Pith. sign in

REVIEW 4 major objections 4 minor 33 references

Training a Perceptual Model for Evaluating Auditory Similarity in Music Adversarial Attack

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

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

Pith's one-line read This paper claims that conditioning a small transformer projection on psychoacoustic perturbation parameters yields music embeddings that match human similarity judgments (Spearman 0.65) and make adversarial training against perceptual atta

desk verdict Interesting idea and a useful human dataset, but the paper underspecifies test-time conditioning and uses a self-referential robustness evaluation, so the headline numbers are not yet convincing. read the letter →

arxiv 2509.04985 v1 pith:HVFJBUIS submitted 2025-09-05 cs.SD eess.AS

Yuxuan Liu , Rui Sang , Peihong Zhang , Zhixin Li , Shengchen Li This is my paper · ORCID
classification cs.SDeess.AS
keywords musicinformationretrievaladversarialattacksperceptualaudiosimilaritypsychoacousticconditioningcontrastivelearningtrainingcoversongidentificationgenreclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that adversarial attacks on music AI succeed because model feature spaces do not track human hearing, and that a lightweight trainable projection head can close that gap. Its listening tests show that standard metrics such as SNR, LSD, PEMO-Q, and Fréchet Audio Distance correlate only weakly with human similarity judgments on adversarial perturbations. The proposed PAMT framework conditions a transformer projection on psychoacoustic perturbation parameters and trains it against pairwise human preferences, reaching a Spearman correlation of 0.65 with subjective scores. When that embedding distance is used as an adversarial training constraint, it raises worst-case robust accuracy on cover-song identification and music genre classification by an average of 9.15% while keeping clean accuracy high. If correct, the result supplies both a perceptual metric for judging music adversarial attacks and a feature space in which defenses are more aligned with human perception.

What carries the argument

The core object is the Psychoacoustically-Conditioned Sequential Contrastive Transformer (PCSCT): a four-layer Transformer projection head that maps 768-dimensional frozen MERT embeddings into a 128-dimensional perceptual space. A Perturbation Parameter Encoder converts known psychoacoustic perturbation parameters into a 64-dimensional conditioning vector, which FiLM layers inject into every transformer layer, letting the model adapt its notion of invariance to the perturbation profile. A sequential contrastive loss then makes the mean-pooled PAMT embeddings close for original-perturbed pairs and far for unrelated clips, so that the L2 distance in this space can serve as a differentiable per

What would settle it

Re-run the perceptual correlation test on the same held-out pairs but withhold the perturbation parameters from PAMT, or feed mismatched parameters. If the Spearman correlation with human 2AFC scores drops from 0.65 toward the 0.2–0.3 range seen for signal-level metrics, the reported alignment depends on knowing the attack in advance.

Watch

Extended reading notes

Core claim

The central claim is that PAMT, built from a frozen MERT encoder and a Psychoacoustically-Conditioned Sequential Contrastive Transformer (PCSCT) projection head, learns music representations whose distances track human auditory similarity. PCSCT is conditioned on a 64-dimensional vector encoding the perturbation's psychoacoustic parameters, injected through feature-wise linear modulation, and trained with a sequential InfoNCE-style contrastive objective to pull original and perturbed audio together while separating different clips. The resulting mean-pooled embeddings define the perceptual distance used to constrain adversarial examples, and the paper reports that this space outperforms exis

Load-bearing premise

The load-bearing premise is that every time PAMT compares two clips, it knows the exact perturbation parameters that turned one into the other; if that information is unavailable, the conditioning input that makes the embeddings perceptual has no defined value.

Editorial extensions

If this is right

  • PAMT's embedding distance can be used as a differentiable perceptual constraint for crafting adversarial examples, not only for defending against them, because it tracks human similarity rankings.
  • Adversarial training inside the PAMT embedding space raises worst-case robust accuracy on Cover Song Identification from 0.315 to 0.535 mAP and on Music Genre Classification from 0.330 to 0.465 accuracy, with only a small drop on clean data.
  • The approach learns perceptual alignment without retraining the large frozen encoder, so new attack types or corpora could be handled by retraining only the lightweight projection head.
  • The high F1 score at distinguishing noticeable from imperceptible perturbations suggests PAMT embeddings could support filtering or re-ranking of adversarial examples at deployment time.
  • The finding that FAD with pre-trained music embeddings already correlates better than signal-level metrics supports the broader direction of using self-supervised music representations as the basis for perceptual evaluation.

