Pith. sign in

REVIEW 2 major objections 6 minor 44 references

Probing Low-Level Acoustic Attribute Encoding in CLAP Audio Embeddings

T0 review · 2 major / 6 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read Low-level acoustic attributes like reverb, loudness, and pitch sit inside frozen CLAP embeddings and can be read out with simple probes.

desk verdict Solid multi-dataset probing map of RT60/LUFS/SC/RP in CLAP: attributes are recoverable, with a clean linear vs non-linear split that mostly generalizes; main caveat is synthetic single-attribute labels, already flagged by the authors. read the letter →

arxiv 2607.03806 v1 pith:6HAI5OOB submitted 2026-07-04 eess.AS cs.AIeess.SP

Héctor Martel , Joe Hennessy-Priest , Taemin Cho This is my paper
classification eess.AScs.AIeess.SP
keywords CLAPaudioembeddingsprobingRT60LUFSspectralcentroidrelativepitchfoundationmodels
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

CLAP and related audio foundation models are used as black-box feature extractors, but it has been unclear which basic physical properties of sound they actually keep. This paper freezes the audio encoder and trains lightweight probes to recover reverberation time (RT60), integrated loudness (LUFS), spectral centroid, and relative pitch from the embeddings alone. Across five datasets that range from pure noise to speech, single notes, and full music mixes, every attribute is recoverable. Reverb, loudness, and relative pitch are largely linear: a single direction in embedding space already predicts them well. Spectral centroid usually needs a non-linear probe. The same two regimes appear in eight other popular audio models, except that architectures that force amplitude invariance throw loudness away by design. The linear directions for reverb and loudness stay roughly aligned across domains; the pitch direction does not. Text prompts that describe reverb also land in the same geometric neighborhood as the audio-derived reverb axis. The practical upshot is that one frozen embedding can supply simultaneous estimates of several low-level acoustic parameters that production tools care about.

What carries the argument

The probing pipeline itself: freeze the audio encoder, apply controlled single-attribute augmentations that generate ground-truth labels, then train linear, shallow MLP, and kernel-ridge probes on the frozen embeddings. Performance gaps between the three probes reveal whether an attribute is linearly or non-linearly encoded, while the linear weight vector supplies an explicit feature axis whose geometric consistency can be measured across datasets.

What would settle it

Train the same linear and non-linear probes on embeddings of completely dry, real-room recordings whose RT60 and LUFS have been measured independently with calibrated physical instruments rather than synthetic convolution and loudness normalization; if recovery collapses or the recovered directions become uncorrelated with the original synthetic axes, the central claim fails.

Watch

Extended reading notes

Core claim

All four low-level acoustic attributes examined—RT60, LUFS, spectral centroid, and relative pitch—are reliably present in frozen CLAP audio embeddings and can be recovered by probes of modest complexity across noise, speech, monophonic notes, and music mixtures. RT60, LUFS, and relative pitch are approximately linearly encoded; spectral centroid generally lies on a curved manifold that requires non-linear probes. Both regimes hold for eight additional foundation models, with the architectural exception that amplitude-normalized models discard loudness entirely.

Load-bearing premise

The synthetic single-attribute augmentations (especially shoebox room impulse responses and independent pitch or bandpass shifts) produce clean ground-truth labels that isolate the intended physical quantity without residual reverb or pipeline artifacts that the probes could exploit instead.

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

2 major / 6 minor

Summary. The paper systematically probes frozen LAION-CLAP audio embeddings for four low-level acoustic attributes—RT60, LUFS, spectral centroid (SC), and relative pitch (RP)—using linear, MLP, and RBF kernel probes across five datasets (white noise, NSynth, VCTK, MusDB18HQ, SonicMaster). Labels are generated by independent single-attribute augmentations so that co-variation is not induced by the pipeline. The central empirical claim is that all four attributes are reliably recoverable; RT60, LUFS, and RP are approximately linearly encoded (Linear R² often high), while SC generally requires non-linear probes. Linear feature axes for RT60 and LUFS show moderate cross-dataset geometric consistency, whereas RP axes are domain-specific. The two encoding regimes largely transfer to eight other audio foundation models on VCTK, with the expected exception that amplitude-invariant architectures discard LUFS. A qualitative text-side check shows RT60-related prompts aligning with the audio-derived RT60 axis.

