Pith. sign in

REVIEW 3 major objections 5 minor 27 references

Pretrained motion and music encoders, aligned by beat-guided contrastive learning and injected via ControlNet into a frozen text-to-audio diffusion model, produce better dance-aligned music from joint positions under scarce paired data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 10:58 UTC pith:YX23K5J5

load-bearing objection Solid systems recipe that combines known pieces for joint-to-music generation; the alignment edge over Li et al. is real but rests on a 36-clip test set without uncertainty. the 3 major comments →

arxiv 2607.10537 v1 pith:YX23K5J5 submitted 2026-07-12 cs.SD cs.MMeess.AS

Dance to Music Generation leveraging Pre-training with Unpaired data and Contrastive Alignment

Ryota Kimura , Sangheon Park , Natalia Polouliakh , Taketo Akama This is my paper
classification cs.SD cs.MMeess.AS
keywords dance-to-music generationcontrastive pretrainingControlNetjoint positionslatent diffusionbeat alignmentunpaired datamotion conditioning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Dance-to-music generation from human joint positions is attractive for choreography support and automatic accompaniment because joints are lightweight, privacy-preserving, and explicit about body dynamics. The obstacle is that accurately synchronized dance–music pairs are expensive and rights-constrained, so pure end-to-end training on the available pairs is weak. This paper claims the workable path is to freeze strong unimodal encoders trained on large unpaired motion and music corpora, align their feature spaces with a beat-guided contrastive stage on the few pairs that exist, and then steer a pretrained latent diffusion audio generator with a ControlNet-style residual branch that never overwrites the generative prior. On the standard AIST++ split the full pipeline raises both objective beat alignment with the input dance and listener ratings of dance match, while remaining competitive on objective audio-quality metrics. A sympathetic reader cares because the same two-stage pattern—unimodal pretraining, lightweight cross-modal alignment, frozen generator—offers a practical route for other motion-conditioned audio tasks that face the same data bottleneck.

Core claim

Combining frozen MotionBERT and MERT features, beat-guided contrastive pretraining that scores segments by maximum frame-wise similarity rather than average pooling, and a zero-initialized ControlNet adapter on a frozen AudioLDM backbone yields higher dance–music Beat Alignment Score and higher subjective dance-alignment ratings than a strong textual-inversion baseline, while remaining competitive on objective sound-quality metrics. Ablations attribute the perceptual alignment gain to both the contrastive stage and the pretrained motion encoder.

What carries the argument

Beat-guided contrastive pretraining plus ControlNet-style motion injection: dance and music features (concatenated with binary beat streams) are fused by dual-path transformers; segment similarity is the max frame-wise dot product; the resulting dance encoder is projected through zero-convolution ControlNet blocks that residual-modulate a frozen latent-diffusion U-Net without destroying its pretrained prior.

Load-bearing premise

The superiority claim rests on a 36-clip test partition and on treating beat-alignment-with-dance scores as more decisive than reference-music rhythmic scores, even though the compared systems use different audio generators.

What would settle it

Re-implement both conditioning methods on the identical frozen audio backbone, re-evaluate on a larger held-out dance set with the same Beat Alignment Score and dance-alignment listener protocol; if the proposed pipeline no longer leads on those two metrics, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Joint-position streams can condition existing text-to-audio diffusion systems without retraining the generator from scratch.
  • Scoring contrastive pairs by peak frame match rather than temporal average is especially useful when paired clips are short and scarce.
  • Unimodal motion pretraining supplies richer conditioning features than training a dance encoder only inside the diffusion loop.
  • Because many valid soundtracks can accompany one choreography, dance–music beat alignment and listener dance-match ratings should be treated as primary over reference-music beat scores.
  • The same two-stage recipe is portable to other scarce-pair cross-modal generation settings that already possess strong unimodal models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same ControlNet adapter could accept other sparse kinematic streams (hand tracking, wearable IMUs) without redesigning the audio backbone.
  • If kinematic beat extractors are noisy, max-frame similarity may lock onto spurious peaks; a soft temporal-attention alignment loss is a natural next test.
  • Scaling the contrastive stage to truly unpaired motion and music corpora (beyond the paired clips used here) could further reduce dependence on rights-constrained synchronized data.
  • Listener preference for a competing generator’s timbre despite worse objective fidelity suggests future work should swap backbones under identical conditioning to isolate alignment gains from generator quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a two-stage dance-to-music generation framework that uses 2D joint positions as motion input. Stage 1 freezes MotionBERT and MERT encoders, fuses them with kinematic/music beat features via a dual-path transformer, and aligns the modalities with a beat-guided contrastive objective that uses frame-wise maximum similarity rather than global average pooling. Stage 2 freezes a pretrained AudioLDM backbone and injects the aligned dance features through a ControlNet-style adapter with zero-initialized convolutions, optionally retaining text prompts. On the AIST++ split of Li et al. (2,744/36/36), the full model reports higher BAS (0.234 vs 0.194) and dance-alignment MOS (2.82 vs 2.66) than a MusicGen textual-inversion baseline, better objective audio metrics (FAD 4.56, CLAP 0.681), and ablations that attribute the perceptual alignment gain to both contrastive pretraining and MotionBERT. Code is released.