Reading between the lines

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

  • If PAMT is to serve as a general perceptual metric on arbitrary music pairs, the perturbation-parameter conditioning vector must be estimated or inferred from the audio itself; the paper's training pipeline always supplies it from a known perturbation module, so unknown-attack settings are a natural next test.
  • The same frozen-encoder-plus-conditioned-head recipe could transfer to other large self-supervised audio encoders, since only the small projection head is trained, though the paper only demonstrates it with one encoder.
  • The finding that 2AFC-derived scores give more reliable supervision than MOS-style ratings suggests future perceptual audio metrics could train directly on pairwise preference data, even when collecting such data is more expensive.
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.

Referee Report

4 major / 4 minor

Summary. The paper proposes PAMT, a projection head built on a frozen MERT encoder, conditioned on psychoacoustic perturbation parameters via FiLM and trained with a sequential contrastive loss. The authors report a Spearman correlation of 0.65 with human 2AFC similarity judgments, outperforming FAD and a MERT+MLP contrastive baseline, and report improved adversarial robustness in Cover Song Identification and Music Genre Classification when models are adversarially trained in PAMT space. The paper is motivated by listening tests showing low correlation of existing metrics with human perception.

Significance. If substantiated, the work would be a useful contribution to perceptually aligned audio similarity metrics and adversarial defenses for MIR. The authors contribute a substantial listening-test dataset (18,000 rated pairs, 45,000 2AFC trials) and compare against a broad set of baseline metrics. The architectural idea of conditioning a contrastive projection head on perturbation parameters is interesting. However, the headline claims rest on an incompletely specified inference-time conditioning mechanism and on a robustness evaluation that may be circular; these issues must be resolved before the contribution can be assessed.

major comments (4)
  1. [Sec. 3.1–3.2, Eq. (1)] The conditioning vector c_perturb is produced by the PPE from perturbation parameters that are available during training, but the paper never specifies how c_perturb is obtained at inference when comparing two arbitrary audio clips. Eq. (3) defines d_PAMT(A',A) using Z_PAMT(A') and Z_PAMT(A), both of which depend on c_perturb via the FiLM layers. Without a stated protocol for selecting c_perturb at test time (e.g., a default vector, an estimated vector, or oracle metadata), Table 2's reported ρ=0.65 and the adversarial-training constraint in Eq. (4) are not reproducible. The authors must specify the inference-time conditioning and ablate the choice, because the claimed improvement over MERT+MLP (0.55 vs. 0.65) may rely on oracle perturbation information.
  2. [Sec. 3.3, Eq. (4); Table 3] The robustness evaluation appears circular. Adversarial examples in Eq. (4) are constrained by d_PAMT(A',A) ≤ ε, while the defense trains fθ on mean_pool(Z_PAMT(A')). Since PAMT is explicitly trained to make Z_PAMT invariant to perceptually small perturbations, the defender is being evaluated against attacks that are constrained by the very distance the model was trained to minimize. The 'Union Robust Acc.' in Table 3 is therefore partly forced by construction. To support the claim that PAMT gives stronger defenses, the authors should include attacks that are not defined via d_PAMT (e.g., standard L∞/L2 PGD, or the perceptual attacks of Section 2 applied uniformly to all defenses) and should report the attack protocol (method, steps, budget ε, white-box/black-box).
  3. [Sec. 2.2; Table 2] The listening-test section reports no inter-rater reliability or agreement statistics. Each of the 18,000 pairs is rated by 'at least five' volunteers, and the 2AFC scores are aggregated over 45,000 trials, but no Krippendorff's alpha, ICC, or similar measure is given. With only five ratings per pair, label noise could substantially attenuate the reported Spearman correlations, making it difficult to know whether PAMT's 0.65 truly reflects perceptual alignment or partly reflects noise in the labels. Please report agreement statistics and confidence intervals for the correlations.
  4. [Sec. 4.5, Table 3] The table reports only 'Clean' and 'Union Robust Acc.' values. The set of attacks used to compute the worst-case performance is not described: which attack algorithms, what perturbation budgets, how the attacks are initialized, and whether they are optimized against each defense or transferred from a single surrogate are all unspecified. Without these details and a common evaluation protocol applied to all defenses, the relative gains of PAMT AT (0.535 mAP / 0.465 Acc) over the baselines cannot be meaningfully interpreted.
minor comments (4)
  1. [Sec. 4.4] The F1 score is described as measuring the model's utility in distinguishing 'noticeable from imperceptible perturbations,' but the definition of the binary classes and the threshold used to derive F1 are not given. Please specify the evaluation protocol.
  2. [Sec. 3.2] The citation for FiLM and Pre-LN is given as [26], but reference [26] is a paper on zero-shot musical stem retrieval, not the original FiLM paper (Perez et al.) or the Pre-LN transformer reference. Please correct the citation.
  3. [Eq. (2)] The notation in the denominator is confusing: 'Batch' appears as a set in the sum, and the term exp(sim(Z_orig, Z_pert)/τ) appears both in the numerator and in the denominator, which is redundant. Please clarify.
  4. [Sec. 2.2] The rating scale is described as '1: identical, 5: dissimilar,' which is the opposite direction of a conventional MOS scale. This is fine, but it should be explicitly noted when interpreting the sign of Spearman correlations.

