Lukthung Classification Using Neural Networks on Lyrics and Audios

Dittaya Wanvarie; Kasina Euchukanonchai; Kawisorn Kamtue; Naruemon Pratanwanich

REVIEW 3 major objections 6 minor 18 references

Lukthung Classification Using Neural Networks on Lyrics and Audios

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Combining lyrics and audio embeddings classifies Thai Lukthung songs with F1 0.86.

desk verdict A genuinely new Lukthung benchmark that fails to test its own never-seen-artist claim because the data is split at song level, not artist level, so the reported F1 is likely optimistic. read the letter →

arxiv 1908.08769 v2 pith:TYTZDTVI submitted 2019-08-23 cs.LG cs.SDeess.ASstat.ML

Kawisorn Kamtue , Kasina Euchukanonchai , Dittaya Wanvarie , Naruemon Pratanwanich This is my paper · ORCID

classification cs.LGcs.SDeess.ASstat.ML

keywords LukthungclassificationThaimusicgenrelyricsaudiospectrogramCNNbag-of-wordsfeaturefusioninformationretrieval

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 the Thai genre Lukthung can be reliably separated from other Thai popular music using only lyrics, only audio, or both. It claims that a simple bag-of-words lyric representation already outperforms standard audio-only classifiers, that a spectrogram CNN using specially shaped filters captures the timbral and rhythmic signature of Lukthung, and that the best result comes from concatenating the learned features of both models into a final classifier, reaching an F1 of about 0.86. The motivation is practical: streaming services that recommend by genre can use the model to avoid recommending Lukthung to listeners who never hear it. The result matters because it is a demonstration that genre classification can work on a non-Western, under-resourced music style using a relatively small dataset.

What carries the argument

The machinery is a two-stream feature-extraction architecture: a bag-of-words MLP for lyrics and a spectrogram CNN with parallel vertical and horizontal filters for audio, whose last-layer representations are concatenated and fed to a small feed-forward classifier. The vertical filters are designed to capture timbral features such as overtone stacks and vibrato, and the horizontal filters capture tempo and rhythm; fusing these learned representations is what carries the combined model's improvement.

What would settle it

Compute the same F1 on an artist-disjoint split, where all songs by one artist go into the same partition; a large drop would show that the 0.86 F1 is inflated by artist-specific cues rather than genre-level recognition.

Watch

Extended reading notes

Core claim

The paper reports that classifying Lukthung from other Thai genres is best done by combining lyrics and audio: the combined model reaches an F1 of 0.8561, beating the lyrics-only BoW-MLP (0.7905) and the audio-only Spectro-CNN (0.7730), and far exceeding classical MFCC baselines, including a random forest that achieves high precision but only 0.1313 recall. The authors also find that the learned features separate the classes in t-SNE, and that false negatives are mainly modern Lukthung songs resembling Pop or Rock and songs whose labels appear incorrect, while false positives are nearby genres such as Life and Lukkrung or non-Lukthung songs sung by Lukthung artists.

Load-bearing premise

The load-bearing premise is that the random song-level split into training and test sets measures genre-level generalization, even though no artist-disjoint split is reported and artists may appear in both sets.

Editorial extensions

If this is right

If the reported F1 holds, Thai streaming services can automatically tag Lukthung songs in large catalogs and exclude them from recommendations for non-Lukthung listeners.
The lyrics-only bag-of-words model demonstrates that a small dataset and simple text features can compete with deeper lyric models, avoiding the need for large pretrained word embeddings.
The error analysis suggests that isolating singing voice and modeling vocal accent or vibrato could improve classification of modern Lukthung songs that otherwise resemble Pop or Rock.
The learned feature representations could be reused to study how Lukthung changes across eras, since the features separate songs by genre and reveal boundary cases.
The combined architecture is a template for genre classification in other regional or under-resourced music traditions where both lyrics and audio are available.

Reading between the lines

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

The paper removes artist names from the lyrics but still evaluates on a random song-level split; a fair test of genre-level generalization would hold out whole artists, and without that split the 0.86 F1 could partially reflect artist identity rather than genre.
The same two-stream fusion could be tested on neighboring Thai genres such as Lukkrung and Mor-lam to see whether the learned features distinguish among country-like genres or only separate Lukthung from the rest.
Because the audio model was adapted from an architecture designed for western music tagging, its filter shapes may be implicitly tuned to the spectrogram statistics of that music; retuning them for Thai vocal vibrato could be a direct next step.
The paper's claim about complementary lyrics and audio features could be tested adversarially by ablating each stream in the combined model and measuring how much the F1 drops.