Significance. The work addresses a genuine data bottleneck in dance-conditioned music generation by combining unimodal pretraining, beat-aware contrastive alignment, and ControlNet-style adaptation of a frozen text-to-audio diffusion model. The integration is practically useful: joint-position conditioning is lightweight and privacy-preserving, and the ControlNet design preserves the generative prior under limited paired data. Public code and a clear two-stage recipe lower the barrier to reproduction. If the reported alignment gains hold under stronger statistical scrutiny and backbone-matched comparisons, the framework would be a solid incremental contribution for choreography support and automatic accompaniment systems.

major comments (3)
  1. Table 1 / §5.2: The central superiority claim (BAS 0.234 vs 0.194; dance-alignment MOS 2.82 vs 2.66) rests on point estimates from a 36-clip test partition with no standard errors, bootstrap intervals, or paired significance tests. The MOS study (35 listeners × 10 random clips) likewise omits inter-rater reliability and per-clip variance. Given the modest absolute deltas and the same tiny split used for all ablations and the baseline, sampling variability alone could reverse the ranking. At minimum the paper should report uncertainty (e.g., bootstrap CIs over the 36 clips and listener-level MOS variance) or a paired test; without that the claim of superior dance-conditioned generation is under-supported.
  2. §5.2 and Table 1: The comparison with Li et al. confounds the proposed conditioning mechanism with the generative backbone (AudioLDM vs MusicGen). The paper correctly notes that BHS/F1/TD favor the baseline and that subjective sound-quality MOS is slightly lower, yet still claims superior dance alignment. Because the ablation against AudioLDM (default) already isolates the conditioning contribution, the cross-backbone ranking should either be de-emphasized or supplemented by a same-backbone re-implementation (or an explicit statement that the claim is limited to the AudioLDM family). As written, the headline comparison overstates what the evidence isolates.
  3. §5.1 ablation “w/o Contrastive Pretraining”: This variant scores higher or comparable on most objective metrics (BCS, F1, BAS, CLAP, FAD) yet drops sharply in dance-alignment MOS (2.82 → 2.39). The paper interprets this as evidence that contrastive pretraining yields “richer” perceptual correspondence. That dissociation is interesting but currently post-hoc; without additional analysis (e.g., qualitative listening examples, beat-level error distributions, or a controlled probe of what the contrastive features capture beyond BAS) the claim that contrastive pretraining is load-bearing for alignment remains only partially substantiated by the numbers in Table 1.
minor comments (5)
  1. §3.2.1 Eqs. (1)–(2): The fusion notation mixes element-wise sum/product with concatenation; a short clarifying sentence or diagram inset would help readers reconstruct the dual-path transformer input dimensions.
  2. §4.1.2: The text-prompt construction (top-1 genre / top-3 instruments / top-1 mood from MTG-Jamendo via CLAP) is reasonable but never ablated; a one-sentence note on whether text is essential or merely auxiliary would strengthen the experimental design.
  3. Table 1 header: “vs. Ground-truth Music” and “vs. Ground-truth Dance” column groups are slightly ambiguous for the MOS columns; a footnote clarifying that dance-alignment MOS is rated against the input dance (not the reference music) would avoid misreading.
  4. §3.1.1: The kinematic beat extraction (Gaussian-smoothed velocity minima) is standard but the exact filter width and local-minima threshold are not stated; adding them would aid exact reproduction.
  5. Abstract and §1: “unpaired data” is emphasized, yet the contrastive stage still uses paired AIST++ clips; a clearer distinction between unimodal pretraining (MotionBERT/MERT) and the paired contrastive stage would avoid overstating the unpaired contribution.