Circularity Check

2 steps flagged · score 6.0 of 10

The PAMT distance used to constrain adversarial attacks is the same distance the PCSCT is trained to minimize, and the metric depends on a conditioning vector c_perturb that is only defined during training; the headline correlation and robust-accuracy gains are therefore partly forced by construction or rely on unavailable oracle inputs.

  1. self definitional [Section 3.3 Eq. (4); Section 3.1 item 5; Eq. (2)]
    "Adversarial examples A' are crafted to maximize the task loss while being constrained by d_PAMT(A', A)≤ε. ... The PCSCT is trained to minimize the distance between representations of original and perturbed audio pairs (Z_orig_PAMT, Z_pert_PAMT) while maximizing their distance to other samples in the batch."

    The attack constraint in Eq. (4) is exactly the L2 distance between mean-pooled PAMT embeddings (Eq. 3), which the PCSCT training objective (Eq. 2) explicitly minimizes for the same perturbation types used in the attack suite. Thus, an adversary constrained by d_PAMT(A',A)≤ε is limited to inputs that the model was trained to map to the original representation. The reported Union Robust Acc. is therefore a measure of how well PAMT satisfied its own invariance objective, not an independent test of robustness. The comparison against Standard AT (Lp) and MERT+MLP AT uses different constraint metrics, so the 9.15% improvement may be an artifact of the evaluation protocol rather than evidence of superior perceptual alignment.

  2. other [Section 3.1 items 2-4; Eq. (1); Eq. (3)]
    "Psychoacoustic Perturbation Module: During training, this module creates a perturbed audio version A_pert ... and outputs the perturbation parameters P_params. ... Perturbation Parameter Encoder (PPE): A 2-layer MLP encodes P_params into a compact 64-dimensional conditioning vector c_perturb."

    Eq. (3) defines the perceptual distance using Z_PAMT, which is the output of the PCSCT conditioned on c_perturb via FiLM (Eq. 1). The paper states the perturbation module runs 'during training' only and never specifies how c_perturb is obtained for arbitrary test pairs. If the test-time c_perturb is the true perturbation parameter, the metric has oracle access to the perturbation type and strength that human raters judge, so the Spearman 0.65 does not establish alignment from audio alone. If a default c_perturb is used, the conditioning innovation is not actually exercised and the advantage over MERT+MLP is unexplained. In either case, the claimed perceptual metric is not computable as a general auditory-similarity metric without an undefined privileged input.

full rationale

The paper's core innovation is a conditioned projection head, but the conditioning vector appears only during training. The distance used to define the perceptual metric and to constrain attacks is the same distance minimized in the contrastive objective. This creates a partial self-definitional loop in the defense evaluation: the attack is constrained by the very metric the model was trained to be invariant to, so the robust-accuracy improvement is expected and not independently validated. In addition, the test-time value of c_perturb is missing, which leaves the headline 0.65 correlation potentially dependent on oracle perturbation metadata. These issues are substantial, but they are not a complete logical tautology: the Spearman correlation is still an empirical result and the contrastive training could fail. The paper does not rely on self-citation. Score 6 reflects that the central claims are partly forced by construction and by an undefined test-time input, though independent empirical content remains.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on hand-chosen perturbation ranges, an InfoNCE temperature, an unspecified inference-time conditioning, and an unvalidated human gold standard. No new physical entities are introduced.

