AI TrackMate: Finally, Someone Who Will Give Your Music More Than Just "Sounds Great!"

Bo-Yu Chen; Chia-Ho Hsiung; Lu-Rong Chen; Yen-Tung Yeh; Yi-Lin Jiang

REVIEW 4 major objections 5 minor 14 references

AI TrackMate: Finally, Someone Who Will Give Your Music More Than Just "Sounds Great!"

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read This paper claims that an LLM supplied with a structured audio report can give music producers useful, track-specific feedback with no training.

desk verdict A likable, honest system demo whose central claim about audio-grounded feedback is plausible but unproven; deserves referee time, not acceptance on the evidence offered. read the letter →

arxiv 2412.06617 v1 pith:UWJGNBWD submitted 2024-12-09 cs.SD cs.HCcs.LGcs.MMeess.AS

Yi-Lin Jiang , Chia-Ho Hsiung , Yen-Tung Yeh , Lu-Rong Chen , Bo-Yu Chen This is my paper · ORCID

classification cs.SDcs.HCcs.LGcs.MMeess.AS

keywords musicproductionfeedbacklargelanguagemodelsaudioanalysisbedroomproducerspromptengineeringGraph-of-Thoughtchordrecognitioninformationretrieval

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

AI TrackMate is proposed as a training-free, plug-and-play system for giving music producers structured feedback on the audio they actually upload. The paper's central claim is that Large Language Models already have enough musical knowledge to act like a producer once raw audio is converted into a readable analysis report, so the system only needs prompting, not fine-tuning. The system analyzes rhythm, harmony, timbre, instrumentation, genre, theme, and emotion with existing audio-analysis tools, then uses a production-oriented prompt to score the track and suggest improvements. The paper demonstrates the system in a web interface and reports an exploratory pilot study with one producer, treating this as a first indication that the approach yields practical, objective self-assessment for bedroom producers.

What carries the argument

The central mechanism is the LLM-Readable Music Report, a structured summary that turns raw audio-analysis output into something an LLM can reason over. Raw metadata, such as time-stamped chord labels, is fed to the LLM; when early interpretations are too shallow, the system computes additional statistical metrics and iterates two to three times until the report includes nuanced features like common chord progressions and pacing. A second LLM evaluates each iteration and selects the clearest, most relevant representation. The report is paired with a Music Production-Oriented Feedback Instruction that uses Graph-of-Thought prompting, person-switching grammar, and a rubric designed to suppress generic praise and force balanced, critical, producer-like feedback. Because everything is done through prompts and reports rather than gradient updates, the system is compatible with any LLM.

What would settle it

Take a set of tracks with known ground-truth chord progressions, keys, and tempos, run them through the analysis module, and check whether the LLM's feedback changes when the numbers in the music report are deliberately corrupted; if the feedback is essentially unchanged, the audio module is not carrying the claimed load.

Watch

Extended reading notes

Core claim

The core discovery the authors are trying to establish is that production-specific musical feedback can be obtained without training or fine-tuning by pairing an LLM's pre-existing knowledge with direct analysis of the user's audio file. Audio is passed through analysis tools, and the resulting metadata is refined into an LLM-Readable Music Report with statistics such as dominant chords, chord-change counts, average chord duration, and common progressions, along with beat, key, tempo, timbre, structure, and emotion information. The LLM is then steered by a Music Production-Oriented Feedback Instruction that applies Graph-of-Thought reasoning and a scoring rubric covering Creativity and Originality, Genre Fidelity, Conveyability, Musical Richness, and Track Memorability. In the pilot dialogue, 75% of the AI's responses combined technical suggestions with emotional or perceptual feedback, which the authors present as evidence that the system can connect objective analysis with artistic intent.

Load-bearing premise

The claim stands on the assumption that the off-the-shelf audio analysis tools are accurate enough on arbitrary user uploads, and that LLMs' built-in musical knowledge is strong enough to turn those numbers into trustworthy producer-level advice.