Circularity Check

0 steps flagged

Empirical ML paper with no circular derivation: claimed gains are measured test-set metrics, not quantities forced by construction from fitted inputs or self-citation.

full rationale

The paper proposes a two-stage engineering pipeline (unimodal pretrained encoders + beat-guided contrastive alignment of motion/music features + ControlNet adapter on frozen AudioLDM) and evaluates it on the AIST++ split with standard metrics (BAS, BHS/BCS/F1/TD, FAD, CLAP, MOS). No equation or claim reduces a reported “prediction” to a parameter fitted on the same quantity: contrastive pretraining (Eqs. 1–3, L_NCE) and the diffusion objective (Eq. 8) are ordinary training losses; the reported BAS/MOS edges vs. Li et al. and ablations are post-hoc measurements on held-out clips, not algebraic identities. Citations (MotionBERT, MERT, AudioLDM, ControlNet, BeatDance, Li et al.) supply reusable components or baselines; none is a uniqueness theorem by overlapping authors that forbids alternatives or smuggles the target result. Preference for BAS/dance-alignment MOS over reference-music rhythmic scores is an explicit task-definition argument (§5.2), not a self-referential redefinition. The work is therefore self-contained against external benchmarks; circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

As an empirical systems paper the load-bearing content is architectural choices and hyper-parameters rather than physical axioms. The free parameters are the usual training knobs; the axioms are standard ML and music-domain assumptions imported from cited work; the invented entities are the specific fusion and similarity constructions introduced for this pipeline.

free parameters (4)
  • contrastive learning rate / epochs / batch / weight decay / embed dim = 1e-6 / 150 / 32 / 0.2 / 256
    AdamW 1e-6, 150 epochs, batch 32, weight decay 0.2, shared dim 256; chosen by hand and affect the quality of the aligned space that later conditions generation.
  • diffusion training steps / lr / batch / guidance / DDIM steps = 300k / 1e-4 / 8 / 3.5 / 200
    300 k steps, Adam 1e-4, batch 8, guidance 3.5, 200 DDIM steps; control final audio quality and alignment strength.
  • time-stretch range and probability, text-dropout probability = [0.8,1.0]@70% / 50%
    Stretch [0.8,1.0] at 70 %, text drop 50 %; data-augmentation and classifier-free choices that influence robustness and controllability.
  • temporal length L and beat reshape factors D_BD, D_BM = L=128
    L=128 frames and integer beat down-sampling factors determine the resolution at which motion and music are aligned and injected.
axioms (5)
  • domain assumption Pretrained MotionBERT and MERT embeddings already encode useful unimodal structure for dance and music respectively.
    Invoked in §3.1; the entire pipeline freezes these extractors and never re-trains them from scratch.
  • domain assumption Kinematic and acoustic beats are reliable rhythmic anchors that improve cross-modal alignment when concatenated and attended.
    Taken from BeatDance [18] and used throughout §3.1–3.2; local-minima velocity beats and Librosa beats are treated as ground-truth rhythm.
  • domain assumption Zero-initialized ControlNet convolutions allow stable adaptation of a frozen diffusion U-Net under limited paired data without destroying the generative prior.
    Cited from ControlNet [23] and applied in §3.3.1 equations (5)–(7).
  • ad hoc to paper The AIST++ train/val/test split of Li et al. is a fair and representative evaluation protocol for dance-to-music generation.
    Adopted in §4.1.1 solely for comparability; the 36-clip test set is never shown to be statistically stable.
  • ad hoc to paper Frame-wise maximum cosine similarity (rather than global average pooling) better captures salient motion–music alignments for contrastive loss.
    Introduced in §3.2.2 equation (3) as a deliberate departure from BeatDance; no independent theoretical justification is given.
invented entities (3)
  • Beat-guided dual-path transformer fusion with explicit positional encoding on the beat stream no independent evidence
    purpose: Produce aligned dance and music embeddings ˆf_D, ˆf_M that serve as the conditioning interface.
    Equations (1)–(2) and the PE augmentation are specific constructions of this paper; independent evidence is limited to the ablation on AIST++.
  • Temporal-max frame-wise similarity matrix for InfoNCE no independent evidence
    purpose: Drive contrastive pretraining by the single most salient frame pair instead of pooled features.
    Equation (3); claimed to avoid rhythmic dilution on short clips, but validated only inside this pipeline.
  • Two-stage dance-to-music ControlNet pipeline on AudioLDM no independent evidence
    purpose: Inject contrastively aligned motion features into a frozen text-to-audio diffusion model while preserving its generative prior.
    Overall architecture of §3; the combination itself is the paper’s main proposed entity.