Significance. If the recoverability and linearity results hold under the stated augmentation regime, the work supplies a clear, quantitative map of which low-level acoustic dimensions sit in shared audio foundation embeddings and how they are geometrically organized. That is directly useful for practitioners who treat CLAP-like models as frozen feature extractors for effect estimation, mix analysis, quality assessment, or text-driven control, and it helps reconcile prior mixed claims about whether CLAP encodes reverberation or loudness. Strengths include multi-dataset coverage, multi-seed reporting, honest failure modes (Linear SC collapse; LUFS ≈ 0 on amplitude-invariant models), independent per-attribute augmentation, feature-axis cosine analysis (Fig. 2), data-efficiency curves (Table 2), and multi-embedder replication (Table 3). The contribution is empirical and interpretive rather than algorithmic, but the evidence base is unusually thorough for a probing study in this area.

major comments (2)
  1. §3.2 and §5: Residual reverberation in MusDB18HQ and SonicMaster is acknowledged as a limitation, but never quantified. Because RT60 targets are defined by the applied synthetic RIR after RMS renormalization, even a modest pre-existing RT60 floor could bias absolute labels and inflate apparent linear recoverability on those two sets. A short control—e.g., estimated RT60 (or energy-decay) statistics on the unaugmented mixes, or a dry-only subset comparison—would make the claim that “the applied RIR dominates” falsifiable rather than asserted. This does not overturn Table 1, but it is load-bearing for interpreting the music-mixture RT60 numbers.
  2. §4 (SC vs RP) and §2.1: The paper treats SC and RP as complementary scales of the same spectral property, yet Linear probes fail on SC on most datasets while succeeding on RP (Table 1). The log-mel / log-semitone explanation is plausible but remains post-hoc; no ablation isolates scale (e.g., regressing log-SC or linear-frequency RP, or matching target distributions). Without that, the strong claim of two distinct “encoding regimes” partly conflates embedding geometry with target reparameterization. A small controlled reparameterization experiment would secure the regime distinction that structures the abstract and conclusion.
minor comments (6)
  1. Table 1: The † marker (R² < −1) is applied inconsistently relative to other negative R² values that are printed explicitly; a single rule in the caption would help.
  2. §4.4 / Table 4: The text experiment is correctly labeled qualitative, but the abstract’s phrasing (“cross-modal consistency”) is slightly stronger than the LUFS failure and the absence of an RP text check warrant. Soften the abstract sentence to match the body.
  3. Figure 3: SC is omitted because the linear probe fails; a brief note that the same projection for MLP/Kernel is out of scope (or a one-panel non-linear illustration) would avoid the impression that SC is simply unvisualizable.
  4. §3.2: Clarify that RP uses the SC-derived value after the SC augmentation (no separate RP augmentation), so SC and RP share the same perturbed waveforms within each dataset; this is implied but easy to miss.
  5. References / related work: [6] is cited as arguing CLAP fails to quantify noise or reverberation; a one-sentence contrast with your RT60 results in the discussion would help readers place the contribution.
  6. Typos / polish: “ind= 512” spacing in §4.1; occasional missing spaces around math (e.g., R 2); “shoebox” geometry is fine but could be named once as rectangular parallelepiped for non-acoustics readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: standard frozen-embedding probing against independently generated augmentation labels; recoverability is empirical, not forced by definition or self-citation.

full rationale

The paper's load-bearing claims are empirical recoverability results (Tables 1–3): probes of increasing complexity are trained on frozen CLAP (and eight other) embeddings to regress RT60, LUFS, SC, and RP labels produced by single-attribute synthetic augmentation (gpuRIR/Sabine RIRs with RMS renormalization; pyloudnorm LUFS; pitch-shift/bandpass SC; RP as log transform of SC). The encoder is never trained on these targets, so probe success is not definitional. Linear feature axes are fit independently per dataset then compared by cosine similarity; the text-side RT60 check applies an audio-trained probe to hand-crafted text embeddings as a qualitative geometric demonstration, not a fitted prediction. Amplitude-invariant models discarding LUFS is correctly attributed to architectural normalization, not a circular claim. No uniqueness theorems, no load-bearing self-citation chain, no ansatz smuggled as derivation, and no renaming of a known result as a first-principles prediction. Limitations (shoebox RIRs, residual reverb in mixes) are stated in §5 and do not create circular reduction of outputs to inputs. Score 0 is the honest finding.