free parameters (3)
  • InfoNCE temperature tau = 0.1
    Chosen by hand with no sensitivity analysis; affects the contrastive objective in Eq. (2).
  • Perturbation parameter ranges = Table 1 ranges
    The six perturbation types and their parameter ranges define the training distribution of PAMT and the human listening test; they are hand-selected to simulate adversarial attacks, and the model's invariance is only demonstrated on these ranges.
  • Distance threshold epsilon in adversarial training = not specified
    Eq. (4) bounds adversarial examples by d_PAMT <= epsilon but epsilon is never given; robust accuracy depends on it.
assumptions (5)
  • domain assumption The six synthetic perturbation types in Table 1 adequately represent the space of imperceptible adversarial attacks on music.
    Used to build the human gold standard and to train PAMT; if real attacks live outside this space, the claimed perceptual alignment and robustness do not transfer.
  • ad hoc to paper At inference, a conditioning vector c_perturb is available to the PCSCT, even though the paper never specifies how it is obtained for arbitrary pairs (Eq. 3).
    The defined distance d_PAMT depends on Z_PAMT which is a function of c_perturb; the paper omits this dependence.
  • domain assumption Human similarity ratings from 200 volunteers, each pair rated at least 5 times, are a reliable gold standard; no inter-rater agreement is reported.
    The target labels for both the correlation evaluation and the training objective are these noisy human scores.
  • domain assumption Frozen MERT-v0 features contain sufficient information about music to support perceptual alignment.
    Everything is built on top of MERT; if MERT misses relevant auditory cues, PAMT cannot recover them.
  • domain assumption A mean-pooled cosine similarity or L2 distance in the PAMT space corresponds to human similarity.
    The model uses time-pooled mean vectors for both contrastive loss and distance; no evidence that this pooling preserves perceptual order.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Training a Perceptual Model for Evaluating Auditory Similarity in Music Adversarial Attack." pith.science (2026). https://pith.science/paper/HVFJBUIS

@misc{pith2026250904985,
  author       = {Pith},
  title        = {Pith review of: Training a Perceptual Model for Evaluating Auditory Similarity in Music Adversarial Attack},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HVFJBUIS}},
  note         = {Machine review of arXiv:2509.04985}
}
read the original abstract

Music Information Retrieval (MIR) systems are highly vulnerable to adversarial attacks that are often imperceptible to humans, primarily due to a misalignment between model feature spaces and human auditory perception. Existing defenses and perceptual metrics frequently fail to adequately capture these auditory nuances, a limitation supported by our initial listening tests showing low correlation between common metrics and human judgments. To bridge this gap, we introduce Perceptually-Aligned MERT Transformer (PAMT), a novel framework for learning robust, perceptually-aligned music representations. Our core innovation lies in the psychoacoustically-conditioned sequential contrastive transformer, a lightweight projection head built atop a frozen MERT encoder. PAMT achieves a Spearman correlation coefficient of 0.65 with subjective scores, outperforming existing perceptual metrics. Our approach also achieves an average of 9.15\% improvement in robust accuracy on challenging MIR tasks, including Cover Song Identification and Music Genre Classification, under diverse perceptual adversarial attacks. This work pioneers architecturally-integrated psychoacoustic conditioning, yielding representations significantly more aligned with human perception and robust against music adversarial attacks.

Figures

Figures reproduced from arXiv: 2509.04985 by the authors.

Figure 1
Figure 1. Spearman Correlation Coefficients (ρ) between objective auditory metrics and human perceptual similarity ratings (mean ± std. dev.). Correlations are shown for both raw MOS-style scores and 2AFC-derived scores. Objective metrics include SNR, LSD, PEMO-Q, FAD with various embeddings (VGGish [11], PANNS-CNN14- 16k [16], PANNS-Wavegram-Logmel [16], MERT-v0 [17]), and speech-domain models (CDPAM [18], NOMAD [25], SESQA … view at source ↗
Figure 2
Figure 2. Architectural Overview of the Proposed Perceptually-Aligned MERT [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 30 canonical work pages

  1. [1]

    In: Asian Conference on Machine Learning

    Abdullah, H., Rahman, M.S., Peeters, C., Gibson, C., Garcia, W., Bindschaedler, V., Shrimpton, T., Traynor, P.: Beyondl_pclipping: Equalization based psychoa- coustic attacks against asrs. In: Asian Conference on Machine Learning. pp. 672–

    work page

  2. [2]