pith-pipeline@v1.1.0-grok45 · 16169 in / 3614 out tokens · 48631 ms · 2026-07-14T10:58:17.503236+00:00 · methodology

0 comments
read the original abstract

Dance-to-music generation is a promising task for applications such as choreography support and automatic accompaniment, where temporal coordination between body movement and sound is essential. In particular, using human joint positions as the motion representation is attractive because they explicitly capture body dynamics while being lightweight, privacy-preserving, and easy to integrate with motion capture and pose-estimation pipelines. A central challenge in this setting, however, is the scarcity of high-quality paired dance-music data, since collecting accurately synchronized pairs is costly and often constrained by copyright and performance rights. This makes it difficult to train end-to-end models solely from paired data. To address this issue, we propose a dance-conditioned music generation framework that efficiently exploits both unpaired and paired data. Our method combines pretrained unimodal encoders for motion and music, beat-guided contrastive pretraining to align their feature spaces, and a ControlNet-style conditioning module on top of a pretrained text-to-audio diffusion model. Experiments on AIST++ demonstrate that the proposed techniques improve both dance-music alignment and audio quality, as confirmed by quantitative and qualitative evaluations. Compared to a state-of-the-art method, our approach achieves superior dance alignment performance and competitive audio quality. Code is available at https://github.com/kmraven/AudioLDM-ControlNet .

Figures

Figures reproduced from arXiv: 2607.10537 by Natalia Polouliakh, Ryota Kimura, Sangheon Park, Taketo Akama.

Figure 1
Figure 1. Figure 1: Overview of the proposed two-stage dance-to-music generation framework. In the first stage, dance and music features are extracted using pretrained encoders with beat information and aligned through beat-guided contrastive learning. In the second stage, the pretrained dance encoder provides motion features that are projected and injected into a ControlNet-style adapter to condition a frozen AudioLDM backbo… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 1 canonical work pages

  1. [1]

    Dance- to-music generation with encoder-based textual inversion,

    S. Li, W. Dong, Y. Zhang, F. Tang, C. Ma, O. Deussen, T.-Y. Lee, and C. Xu, “Dance- to-music generation with encoder-based textual inversion,” inSIGGRAPH Asia 2024 Conference Papers. ACM, 2024, pp. 135:1–135:11. [Online]. Available: https://arxiv.org/abs/2401.17800

    Pith/arXiv arXiv 2024

  2. [2]

    Momu-diffusion: On learning long-term motion-music synchro- nization and correspondence,

    F. You, M. Fang, L. Tang, R. Huang, Y. Wang, and Z. Zhao, “Momu-diffusion: On learning long-term motion-music synchro- nization and correspondence,”arXiv preprint arXiv:2411.01805, 2024. [Online]. Available: https://arxiv.org/abs/2411.01805

    Pith/arXiv arXiv 2024

  3. [3]

    Dance2music: Au- tomatic dance-driven music generation,

    G. Aggarwal and D. Parikh, “Dance2music: Au- tomatic dance-driven music generation,”arXiv preprint arXiv:2107.06252, 2021

    Pith/arXiv arXiv 2021

  4. [4]

    Quantized GAN for complex music generation from dance videos,

    Y. Zhu, K. Olszewski, Y. Wu, P. Achlioptas, M. Chai, Y. Yan, and S. Tulyakov, “Quantized GAN for complex music generation from dance videos,” inComputer Vision – ECCV 2022. Springer, 2022, pp. 182–199. [Online]. Available: https://arxiv.org/abs/2204.00604