Editorial extensions

If this is right

A producer can upload an audio file or a YouTube link and receive a structured score and concrete improvement suggestions without any model training.
The same pipeline should work with newer or differently sized LLMs, since only the prompt and the report change.
The feedback loop is conversational, so a producer can ask follow-up questions about chord changes, emotional flow, or mixing decisions after the initial analysis.
Because the prompting layer is separate from the analysis layer, future improvements in either audio analysis or LLM capability can be absorbed without redesigning the system.
Extending the system to more genres, real-time analysis, and DAW integration are directions the authors explicitly identify as next steps.

Reading between the lines

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

If the pipeline works as claimed, a natural next step is to expose the underlying statistics to the user, so producers can verify whether the AI's advice actually follows from the analysis rather than from the LLM's musical priors.
The same report-and-prompt structure could be turned into a comparative tool, letting producers A/B test two mixes or ask how a suggested chord change would alter the reported emotional profile.
A measurable consequence of the architecture is that feedback quality should track the accuracy of the audio-analysis tools, so genres or production styles those tools struggle with should produce weaker advice even when the LLM itself knows the genre well.

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

Readable demo paper, not a validation study. It does one thing well: describes a plug-and-play, training-free system that turns standard MIR outputs into an LLM-readable report and uses prompt engineering to get production-style feedback. The iterative refinement loop—feed raw features, add derived stats, let a second LLM pick the clearest rendering—is a sensible practical trick. The pilot transcript gives a flavor of the interaction, and the authors are careful to call it a pilot.

The soft spot is exactly where the stress-test puts it: the claim that insights are 'production-specific' because they come from the audio analysis is never demonstrated. No accuracy numbers for All-In-One, Madmom, autochord, or Essentia on the demo tracks. No ablation where the LLM gets the report versus just the prompt and the user's question. The single producer's self-report doesn't tell us whether the feedback is grounded or just plausible-sounding. Also §2.2 says the iterative refinement 'demonstrably enhances' understanding, but there's no comparison shown. That's a minor overstatement, not a fraud.

In proportion: this is an honest system demo with a modest pilot, not a rigorous evaluation. The architecture is clear and the limitations section is unusually candid. The authors don't claim more than exploratory. So the right verdict is: the paper is acceptable as a demonstration, but not as evidence that the system works as advertised until there's a real evaluation.

The secondary-LLM selection step is a mild self-reference, but it's not a derivation, so it doesn't bother me much. The citation to ComposerX as evidence that LLMs have inherent musical knowledge is thin, but again minor.

This paper is for people building AI-assisted music tools and for workshop/demo audiences. I'd bring it to a reading group discussion on LLM grounding, but I wouldn't cite it as evidence in my own work. I would send it to a serious referee: it's a working system with a live demo and a clear writeup, and the field benefits from such demos. But the referee should ask for a baseline that strips the report, accuracy checks on the MIR features, and a small multi-user study before acceptance.

Referee Report

4 major / 5 minor

Summary. The paper presents AI TrackMate, a training-free LLM-based chatbot that gives music-production feedback by combining audio analysis tools (All-In-One, Madmom, autochord, Essentia, AudioCommons timbral models) with an LLM. The pipeline consists of a Music Analysis Module that produces a structured report, an LLM-Readable Music Report built through iterative refinement, and Music Production-Oriented Feedback Instructions using Graph-of-Thought, person switching, and a custom feedback mechanism. The authors describe a web interface, report a pilot study with one music producer, and claim that this approach provides production-specific insights that distinguish it from text-only LLM systems. The paper concludes with limitations and future work directions.

