 2018) , pages 383–391, 2018

2018
[39]

Panda, R

R. Panda, R. Malheiro, and R. P. Paiva. Novel audio features for music emotion recognition. IEEE Transactions on Affective Computing , 11(4):614–626, 2018

2018
[40]

Pham, M.-H

P.-T. Pham, M.-H. Huynh, H.-D. Nguyen, and M.-T. Tran. Selab-hcmus at mediaeval 2021: Music theme and emotion classification with co- teaching training strategy. In MediaEval, 2021

2021
[41]

J. Qiu, C. Chen, and T. Zhang. A novel multi-task learning method for symbolic music emotion recognition. arXiv:2201.05782, 2022

2022 arXiv
[42]

Raffel, B

C. Raffel, B. McFee, E. J. Humphrey, J. Salamon, O. Nieto, D. Liang, D. P. Ellis, and C. C. Raffel. Mir eval: A transparent implementation of common mir metrics. In ISMIR, volume 10, page 2014, 2014

2014
[43]

Rajesh and N

S. Rajesh and N. Nalini. Musical instrument emotion recognition using deep recurrent neural network. Procedia Comput. Sci., 167:16–25, 2020

2020
[44]

J. A. Russell. A circumplex model of affect. Journal of personality and social psychology, 39(6):1161, 1980

1980
[45]

Shelke and M

S. Shelke and M. Patil. Exploring machine learning techniques for music emotion classification: A comprehensive review. In 2024 11th Int. Conf. on Computing for Sustainable Global Development (INDIACom) , pages 1188–1195. IEEE, 2024

2024
[46]

Soleymani, M

M. Soleymani, M. N. Caro, E. M. Schmidt, C.-Y . Sha, and Y .-H. Yang. 1000 songs for emotional analysis of music. In Proc. of the 2nd ACM Int. workshop on Crowdsourcing for multimedia , pages 1–6, 2013

2013
[47]

S. A. Suresh Kumar and R. Rajan. Transformer-based automatic music mood classification using multi-modal framework. Journal of Computer Science & Technology, 23, 2023

2023
[48]

H. H. Tan. Semi-supervised music emotion recognition using noisy student training and harmonic pitch class profiles. arXiv:2112.00702, 2021

2021 arXiv
[49]

G. Tong. Multimodal music emotion recognition method based on the combination of knowledge distillation and transfer learning. Scientific Programming, 2022(1):2802573, 2022

2022
[50]

Tovstogan, D

P. Tovstogan, D. Bogdanov, and A. Porter. Mediaeval 2021: Emotion and theme recognition in music using jamendo. In MediaEval, 2021

2021
[51]

H. Tran, T. Le, A. Do, T. Vu, S. Bogaerts, and B. Howard. Emotion- aware music recommendation. In Proc. of the AAAI Conf. on Artificial Intelligence, volume 37, pages 16087–16095, 2023

2023
[52]

Turnbull, L

D. Turnbull, L. Barrington, D. Torres, and G. Lanckriet. Towards musical query-by-semantic-description using the cal500 data set. In Proc. of the 30th annual Int. ACM SIGIR Conf. on Research and development in information retrieval, pages 439–446, 2007

2007
[53]

Wang, J.-C

S.-Y . Wang, J.-C. Wang, Y .-H. Yang, and H.-M. Wang. Towards time- varying music auto-tagging based on cal500 expansion. In 2014 IEEE Int. Conf. on Multimedia and Expo (ICME) , pages 1–6. IEEE, 2014

2014
[54]

Yizhi, R

L. Yizhi, R. Yuan, G. Zhang, Y . Ma, X. Chen, H. Yin, C. Xiao, C. Lin, A. Ragni, E. Benetos, et al. Mert: Acoustic music understanding model with large-scale self-supervised training. In The Twelfth International Conference on Learning Representations , 2023

2023
[55]

Zhang, H

K. Zhang, H. Zhang, S. Li, C. Yang, and L. Sun. The pmemo dataset for music emotion recognition. In Proc. of the 2018 acm on Int. Conf. on multimedia retrieval , pages 135–142, 2018

2018

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

K. R. Agres, R. S. Schaefer, A. V olk, S. van Hooren, A. Holzapfel, S. Dalla Bella, M. M ¨uller, M. De Witte, D. Herremans, R. Ramirez Me- lendez, et al. Music, computing, and health: a roadmap for the current and future roles of music technology for health care and well-being. Music & Science , 4:2059204321997709, 2021

work page 2021

[2] [2]

Aljanaki, F

A. Aljanaki, F. Wiering, and R. C. Veltkamp. Studying emotion induced by music through a crowdsourcing game. Information Processing & Management, 52(1):115–128, 2016

work page 2016

[3] [3]

Aljanaki, Y .-H

A. Aljanaki, Y .-H. Yang, and M. Soleymani. Developing a benchmark for emotional analysis of music. PloS one, 12(3):e0173392, 2017

work page 2017

[4] [4]

Bogdanov, M

D. Bogdanov, M. Won, P. Tovstogan, A. Porter, and X. Serra. The mtg-jamendo dataset for automatic music tagging. ICML, 2019

work page 2019

[5] [5]

V . Bour. Frequency dependent convolutions for music tagging. In MediaEval, 2021

work page 2021

[6] [6]

X. Cai, J. Yuan, R. Zheng, L. Huang, and K. Church. Speech emotion recognition with multi-task learning. In Interspeech, volume 2021, pages 4508–4512. Brno, 2021

work page 2021

[7] [7]

C ¸ ano and M

E. C ¸ ano and M. Morisio. Moodylyrics: A sentiment annotated lyrics dataset. In Proceedings of the 2017 international conference on intelligent systems, metaheuristics & swarm intelligence , pages 118– 124, 2017

work page 2017

[8] [8]

R. Caruana. Multitask learning. Machine learning, 28:41–75, 1997

work page 1997