    Pith/arXiv arXiv 2022

  5. [5]

    Dance2midi: Dance-driven multi-instrument mu- sic generation,

    B. Han, Y. Li, Y. Shen, Y. Ren, and F. Han, “Dance2midi: Dance-driven multi-instrument mu- sic generation,”Computational Visual Media, vol. 10, no. 4, pp. 791–802, 2024

    2024

  6. [6]

    Motion to dance music generation using latent diffusion model,

    V. Tan, J. Nam, J. Nam, and J. Noh, “Motion to dance music generation using latent diffusion model,” inSIGGRAPH Asia 2023 Technical Communications. ACM, 2023, pp. 1–4. [Online]. Available: https://dmdproject.github.io/

    2023

  7. [7]

    Dance2music-diffusion: Leveraging latent diffusion models for music generation from dance videos,

    C. Zhang and Y. Hua, “Dance2music-diffusion: Leveraging latent diffusion models for music generation from dance videos,”EURASIP Journal on Audio, Speech, and Music Processing, vol. 2024, p. 48, 2024. [Online]. Available: https: //doi.org/10.1186/s13636-024-00370-6

    work page doi:10.1186/s13636-024-00370-6 2024

  8. [8]

    Ai choreographer: Music conditioned 3d dance generation with aist++,

    R. Li, S. Yang, D. A. Ross, and A. Kanazawa, “Ai choreographer: Music conditioned 3d dance generation with aist++,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 13401–13412

    2021

  9. [9]

    Jukebox: A generative model for music,

    P. Dhariwal, H. Jun, C. Payne, J. W. Kim, A. Radford, and I. Sutskever, “Jukebox: A generative model for music,”arXiv preprint arXiv:2005.00341, 2020. [Online]. Available: https: //arxiv.org/abs/2005.00341

    Pith/arXiv arXiv 2005

  10. [10]

    Mu- siclm: Generating music from text,

    A. Agostinelli, T. I. Denk, Z. Borsos, J. En- gel, M. Verzetti, A. Caillon, Q. Huang, A. Jansen, A. Roberts, M. Tagliasacchi, M. Sharifi, N. Zeghidour, and C. Frank, “Mu- siclm: Generating music from text,”arXiv preprint arXiv:2301.11325, 2023. [Online]. Avail- able: https://arxiv.org/abs/2301.11325

    Pith/arXiv arXiv 2023

  11. [11]

    Simple and controllable music gen- eration,

    J. Copet, F. Kreuk, I. Gat, T. Remez, D. Kant, G. Synnaeve, Y. Adi, and A. Dé- fossez, “Simple and controllable music gen- eration,” inAdvances in Neural Information Processing Systems, 2023. [Online]. Available: https://openreview.net/forum?id=jtiQ26sCJi

    2023

  12. [12]

    Noise2music: Text-conditioned music generation with diffusion models,

    Q. Huang, D. S. Park, T. Wang, T. I. Denk, A. Ly, N. Chen, Z. Zhang, Z. Zhang, J. Yu, C. Frank, J. Engel, Q. V. Le, W. Chan, Z. Chen, and W. Han, “Noise2music: Text-conditioned music generation with diffusion models,”arXiv preprint arXiv:2302.03917, 2023. [Online]. Available: https: //arxiv.org/abs/2302.03917

    Pith/arXiv arXiv 2023

  13. [13]

    Moûsai: Text-to-music genera- tion with long-context latent diffusion,

    F. Schneider, O. Kamal, Z. Jin, and B. Schölkopf, “Moûsai: Text-to-music genera- tion with long-context latent diffusion,”arXiv preprint arXiv:2301.11757, 2023. [Online]. Avail- able: https://arxiv.org/abs/2301.11757 7

    Pith/arXiv arXiv 2023

  14. [14]

    AudioLDM: Text-to-audio generation with la- tent diffusion models,

    H. Liu, Z. Chen, Y. Yuan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. D. Plumbley, “AudioLDM: Text-to-audio generation with la- tent diffusion models,” inProceedings of the 40th International Conference on Machine Learn- ing. PMLR, 2023, pp. 21450–21474. [Online]. Available: https://arxiv.org/abs/2301.12503