    In: Machine Learning for Music Discovery Workshop, International Conference on Machine Learning (ICML)

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

    work page 2019

  3. [3]

    In: 2017 IEEE Symposium on Security and Privacy (SP)

    Carlini, N., Wagner, D.: Towards evaluating the robustness of neural networks. In: 2017 IEEE Symposium on Security and Privacy (SP). pp. 39–57. IEEE (2017)

    2017

  4. [4]

    Multimedia Tools and Applications82(17), 26451–26463 (2023)

    Chen, L., Wang, R., Dong, L., Yan, D.: Imperceptible adversarial audio steganog- raphy based on psychoacoustic model. Multimedia Tools and Applications82(17), 26451–26463 (2023)

    work page 2023

  5. [5]

    In: 18th International Society for Music Information Retrieval Con- ference (2017)

    Defferrard, M., Benzi, K., Vandergheynst, P., Bresson, X.: FMA: A dataset for music analysis. In: 18th International Society for Music Information Retrieval Con- ference (2017)

    work page 2017

  6. [6]

    In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Du, X., Yu, Z., Zhu, B., Chen, X., Ma, Z.: Bytecover: Cover song identification via multi-loss training. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 551–555. IEEE (2021) PAMT Perception Model 11

    work page 2021

  7. [7]

    In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security

    Duan, R., Qu, Z., Zhao, S., Ding, L., Liu, Y., Lu, Z.: Perception-aware attack: Cre- ating adversarial music via reverse-engineering human perception. In: Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security. pp. 905–919 (2022)

    work page 2022

  8. [8]

    In: 2020 4th International Conference on Electronics, Communication and Aerospace Technology (ICECA)

    Ghildiyal,A.,Singh,K.,Sharma,S.:Musicgenreclassificationusingmachinelearn- ing. In: 2020 4th International Conference on Electronics, Communication and Aerospace Technology (ICECA). pp. 1368–1372. IEEE (2020)

    work page 2020

Show all 33 references
  1. [9]

    IEEE Transactions on Acoustics, Speech, and Signal Processing24(5), 380–391 (1976)

    Gray, A., Markel, J.: Distance measures for speech processing. IEEE Transactions on Acoustics, Speech, and Signal Processing24(5), 380–391 (1976)

    1976

  2. [10]

    Advances in Neural Information Processing Systems34, 20916–20928 (2021)

    Greshler, G., Shaham, T., Michaeli, T.: Catch-a-waveform: Learning to generate audio from a single short example. Advances in Neural Information Processing Systems34, 20916–20928 (2021)

    2021

  3. [11]

    In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2017)

    Hershey, S., Chaudhuri, S., Ellis, D.P., Gemmeke, J.F., Jansen, A., Moore, R.C., Plakal, M., Platt, D., Saurous, R.A., Seybold, B.: CNN architectures for large- scale audio classification. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Pro...

    2017

  4. [12]

    In: Interspeech

    Hu, S., Zhang, B., Lu, J., Jiang, Y., Wang, W., Kong, L., Zhao, W., Jiang, T.: Wideresnet with joint representation learning and data augmentation for cover song identification. In: Interspeech. pp. 4187–4191 (2022)

    2022

  5. [13]

    IEEE Transactions on Audio, Speech, and Language Processing14(6), 1902–1911 (2006)

    Huber, R., Kollmeier, B.: PEMO-Q—A new method for objective audio quality assessment using a model of auditory perception. IEEE Transactions on Audio, Speech, and Language Processing14(6), 1902–1911 (2006)

    1902

  6. [14]

    Hung,Y.N.,Yang,C.H.H.,Chen,P.Y.,Lerch,A.:Low-resourcemusicgenreclassifi- cationwithcross-modalneuralmodelreprogramming.In:ICASSP2023-2023IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2023)

    2023

  7. [15]

    In: Proceed- ings of Interspeech (2019)

    Kilgour, K., Zuluaga, M., Roblek, D., Sharifi, M.: Fréchet Audio Distance: A reference-free metric for evaluating music enhancement algorithms. In: Proceed- ings of Interspeech (2019)

    2019

  8. [16]

    IEEE/ACM Transactions on Audio, Speech, and Language Processing28, 2880–2894 (2020)

    Kong, Q., Cao, Y., Iqbal, T., Wang, Y., Wang, W., Plumbley, M.D.: PANNs: Large- scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing28, 2880–2894 (2020)