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 useful thing here is that the paper opens a new dataset and task: it is the first to train neural classifiers for Lukthung, a major Thai genre, using both lyrics and audio. The audio model adapts the Pons et al. filter design to preserve vibrato by using small frequency-band horizontal filters before pooling, which is a sensible domain adaptation. The combined model, which concatenates the last-layer features from the BoW-MLP and the Spectro-CNN, does beat both single-modality models on the reported test set (F1 0.856 vs 0.773 for audio and 0.791 for lyrics). That result, if reliable, matters for Thai streaming services.

The soft spot is the evaluation. The paper says the goal is to classify songs from unknown or never-seen-before artists, but the only split is a random 55/20/25 split of songs. With roughly 10.5k songs across six genres, the number of artists is small enough that the same artists almost certainly appear in train and test. The BoW model can memorize artist-specific vocabulary, and the audio model can exploit production and vocal signatures that are artist-specific rather than genre-specific. The paper's own error analysis supports this: it explicitly notes false positives that are "Lukthung songs sung with Pop/Rock artists," which is a direct sign that artist identity is a confound. So the central improvement claim is plausible but not established for the stated generalization scenario.

Other issues are more minor: no confidence intervals or significance tests; the baselines are audio-only with default scikit-learn parameters, so the "combined beats traditional" comparison is not against a strong multimodal baseline; and there is no lyrics baseline from the traditional-methods side. The chorus start is approximated at 30% of the song, which could add noise. The paper does not release code or data, but for a Tencent-related project that is not unusual.

My overall take: the work is a reasonable first application, not a methodological advance. The central flaw is fixable: add an artist-disjoint split, report variance or per-artist statistics, and include a standard lyrics baseline. I would not desk-reject it, but I would send it to review with the expectation of major revision, primarily to make the evaluation match the claim.

Who is this for? Researchers working on non-Western music classification, and to a lesser extent anyone who wants a concrete example of the artist-leakage problem in music tagging. The paper is worth reading for that cautionary lesson, and worth a serious referee.

Referee Report

3 major / 6 minor

Summary. The paper presents a binary classification system for distinguishing Lukthung, a Thai country music genre, from five other genres using lyrics and audio. A bag-of-words representation fed to a two-layer MLP is used for lyrics; a spectrogram-based CNN with vertical and horizontal filters is used for audio; and a combined model concatenates the last-layer features of both. The dataset consists of 10,547 expert-tagged Thai songs from 1985 to 2019. On a random 55/20/25 split, the combined model achieves F1 = 0.8561, outperforming the audio-only baselines (RF, LR, SVM) and the individual proposed models.

Significance. If the reported performance holds under rigorous evaluation, the paper would be a useful demonstration of multimodal genre classification for a non-Western music genre, with practical implications for music streaming services. The authors make several commendable choices: a sizable expert-annotated corpus, explicit architectural adaptations to Lukthung-specific traits such as vibrato and traditional instruments, and an analysis of false positives and false negatives. However, the evaluation protocol currently does not support the paper's generalization claim to never-before-seen artists, and the statistical significance of the gains is not established.

major comments (3)

[IV-A, VI] The paper claims in Section IV-A that removing artist names allows classification of 'songs from unknown or never-seen-before artists,' but Section VI evaluates only a random song-level split (55/20/25) with no artist-disjoint partition. Because lyrics vocabulary, songwriting habits, and audio production are artist-specific, a random split can place songs by the same artist in both training and test, allowing the models to memorize artist identity rather than genre. The error analysis in Section VI reports false positives including 'Lukthung songs sung with Pop/Rock artists,' which indicates that artist identity is a confound. The authors should either provide an artist-disjoint split with per-artist statistics or retract the generalization claim.
[VI, Table II] The F1 differences between the combined model (0.8561), BoW-MLP (0.7905), and Spectro-CNN (0.7730) are reported without confidence intervals, significance tests, or multiple runs. A single random split cannot establish that the observed gains are not due to chance; the authors should report mean and standard deviation over several seeds, or at least a paired significance test, to support the claim that combining features substantially improves performance.
[VI, Table II] The comparison baselines are limited to audio-only classifiers (RF, LR, SVM) with default Scikit-learn parameters and MFCC features; no lyrics-only baselines (e.g., TF-IDF with SVM) or standard multimodal baselines are included. Consequently, the abstract's claim that the proposed models 'outperform all of the standard classifiers' rests on a narrow set of weak baselines. Adding a lyrics baseline and a late-fusion baseline would make the comparison more informative and the claim more accurate.

minor comments (6)