    Pith/arXiv arXiv 2023

  15. [15]

    Music controlnet: Multiple time-varying controls for music generation,

    S.-L. Wu, C. Donahue, S. Watanabe, and N. J. Bryan, “Music controlnet: Multiple time-varying controls for music generation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 2692–2703,

  16. [16]

    Available: https://doi.org/10

    [Online]. Available: https://doi.org/10. 1109/TASLP.2024.3399026

    arXiv 2024

  17. [17]

    Learning transferable visual mod- els from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual mod- els from natural language supervision,” in Proceedings of the 38th International Con- ference on Machine Learning. PMLR, 2021, pp. 8748–8763. [Online]. Available: https: //proceeding...

    2021

  18. [18]

    Hierarchical text-conditional image generation with CLIP latents,

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchical text-conditional image generation with CLIP latents,”arXiv preprint arXiv:2204.06125, 2022. [Online]. Available: https: //arxiv.org/abs/2204.06125

    Pith/arXiv arXiv 2022

  19. [19]

    Beatdance: A beat-based model-agnostic contrastive learning framework for music-dance retrieval,

    K. Yang, X. Zhou, X. Tang, R. Diao, H. Liu, J. He, and Z. Fan, “Beatdance: A beat-based model-agnostic contrastive learning framework for music-dance retrieval,” inProceedings of the 2024 International Conference on Multimedia Retrieval. ACM, 2024, pp. 11–19. [Online]. Available: https://arxiv.org/abs/2310.10300

    Pith/arXiv arXiv 2024

  20. [20]

    Motionbert: A unified perspective on learning human motion representations,

    W. Zhu, X. Ma, Z. Liu, L. Liu, W. Wu, and Y. Wang, “Motionbert: A unified perspective on learning human motion representations,” inPro- ceedings of the IEEE/CVF international confer- ence on computer vision, 2023, pp. 15085–15099

    2023

  21. [21]

    Mert: Acoustic music understanding model with large-scaleself-supervisedtraining,

    Y. Li, R. Yuan, G. Zhang, Y. Ma, X. Chen, H. Yin, C. Xiao, C. Lin, A. Ragni, E. Benetoset al., “Mert: Acoustic music understanding model with large-scaleself-supervisedtraining,”arXiv preprint arXiv:2306.00107, 2023

    Pith/arXiv arXiv 2023

  22. [22]

    li- brosa: Audio and music signal analysis in python

    B. McFee, C. Raffel, D. Liang, D. P. Ellis, M. McVicar, E. Battenberg, O. Nietoet al., “li- brosa: Audio and music signal analysis in python.” SciPy, vol. 2015, no. 18-24, p. 7, 2015

    2015

  23. [23]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polo- sukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  24. [24]

    Adding con- ditional control to text-to-image diffusion mod- els,

    L. Zhang, A. Rao, and M. Agrawala, “Adding con- ditional control to text-to-image diffusion mod- els,” inProceedings of the IEEE/CVF interna- tional conference on computer vision, 2023, pp. 3836–3847

    2023

  25. [25]

    The mtg-jamendo dataset for automatic music tagging,

    D. Bogdanov, M. Won, P. Tovstogan, A. Porter, and X. Serra, “The mtg-jamendo dataset for automatic music tagging,” inMachine Learning for Music Discovery Workshop, International Conference on Machine Learning (ICML 2019), Long Beach, CA, United States, 2019. [Online]. Available: http://hdl.handle.net/10230/42015

    2019

  26. [26]

    Clap learning audio concepts from natural language supervision,

    B. Elizalde, S. Deshmukh, M. Al Ismail, and H. Wang, “Clap learning audio concepts from natural language supervision,” inICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  27. [27]

    Fr\’echet audio distance: A metric for evaluating music enhancement algorithms,

    K. Kilgour, M. Zuluaga, D. Roblek, and M. Sharifi, “Fr\’echet audio distance: A metric for evaluating music enhancement algorithms,”arXiv preprint arXiv:1812.08466, 2018. 8

    Pith/arXiv arXiv 2018