Significance. If the load-bearing assumptions hold—namely that the MIR tools are accurate on arbitrary user uploads and that the LLM's feedback is genuinely driven by the audio-derived report—the system could be a useful, plug-and-play assistant for independent producers. The paper has clear strengths: it integrates established MIR libraries, provides a public demo link, explicitly discusses limitations, and avoids claiming that the system replaces human expertise. However, the evidence presented is anecdotal. The central claim that AI TrackMate offers production-specific feedback is supported only by self-reported observations from a single producer and by illustrative LLM dialogue excerpts. No quantitative evaluation of the analysis modules, no ablation of the audio-report component, and no inter-rater or accuracy checks are provided.

major comments (4)

[§2.1 and §2.2] The load-bearing premise is that the MIR tools—All-In-One, Madmom, autochord, Essentia, and the AudioCommons timbral models—produce accurate features on arbitrary user uploads or YouTube audio. The paper reports no accuracy numbers for any of these tools on the demo tracks or in the pilot study. Many of these tools are known to be brittle outside their training distribution, and YouTube rips add compression artifacts. If the report is noisy, the LLM is reasoning from incorrect data, which directly undermines the production-specific claim. Please add per-module accuracy or error-rate measurements against ground truth for a representative set of tracks, or at least a human sanity-check of the extracted features for the tracks used in the pilot.
[§2.2] The sentence 'This method demonstrably enhances the LLM's capacity to generate meaningful, musician-relevant insights from complex musical data' is not supported by any comparison. The iterative refinement is judged by 'a secondary LLM' that selects the most insightful representation, which introduces a self-referential risk: the same model class is evaluating its own output. No examples of the per-iteration reports, no defined selection criteria, and no inter-annotator agreement are provided. Please show the report versions across iterations and evaluate the final output against a human-annotated gold standard or at least a fixed rubric.
[§3.2 and §3.3] The pilot study consists of a single producer's self-report and an illustrative dialogue. There is no baseline condition, no blinded comparison, and no evidence that the reported 'production-specific' details would not also appear if the audio report were omitted. Since the paper's distinguishing claim is the combination of audio analysis with LLM knowledge, an ablation is necessary: feed the same user queries to the LLM without the audio-derived report and compare the specificity, correctness, and actionability of the feedback. Without such a comparison, the system could be reducing to a text-only LLM with a plausible dashboard.
[§2.3.2] Under 'Feedback Mechanism,' the paper states 'Our research demonstrates that this refined feedback mechanism not only provides musicians with practical, applicable advice but also fosters a more engaging and productive dialogue.' No data or procedure for this demonstration is presented. Please either provide the iterative-testing evidence, specify the evaluation protocol, or soften this claim to a design hypothesis.

minor comments (5)

[§1] The paper cites ComposerX [7] to support the claim that LLMs have 'inherent musical knowledge,' but ComposerX is a symbolic composition system, not a music-feedback or music-understanding benchmark. Consider citing evaluations of LLM musical reasoning or include a short validation of the LLM's production knowledge.
[§3.2] The statistic '75% of AI responses combined technical suggestions with emotional/perceptual feedback' is used without defining what counts as a response, how the units were segmented, or who coded the categories. Please provide the coding scheme and ideally inter-rater reliability.
[§2.2] The description of the iterative refinement process is underspecified: 'typically incorporate' and '2-3 iterations' should be replaced with the exact number of iterations, the set of statistical metrics added at each step, and the stopping criterion.
[§3.3] There are minor grammatical issues, e.g., 'For novice musicians in particular, they suggested the system might offer guidance'—the comma splice makes the subject unclear. Proofreading for such constructions would improve clarity.
[Figure 3] Figure 3 is described in the caption but no screenshot appears in the paper text. Including the actual UI screenshot would help readers assess the interaction design.

Circularity Check

1 steps flagged · score 3.0 of 10

Mild self-referential refinement loop in Section 2.2; otherwise the system build is independent and no fitted-input prediction.

self definitional [Section 2.2, 'LLM Readable Music Report']
"The process concludes with a secondary LLM that evaluates the output of each iteration for clarity, accuracy, and relevance, selecting the most insightful representation. This method demonstrably enhances the LLM’s capacity to generate meaningful, musician-relevant insights from complex musical data."