[V-B] In Section V-B, the phrase 'We used the chorus part solely not only for computational reasons but also because...' is ungrammatical; please rephrase to 'not only... but also...'.
[References] Reference [16] misspells 'October' as 'Octoboer' and Reference [11] misspells 'International' as 'Internataional'.
[VI, Table II] Table II is garbled for the combined model row: the Input column reads 'Lyrics & Combined model' and 'Audios' appears after the F1 value. Please reformat the table.
[V-A, Eq. (1)] The normalization in Eq. (1) divides by $\max \log c_i$, which is undefined when the maximum word count of a song is 1; please clarify how such cases are handled.
[VI] The claim that the best scores are highlighted in bold should be verified; the bold formatting is not visible in the manuscript text, and the final version should ensure consistent highlighting.
[III] In Section III, 'There are few songs before 2000s' should be 'before the 2000s' or 'before 2000'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the F1 comparison is an empirical evaluation, not a quantity forced by definitions, fitted inputs, or self-citations.

full rationale

The paper's derivation chain is a standard supervised machine-learning pipeline: lyrics are converted to bag-of-words vectors and fed to an MLP; audio is converted to mel spectrograms and fed to a CNN; the combined model concatenates intermediate learned features and trains a feed-forward classifier. The reported F1 score of 0.856 is measured on a held-out test split and compared against baselines in Table II. No parameter is fitted to the test set and then renamed as a prediction, and no mathematical identity equates the evaluation quantity with a training objective. The claim that combining lyrics and audio features improves performance is an empirical observation from the table, not a consequence of the models' definitions. Citations such as [5], [17], and [18] are to external prior work, and the paper explicitly modifies the architecture from [5]; no load-bearing premise is justified solely by a self-citation. The artist-overlap concern about the random song-level split is a potential threat to external validity and generalization to never-seen-before artists, not a form of circularity: the evaluation remains an honest estimate on held-out songs, even if not on held-out artists. No equation is shown to equal another by construction, and no fitted input is presented as a prediction. Therefore no circular steps are identified, and the circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper's central performance claim depends on a string of hand-chosen architecture parameters and dataset assumptions. These are standard for empirical ML papers, but none are derived or independently validated, which caps the strength of the claim that the method generalizes beyond this private dataset.

free parameters (6)

BoW-MLP hidden layer size = 100 neurons per hidden layer
Chosen by hand; no architecture search or sensitivity analysis is reported, and capacity affects the lyrics classifier's F1.
Audio CNN filter configuration = sizes and counts in Table I (e.g., 115x7 x32, 7x165 x32)
Hand-designed based on [5] with modifications for vibrato; no ablation over filter shapes is shown.
Combined model input dimension = 800 (100 lyrics features + 700 audio features)
Determined by the last-layer sizes of the two submodels; no rationale is given for the exact concatenation dimensionality.
Vocabulary pruning thresholds = min count 10, max word length 20
Ad hoc thresholds applied to the 85k-song external corpus; effect on classification F1 is not evaluated.
Chorus start fraction = 30% of total song duration
Used to extract the 10-second audio clip; an approximation that could misalign the clip for songs with non-standard structure.
Spectrogram extraction parameters = 128 mel bins, n_fft 2048, hop 512, sample rate 22050 Hz, band 300-8000 Hz
Standard MIR choices adopted without comparative justification; the model input depends on them.

assumptions (5)

domain assumption The music expert genre labels in the dataset are correct enough to serve as ground truth.
Section III describes expert tagging; the paper's own false-negative analysis in Section VI notes that some non-Lukthung songs may be incorrectly labeled as Lukthung, so label noise affects the reported F1.
domain assumption A random song-level split into train/validation/test estimates performance on never-seen artists.
Section VI performs a random split; Section IV-A claims generalization to unknown artists, but no artist-disjoint split is implemented, so this assumption is unverified.
domain assumption The 10-second clip beginning at 30% of the song duration approximates the chorus with the richest instrumental content.
Stated in Section IV-B without validation; the audio model sees only this fixed excerpt.
domain assumption The externally built bag-of-words vocabulary from about 85k songs is a suitable feature space for lyrics.
Section V-A constructs the vocabulary from a larger unlabeled corpus; tokenizer and vocabulary effects on the downstream F1 are not ablated.
domain assumption Mel spectrograms capture the timbral and temporal cues (overtones, vibrato, rhythm) needed to separate Lukthung.
Section V-B justifies the CNN input visually; no comparison to other audio representations is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Lukthung Classification Using Neural Networks on Lyrics and Audios." pith.science (2026). https://pith.science/paper/TYTZDTVI