Assumptions & free parameters 4 free parameters · 6 assumptions · 1 invented entities

Empirical probing paper: the claim rests on standard acoustic definitions, public models/datasets, and experimental design choices rather than new physical entities. Load-bearing assumptions are that synthetic single-attribute augmentations yield faithful labels and that final-layer frozen embeddings are a valid place to measure encoding. Free parameters are probe/training hyperparameters and augmentation ranges, not constants fitted to invent the main result.

free parameters (4)
  • MLP hidden size (64) and KRR γ=1.0, λ=1e-3, M≤10^4
    Probe capacity and kernel settings chosen by authors; performance margins between Linear/MLP/Kernel depend on these fixed choices, though multiple probe classes reduce single-hyperparameter risk.
  • RT60 augmentation range U(0.0, 2.0)s and shoebox room dims [4,12]m via Sabine/gpuRIR
    Defines the label distribution and physical model of reverb the probes learn; outside this range (e.g. cathedral >2s) text predictions already fail.
  • LUFS target range U(-40, -10) and SC pitch-shift U(-6,6) semitones / bandpass fc U(500,5000)Hz
    Hand-chosen balanced ranges that determine what 'recoverable' means in the reported R²/MAE.
  • Linear/MLP training: AdamW lr=1e-3, wd=1e-3, 100 epochs, batch 256, early stop patience 10
    Optimization settings that can affect absolute probe scores; mitigated by 10-seed averaging and Kernel non-gradient baseline.
assumptions (6)
  • domain assumption Integrated loudness follows ITU-R BS.1770 (pyloudnorm) and is the right perceptual loudness target.
    §2.1; LUFS encoding claims depend on this standard definition of loudness.
  • domain assumption RT60 from synthetic shoebox RIRs via Sabine-derived reflection coefficients is a valid global reverb descriptor for probing.
    §3.2 and limitation in §5; central RT60 recoverability results use this generative model of rooms.
  • domain assumption Spectral centroid (and log-semitone RP) are appropriate domain-agnostic proxies for spectral content / pitch-like structure, including in polyphonic mixes where f0 is ill-posed.
    §2.1; SC/RP conclusions inherit this measurement choice.
  • ad hoc to paper Independent single-attribute augmentation prevents co-variation confounds so correlations between probe directions reflect embedding geometry.
    §3.2 design claim; if violated, axis consistency and multi-attribute conclusions weaken.
  • domain assumption Final-layer frozen embeddings are sufficient to characterize what the model encodes about these attributes.
    §2 methodology and §5 limitation (no intermediate layers).
  • standard math Standard linear algebra / regression metrics (MAE, R², Pearson r, cosine of weight vectors) correctly quantify linear vs non-linear encoding.
    §2.2–2.3 evaluation framework.
invented entities (1)
  • feature axis (linear probe weight vector w as attribute direction) independent evidence
    purpose: Operationalize geometric consistency of RT60/LUFS/RP encoding across datasets and for text projection demos.
    Not a new physical object; it is the standard linear-probe direction. Treated as an analysis construct with independent handle via cross-dataset cosine similarity and projection plots.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probing Low-Level Acoustic Attribute Encoding in CLAP Audio Embeddings." pith.science (2026). https://pith.science/paper/6HAI5OOB

@misc{pith2026260703806,
  author       = {Pith},
  title        = {Pith review of: Probing Low-Level Acoustic Attribute Encoding in CLAP Audio Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6HAI5OOB}},
  note         = {Machine review of arXiv:2607.03806}
}
read the original abstract