The claimed demonstration of enhancement rests on an LLM from the same model family selecting among outputs that the same family generated. If 'most insightful' is operationalized as the secondary LLM's preference, the selection maximizes that criterion by construction, so the claimed improvement over the raw report is self-confirming rather than externally measured. No human ground truth, independent scoring rubric, or comparison against a no-refinement baseline is given in Section 2.2 or in the pilot study, so the validation loop does not independently establish the enhancement.

full rationale

AI TrackMate is a system paper rather than a mathematical derivation, so most circularity patterns do not apply. The audio analysis tools (All-In-One, Madmom, autochord, AudioCommons, Essentia) are external and cited rather than fitted to the paper's own data, and the LLM feedback is generated, not statistically predicted from fitted parameters. There is also no self-citation chain: the cited LLM-music works, including ComposerX [7], have disjoint author sets from the present paper. The only genuine circularity is a validation loop in Section 2.2, where the iterative refinement is judged by a secondary LLM from the same model family, making the 'demonstrably enhances' claim self-referential. This does not compromise the core system construction, which remains an independent integration of existing components, so the score is moderate rather than high. The lack of accuracy checks on the MIR features and the single-participant pilot are empirical limitations, not circularity.

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

The central claim rests on the reliability of off-the-shelf MIR tools, the musical competence of LLMs, and prompt designs that are not formally evaluated. There are no fitted constants or new physical entities.

assumptions (4)

domain assumption LLMs possess sufficient inherent musical knowledge to interpret MIR metadata and produce useful production feedback.
Invoked in Section 2.2 and the Conclusion; cited to ComposerX [7] but not demonstrated for feedback quality.
domain assumption The audio analysis tools produce accurate measurements on arbitrary user tracks.
Section 2.1 lists All-In-One, Madmom, autochord, Essentia, and AudioCommons timbral models, but the paper reports no accuracy check on the demo tracks or user uploads.
domain assumption Graph-of-Thought prompting improves music reasoning specifically.
Used in Section 2.3.2; the original GoT paper is for general reasoning, and transfer to music production is assumed without testing.
ad hoc to paper The hand-designed scoring categories and prompt instructions (GoT, person switching, feedback mechanism) improve feedback quality over simpler prompts.
Section 2.3 introduces these design choices without ablations or comparisons.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AI TrackMate: Finally, Someone Who Will Give Your Music More Than Just "Sounds Great!"." pith.science (2026). https://pith.science/paper/UWJGNBWD

@misc{pith2026241206617,
  author       = {Pith},
  title        = {Pith review of: AI TrackMate: Finally, Someone Who Will Give Your Music More Than Just "Sounds Great!"},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UWJGNBWD}},
  note         = {Machine review of arXiv:2412.06617}
}

read the original abstract

The rise of "bedroom producers" has democratized music creation, while challenging producers to objectively evaluate their work. To address this, we present AI TrackMate, an LLM-based music chatbot designed to provide constructive feedback on music productions. By combining LLMs' inherent musical knowledge with direct audio track analysis, AI TrackMate offers production-specific insights, distinguishing it from text-only approaches. Our framework integrates a Music Analysis Module, an LLM-Readable Music Report, and Music Production-Oriented Feedback Instruction, creating a plug-and-play, training-free system compatible with various LLMs and adaptable to future advancements. We demonstrate AI TrackMate's capabilities through an interactive web interface and present findings from a pilot study with a music producer. By bridging AI capabilities with the needs of independent producers, AI TrackMate offers on-demand analytical feedback, potentially supporting the creative process and skill development in music production. This system addresses the growing demand for objective self-assessment tools in the evolving landscape of independent music production.

Figures

Figures reproduced from arXiv: 2412.06617 by the authors.