@misc{pith2026190808769,
  author       = {Pith},
  title        = {Pith review of: Lukthung Classification Using Neural Networks on Lyrics and Audios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TYTZDTVI}},
  note         = {Machine review of arXiv:1908.08769}
}

abstract

Music genre classification is a widely researched topic in music information retrieval (MIR). Being able to automatically tag genres will benefit music streaming service providers such as JOOX, Apple Music, and Spotify for their content-based recommendation. However, most studies on music classification have been done on western songs which differ from Thai songs. Lukthung, a distinctive and long-established type of Thai music, is one of the most popular music genres in Thailand and has a specific group of listeners. In this paper, we develop neural networks to classify such Lukthung genre from others using both lyrics and audios. Words used in Lukthung songs are particularly poetical, and their musical styles are uniquely composed of traditional Thai instruments. We leverage these two main characteristics by building a lyrics model based on bag-of-words (BoW), and an audio model using a convolutional neural network (CNN) architecture. We then aggregate the intermediate features learned from both models to build a final classifier. Our results show that the proposed three models outperform all of the standard classifiers where the combined model yields the best $F_1$ score of 0.86, allowing Lukthung classification to be applicable to personalized recommendation for Thai audience.

Figures

Figures reproduced from arXiv: 1908.08769 by the authors.

**Figure 2.** Overall model architecture [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Mel spectrograms (shown only half of the input time length) along [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Feature values of selected songs, taken from the input layer of the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Feature embeddings from the last hidden layer in the combined model [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 16 canonical work pages

[1]

Luk thung: The culture and politics of thailands most popular music,

B. Chan, “Luk thung: The culture and politics of thailands most popular music,” Journal of the Malaysian Branch of the Royal Asiatic Society , vol. 89, pp. 171–172, January 2016

work page 2016
[2]

Lukthung: Authenticity and modernity in thai country music,

A. Jirattikorn, “Lukthung: Authenticity and modernity in thai country music,” Asian Music , vol. 37, pp. 24–50, December 2006

work page 2006
[3]

Musical genre classiﬁcation of audio signals,

G. Tzanetakis and P. Cook, “Musical genre classiﬁcation of audio signals,” IEEE Transactions on Speech and Audio Processing , vol. 10, no. 5, pp. 293–302, July 2002

work page 2002
[4]

Music genre recognition using spectrograms,

Y . M. G. Costa, L. S. Oliveira, A. L. Koericb, and F. Gouyon, “Music genre recognition using spectrograms,” in 2011 18th International Conference on Systems, Signals and Image Processing (IWSSIP) , June 2011, pp. 1–4

work page 2011
[5]

End-to-end learning for music audio tagging at scale,

J. Pons, O. Nieto, M. Prockup, E. M. Schmidt, A. F. Ehmann, and X. Serra, “End-to-end learning for music audio tagging at scale,” CoRR, vol. abs/1711.02520, 2017

arXiv 2017
[6]

Music artist style identiﬁcation by semi- supervised learning from both lyrics and content,

T. Li and M. Ogihara, “Music artist style identiﬁcation by semi- supervised learning from both lyrics and content,” in Proceedings of the 12th ACM International Conference on Multimedia , January 2004, pp. 364–367

work page 2004
[7]

Lyrics-based analysis and classiﬁcation of music,

M. Fell and C. Sporleder, “Lyrics-based analysis and classiﬁcation of music,” in Proceedings of the 25th International Conference on Computational Linguistics: Technical Papers (COLING 2014) . Dublin, Ireland: Dublin City University and Association for Computational Linguistics, August 2014, pp. 620–631

work page 2014
[8]

Music genre classiﬁcation by ensembles of audio and lyrics features,

R. Mayer and A. Rauber, “Music genre classiﬁcation by ensembles of audio and lyrics features,” in Proceedings of the 12th International Society for Music Information Retrieval Conference (ISMIR 2011) , Miami, Florida, USA, October 2011, pp. 675–680

work page 2011

Show all 18 references

[9]

Turkish music genre classiﬁcation using audio and lyrics features,

¨O. oban, “Turkish music genre classiﬁcation using audio and lyrics features,” S¨uleyman Demirel University Journal of Natural and Applied Sciences, vol. 21, no. 2, pp. 322–331, May 2017

2017
[10]

Audio musical genre classiﬁcation using convolutional neural networks and pitch and tempo transformations,

L. Li, “Audio musical genre classiﬁcation using convolutional neural networks and pitch and tempo transformations,” Master’s thesis, City University of Hong Kong, Hong Kong, 2010

2010
[11]