Audio foundation models are widely adopted as general-purpose feature extractors, yet the internal structure of their learned representations remains insufficiently understood. In this work, we analyze CLAP audio embeddings through a probing framework, studying the encoding of three fundamental perceptual dimensions: reverberation (RT60), loudness (LUFS), and spectral content, measured via spectral centroid (SC) and relative pitch (RP). Probes of increasing complexity are trained to predict each attribute from frozen embeddings across five datasets spanning noise, speech, monophonic musical notes, and music mixtures. Our primary finding is that all of these attributes are reliably recoverable from the CLAP embedding space across the examined datasets. Within this global picture, two encoding regimes emerge: RT60, LUFS, and RP are approximately linearly encoded, while SC requires non-linear probes. Both regimes generalize across eight additional audio foundation models, with the notable exception that amplitude-invariant architectures discard loudness entirely by construction. The identified linear feature directions are geometrically consistent across datasets for RT60 and LUFS, while highly domain-specific for RP. Finally, we provide a qualitative demonstration of cross-modal consistency, showing that text embeddings of acoustic descriptors align geometrically with the identified RT60 feature direction.

Figures

Figures reproduced from arXiv: 2607.03806 by the authors.

Figure 1
Figure 1. Overview of the probing methodology. Interpretability approaches for acoustic models have favored qual￾itative concept alignment [17] over quantitative prediction of phys￾ical descriptors. Within the CLAP literature, studies have shown that high-level features such as timbre [18] and emotion [19] are effectively captured. However, low-level acoustic attributes have received far less attention. [20] shows that applyi… view at source ↗
Figure 2
Figure 2. Pairwise cosine similarity of linear probe weight vectors w ∈ R 512 trained independently on each dataset, for RT60 (left), LUFS (center), and RP (right). As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. VCTK test-set embeddings projected onto the linear-probe weight vector (x-axis) vs. ground-truth value (GT, y-axis), colored by target value. Both RT60 and LUFS show a monotone band with mild curvature, confirming a predominantly linear structure consistent with R 2 = 0.92 for the Linear probe on VCTK-Corpus ( [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: PCA projection of LAION-CLAP embeddings for the VCTK-Corpus test set, colored by RT60. Projected text descrip￾tions confirm that the qualitative descriptors dry and long reverb land at geometrically coherent positions along the RT60 axis iden￾tified by the Linear probe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 4 linked inside Pith

  1. [1]

    INTRODUCTION CLAP [1, 2] is a prominent audio-language foundation model that leverages contrastive learning to align audio and text in a shared embedding space. Trained on large collections of audio-text pairs, CLAP learns embeddings that capture both acoustic characteris- tics and semantic content, enabling applications such as audio re- trieval, caption...

    2026

  2. [2]

    METHODOLOGY An overview of the probing methodology is presented in Figure 1. Each waveformx∈R C×T , withCchannels andTsamples, is transformed with an attribute-dependent data augmentation and downmixed to mono obtainingx ′ ∈R 1×T .x ′ is then encoded intoz∈R d (d= 512), which serves as the sole input to all probes. The probes produce the predictions for s...

    arXiv 2026

  3. [3]

    EXPERIMENTAL SETTINGS 3.1. Datasets To test generalization across domains, we evaluate probes on five datasets spanning noise, speech, and music at increasing complex- ity, from monophonic notes to full music mixtures. All audio is resampled to 48 kHz to match the sample rate of LAION-CLAP. White Noise(100k samples) is a synthetic control of gener- ated w...

    2026

  4. [4]

    RESULTS Results are shown in Table 1. A global trend can be observed: non-linear probes (MLP, Kernel) consistently outperform the Lin- ear probe across all features and datasets, with the performance margin growing from modest gains for RT60 and LUFS, to sub- stantial gains for SC, and moderate gains for RP. This implies that the geometry of CLAP’s encodi...

    2026

  5. [5]

    CONCLUSION We presented a systematic probing study of three fundamental per- ceptual dimensions, reverberation (RT60), loudness (LUFS), and spectral content (SC and RP), in CLAP audio embeddings, us- ing Linear, MLP, and Kernel Ridge Regression probes across five datasets spanning noise, speech, and music. Our primary finding is that all the attributes we...