    2020

  9. [17]

    arXiv preprint arXiv:2306.00107 (2023)

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

    2023 arXiv

  10. [18]

    In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Manocha, P., Jin, Z., Zhang, R., Finkelstein, A.: Cdpam: Contrastive learning for perceptual audio similarity. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 196–200. IEEE (2021)

    2021

  11. [19]

    IEEE Journal of Selected Topics in Signal Processing15(1), 120–131 (2020)

    Marafioti, A., Majdak, P., Holighaus, N., Perraudin, N.: GACELA: A generative adversarial context encoder for long audio inpainting of music. IEEE Journal of Selected Topics in Signal Processing15(1), 120–131 (2020)

    2020

  12. [20]

    In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Moliner, E., Lehtinen, J., Välimäki, V.: Solving audio inverse problems with a dif- fusion model. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2023)

    2023

  13. [21]

    IEEE Intelligent Systems36(5), 57–68 (2020)

    Moscato, V., Picariello, A., Sperli, G.: An emotional recommender system for mu- sic. IEEE Intelligent Systems36(5), 57–68 (2020)

    2020

  14. [22]

    IEEE/ACM Transactions on Audio, Speech, and Language Processing (2023) 12 Y

    Nayem,K.M.,Williamson,D.S.:Attention-basedspeechenhancementusinghuman quality perception modelling. IEEE/ACM Transactions on Audio, Speech, and Language Processing (2023) 12 Y. Liu et al

    2023

  15. [23]

    Transactions of the International Society for Music Information Retrieval4(1), 93–105 (2021)

    Prinz, K., Flexer, A., Widmer, G.: On end-to-end white-box adversarial attacks in music information retrieval. Transactions of the International Society for Music Information Retrieval4(1), 93–105 (2021)

    2021

  16. [24]

    In: Interna- tional Conference on Machine Learning

    Qin, Y., Carlini, N., Cottrell, G., Goodfellow, I., Raffel, C.: Imperceptible, robust, and targeted adversarial examples for automatic speech recognition. In: Interna- tional Conference on Machine Learning. pp. 5231–5240. PMLR (2019)

    2019

  17. [25]

    In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Ragano, A., Skoglund, J., Hines, A.: Nomad: Unsupervised learning of perceptual embeddings for speech enhancement and non-matching reference audio quality assessment. In: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 10...

    2024

  18. [26]

    In: ICASSP 2025- 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Riou, A., Gagneré, A., Hadjeres, G., Lattner, S., Peeters, G.: Zero-shot musical stem retrieval with joint-embedding predictive architectures. In: ICASSP 2025- 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2025)

    2025

  19. [27]

    In: International Conference on Machine Learning

    Saadatpanah, P., Shafahi, A., Goldstein, T.: Adversarial attacks on copyright de- tection systems. In: International Conference on Machine Learning. pp. 8307–8315. PMLR (2020)

    2020

  20. [28]

    Frontiers in Applied Mathematics and Statistics5, 457883 (2019)

    Schedl, M.: Deep learning in music recommendation systems. Frontiers in Applied Mathematics and Statistics5, 457883 (2019)

    2019

  21. [29]

    arXiv preprint arXiv:1808.05665 (2018)

    Schönherr, L., Kohls, K., Zeiler, S., Holz, T., Kolossa, D.: Adversarial attacks against automatic speech recognition systems via psychoacoustic hiding. arXiv preprint arXiv:1808.05665 (2018)

    2018 arXiv

  22. [30]

    In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Serrà, J., Pons, J., Pascual, S.: Sesqa: semi-supervised learning for speech quality assessment. In: ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 381–385. IEEE (2021)

    2021

  23. [31]

    IEEE Trans- actions on speech and audio processing10(5), 293–302 (2002)

    Tzanetakis, G., Cook, P.: Musical genre classification of audio signals. IEEE Trans- actions on speech and audio processing10(5), 293–302 (2002)

    2002

  24. [32]

    In: 2022 29th Asia-Pacific Software Engineering Conference (APSEC)

    Wu, X., Rajan, A.: Catch me if you can: Blackbox adversarial attacks on automatic speech recognition using frequency masking. In: 2022 29th Asia-Pacific Software Engineering Conference (APSEC). pp. 169–178. IEEE (2022)

    2022

  25. [33]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 586–595 (2018)

    2018

Pith tools

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