**Figure 1.** The system comprises three layers: (1) User Interface for audio upload, query input, and feedback reception; (2) Data Processing for handling raw audio and text; and (3) AI Analysis, featuring a Music Analysis Module that transforms raw audio into LLM-readable report, and an LLM that processes these reports along with user queries. Guided by music production-oriented feedback instructions, the LLM generates insights… view at source ↗

**Figure 2.** Graph of Thoughts (GoT) approach applies to analyze dominant instruments’ impact [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The user interface consists of: (1) Audio input components for file upload or YouTube link input. (2) A chat window displaying the LLM’s scoring and suggestions. (3) A text input for user questions and LLM responses. 3 Pilot Study We conducted an exploratory pilot study with a music producer to gather initial insights about our system’s approach and potential impact. The study consisted of a questionnaire, an onboar… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 9 canonical work pages

[1]

Gradio: Hassle-free sharing and testing of ml models in the wild

Abubakar Abid, Ali Abdalla, Ali Abid, Dawood Khan, Abdulrahman Alfozan, and James Zou. Gradio: Hassle-free sharing and testing of ml models in the wild. In Proceedings of the Machine Learning Conference , 2019. doi: 10.48550/arXiv.1906.02569. URL https: //arxiv.org/abs/1906.02569

arXiv doi:10.48550/arxiv.1906.02569 2019
[2]

Autochord: Automatic chord recognition library and chord visu- alization app

Christopher John Bayron. Autochord: Automatic chord recognition library and chord visu- alization app. In Proceedings of the International Society for Music Information Retrieval Conference (ISMIR), Manila, Philippines, 2021

work page 2021
[3]

Graph of thoughts: Solving elaborate problems with large language models

Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38(16), pages 17682–17690, 2024

work page 2024
[4]

mad- mom: a new Python Audio and Music Signal Processing Library

Sebastian Böck, Filip Korzeniowski, Jan Schlüter, Florian Krebs, and Gerhard Widmer. mad- mom: a new Python Audio and Music Signal Processing Library. In Proceedings of the 24th ACM International Conference on Multimedia, pages 1174–1178, Amsterdam, The Netherlands, 10 2016. doi: 10.1145/2964284.2973795

arXiv 2016
[5]

Mayor, Gerard Roma, Justin Salamon, J

Dmitry Bogdanov, Nicolas Wack, Emilia Gómez, Sankalp Gulati, Perfecto Herrera, O. Mayor, Gerard Roma, Justin Salamon, J. R. Zapata, and Xavier Serra. Essentia: an audio analysis library for music information retrieval. In Proceedings of the International Society for Music Information Retrieval Conference (ISMIR), pages 493–498, 2013

work page 2013
[6]

Music production & self-producing: A case study

Darro Chea. Music production & self-producing: A case study. Master’s thesis, Berklee College of Music, Valencia Campus, 2017

work page 2017
[7]

Composerx: Multi-agent symbolic music composition

Qixin Deng, Qikai Yang, Ruibin Yuan, Yipeng Huang, Yi Wang, Xubo Liu, Zeyue Tian, Jiahao Pan, Ge Zhang, Hanfeng Lin, Yizhi Li, Yinghao Ma, Jie Fu, Chenghua Lin, Emmanouil Benetos, Wenwu Wang, Guangyu Xia, Wei Xue, and Yike Guo. Composerx: Multi-agent symbolic music composition. arXiv:2404.18081, 2024

arXiv 2024
[8]

Lp-musiccaps: Llm-based pseudo music captioning

SeungHeon Doh, Keunwoo Choi, Jongpil Lee, and Juhan Nam. Lp-musiccaps: Llm-based pseudo music captioning. In Proceedings of the International Society for Music Information Retrieval Conference (ISMIR), 2023

work page 2023

Show all 14 references

[9]

Josh Gardner, Simon Durand, Daniel Stoller, and Rachel M. Bittner. Llark: A multimodal instruction-following language model for music. In Proceedings of the International Conference on Machine Learning, 2023

2023
[10]

All-in-one metrical and functional structure analysis with neigh- borhood attentions on demixed audio

Taejun Kim and Juhan Nam. All-in-one metrical and functional structure analysis with neigh- borhood attentions on demixed audio. In IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2023

2023
[11]

Any sound you can imagine? The bedroom producer, creativity and popular music education

Paul Thompson. Any sound you can imagine? The bedroom producer, creativity and popular music education. Journal of Popular Music Education, 2023

2023
[12]

Towards an understanding of creativity in independent music production

Daniel Walzer. Towards an understanding of creativity in independent music production. Creative Industries Journal, 2021

2021
[13]

Autogen: Enabling next-gen llm applications via multi-agent conversation

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation. arXiv preprint ...