Automatic musical pattern feature extraction using convolutional neural network,

T. L. Li, A. B. Chan, and A. H. Chun, “Automatic musical pattern feature extraction using convolutional neural network,” in Proceedings of the Internataional Conference on Data Mining and Applications (IMECS) , March 2010

2010
[12]

Audio music genre classiﬁcation using different classiﬁers and feature selection methods,

Y . Yaslan and Z. Cataltepe, “Audio music genre classiﬁcation using different classiﬁers and feature selection methods,” in Proceedings of the 18th International Conference on Pattern Recognition (ICPR 2006) , vol. 2, January 2006, pp. 573–576

2006
[13]

Deep convolutional neural networks for lvcsr,

T. N. Sainath, A.-r. Mohamed, B. Kingsbury, and B. Ramabhadran, “Deep convolutional neural networks for lvcsr,” in 2013 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP), May 2013, pp. 8614–8618

2013
[14]

End-to-end learning for music audio,

S. Dieleman and B. Schrauwen, “End-to-end learning for music audio,” in 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), May 2014, pp. 6964–6968

2014
[15]

A deep bag-of-features model for music auto-tagging,

J. Nam, J. Herrera, and K. Lee, “A deep bag-of-features model for music auto-tagging,” CoRR, vol. abs/1508.04999, 2016

2016 arXiv
[16]

Transfer learning by supervised pre-training for audio-based music classiﬁcation,

A. van den Oord, S. Dieleman, and B. Schrauwen, “Transfer learning by supervised pre-training for audio-based music classiﬁcation,” in Proceedings of the 15th Conference of the International Society for Music Information Retrieval, Proceedings (ISMIR) , Octoboer 2014, pp. 29–34

2014
[17]

Lyrics-based music genre classiﬁcation using a hier- archical attention network,

A. Tsaptsinos, “Lyrics-based music genre classiﬁcation using a hier- archical attention network,” in Proceedings of the 18th International Society for Music Information Retrieval Conference (ISMIR 2017) , Suzhou, China, October 2017, pp. 694–701

2017
[18]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2016, pp. 770–778

2016

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Luk thung: The culture and politics of thailands most popular music,

B. Chan, “Luk thung: The culture and politics of thailands most popular music,” Journal of the Malaysian Branch of the Royal Asiatic Society , vol. 89, pp. 171–172, January 2016

work page 2016

[2] [2]

Lukthung: Authenticity and modernity in thai country music,

A. Jirattikorn, “Lukthung: Authenticity and modernity in thai country music,” Asian Music , vol. 37, pp. 24–50, December 2006

work page 2006

[3] [3]

Musical genre classiﬁcation of audio signals,

G. Tzanetakis and P. Cook, “Musical genre classiﬁcation of audio signals,” IEEE Transactions on Speech and Audio Processing , vol. 10, no. 5, pp. 293–302, July 2002

work page 2002

[4] [4]

Music genre recognition using spectrograms,

Y . M. G. Costa, L. S. Oliveira, A. L. Koericb, and F. Gouyon, “Music genre recognition using spectrograms,” in 2011 18th International Conference on Systems, Signals and Image Processing (IWSSIP) , June 2011, pp. 1–4

work page 2011

[5] [5]

End-to-end learning for music audio tagging at scale,

J. Pons, O. Nieto, M. Prockup, E. M. Schmidt, A. F. Ehmann, and X. Serra, “End-to-end learning for music audio tagging at scale,” CoRR, vol. abs/1711.02520, 2017

arXiv 2017

[6] [6]

Music artist style identiﬁcation by semi- supervised learning from both lyrics and content,

T. Li and M. Ogihara, “Music artist style identiﬁcation by semi- supervised learning from both lyrics and content,” in Proceedings of the 12th ACM International Conference on Multimedia , January 2004, pp. 364–367

work page 2004

[7] [7]

Lyrics-based analysis and classiﬁcation of music,

M. Fell and C. Sporleder, “Lyrics-based analysis and classiﬁcation of music,” in Proceedings of the 25th International Conference on Computational Linguistics: Technical Papers (COLING 2014) . Dublin, Ireland: Dublin City University and Association for Computational Linguistics, August 2014, pp. 620–631

work page 2014

[8] [8]

Music genre classiﬁcation by ensembles of audio and lyrics features,

R. Mayer and A. Rauber, “Music genre classiﬁcation by ensembles of audio and lyrics features,” in Proceedings of the 12th International Society for Music Information Retrieval Conference (ISMIR 2011) , Miami, Florida, USA, October 2011, pp. 675–680

work page 2011