2023 arXiv
[14]

Chatmusician: Understanding and generating music intrinsically with llm

Ruibin Yuan, Hanfeng Lin, Yi Wang, Zeyue Tian, Shangda Wu, Tianhao Shen, Ge Zhang, Yuhang Wu, Cong Liu, and Ziya Zhou. Chatmusician: Understanding and generating music intrinsically with llm. arXiv:2402.16153, 2024. 7

2024 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Gradio: Hassle-free sharing and testing of ml models in the wild

Abubakar Abid, Ali Abdalla, Ali Abid, Dawood Khan, Abdulrahman Alfozan, and James Zou. Gradio: Hassle-free sharing and testing of ml models in the wild. In Proceedings of the Machine Learning Conference , 2019. doi: 10.48550/arXiv.1906.02569. URL https: //arxiv.org/abs/1906.02569

arXiv doi:10.48550/arxiv.1906.02569 2019

[2] [2]

Autochord: Automatic chord recognition library and chord visu- alization app

Christopher John Bayron. Autochord: Automatic chord recognition library and chord visu- alization app. In Proceedings of the International Society for Music Information Retrieval Conference (ISMIR), Manila, Philippines, 2021

work page 2021

[3] [3]

Graph of thoughts: Solving elaborate problems with large language models

Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38(16), pages 17682–17690, 2024

work page 2024

[4] [4]

mad- mom: a new Python Audio and Music Signal Processing Library

Sebastian Böck, Filip Korzeniowski, Jan Schlüter, Florian Krebs, and Gerhard Widmer. mad- mom: a new Python Audio and Music Signal Processing Library. In Proceedings of the 24th ACM International Conference on Multimedia, pages 1174–1178, Amsterdam, The Netherlands, 10 2016. doi: 10.1145/2964284.2973795

arXiv 2016

[5] [5]

Mayor, Gerard Roma, Justin Salamon, J

Dmitry Bogdanov, Nicolas Wack, Emilia Gómez, Sankalp Gulati, Perfecto Herrera, O. Mayor, Gerard Roma, Justin Salamon, J. R. Zapata, and Xavier Serra. Essentia: an audio analysis library for music information retrieval. In Proceedings of the International Society for Music Information Retrieval Conference (ISMIR), pages 493–498, 2013

work page 2013

[6] [6]

Music production & self-producing: A case study

Darro Chea. Music production & self-producing: A case study. Master’s thesis, Berklee College of Music, Valencia Campus, 2017

work page 2017

[7] [7]

Composerx: Multi-agent symbolic music composition

Qixin Deng, Qikai Yang, Ruibin Yuan, Yipeng Huang, Yi Wang, Xubo Liu, Zeyue Tian, Jiahao Pan, Ge Zhang, Hanfeng Lin, Yizhi Li, Yinghao Ma, Jie Fu, Chenghua Lin, Emmanouil Benetos, Wenwu Wang, Guangyu Xia, Wei Xue, and Yike Guo. Composerx: Multi-agent symbolic music composition. arXiv:2404.18081, 2024

arXiv 2024

[8] [8]

Lp-musiccaps: Llm-based pseudo music captioning

SeungHeon Doh, Keunwoo Choi, Jongpil Lee, and Juhan Nam. Lp-musiccaps: Llm-based pseudo music captioning. In Proceedings of the International Society for Music Information Retrieval Conference (ISMIR), 2023

work page 2023