  6. [6]

    Clap: Learning audio concepts from natural language su- pervision,

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

    2023

  7. [7]

    Natural language supervision for general-purpose audio representations,

    B. Elizalde, S. Deshmukh, and H. Wang, “Natural language supervision for general-purpose audio representations,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 336–340

    2024

  8. [8]

    Transformation of audio embeddings into interpretable, concept-based representations,

    A. Zhang, E. Thomaz, and L. Lu, “Transformation of audio embeddings into interpretable, concept-based representations,” 2025. [Online]. Available: https://arxiv. org/abs/2504.14076

    arXiv 2025

Show all 44 references
  1. [9]

    Drcap: Decoding clap latents with retrieval- augmented generation for zero-shot audio captioning,

    X. Li, W. Chen, Z. Ma, X. Xu, Y . Liang, Z. Zheng, Q. Kong, and X. Chen, “Drcap: Decoding clap latents with retrieval- augmented generation for zero-shot audio captioning,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  2. [10]

    Pam: Prompting audio-language models for audio quality assessment,

    S. Deshmukh, D. Alharthi, B. Elizalde, H. Gamper, M. Al Is- mail, R. Singh, B. Raj, and H. Wang, “Pam: Prompting audio-language models for audio quality assessment,” in Proc. Interspeech 2024, 2024, pp. 3320–3324

    2024

  3. [11]

    Listen through the sound: Generative speech restoration leveraging acoustic context representation,

    S.-W. Chung and M.-S. Choi, “Listen through the sound: Generative speech restoration leveraging acoustic context representation,” inProc. Interspeech 2025, 2025, pp. 4843– 4847

    2025

  4. [12]

    Text2FX: Harnessing clap embeddings for text-guided audio effects,

    A. Chu, P. O’Reilly, J. Barnett, and B. Pardo, “Text2FX: Harnessing clap embeddings for text-guided audio effects,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  5. [13]

    Multimodal room impulse response generation through latent rectified flow matching,

    A. V osoughi, Y . Zang, Q. Yang, N. Paek, R. Leistikow, and C. Xu, “Multimodal room impulse response generation through latent rectified flow matching,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2026, pp. 14 627–14 631

    2026

  6. [14]

    TokenSynth: A token-based neural synthesizer for instrument cloning and text-to-instrument,

    K. Kim, J. Koo, S. Lee, H. Joung, and K. Lee, “TokenSynth: A token-based neural synthesizer for instrument cloning and text-to-instrument,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  7. [15]

    FlowSynth: Instru- ment generation through distributional flow matching and test-time search,

    Q. Yang, R. Leistikow, and Y . Zang, “FlowSynth: Instru- ment generation through distributional flow matching and test-time search,”arXiv preprint arXiv:2510.21667, 2025

    2025

  8. [16]

    Make-An-Audio: Text-to- audio generation with prompt-enhanced diffusion models,

    R. Huang, J. Huang, D. Yang, Y . Ren, L. Liu, M. Li, Z. Ye, J. Liu, X. Yin, and Z. Zhao, “Make-An-Audio: Text-to- audio generation with prompt-enhanced diffusion models,” inInternational Conference on Machine Learning (ICML). PMLR, 2023, pp. 13 916–13 932

    2023

  9. [17]

    T-clap: Temporal-enhanced contrastive language-audio pretraining,

    Y . Yuan, Z. Chen, X. Liu, H. Liu, X. Xu, D. Jia, Y . Chen, M. D. Plumbley, and W. Wang, “T-clap: Temporal-enhanced contrastive language-audio pretraining,” inIEEE 34th Inter- national Workshop on Machine Learning for Signal Process- ing (MLSP). IEEE, 2024, pp. 1–6

    2024

  10. [18]

    Spatial-clap: Learning spatially-aware audio–text embeddings for multi-source conditions,

    K. Seki, Y . Okamoto, K. Yamaoka, Y . Saito, S. Takamichi, and H. Saruwatari, “Spatial-clap: Learning spatially-aware audio–text embeddings for multi-source conditions,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 14 742–14 746

    2026

  11. [19]

    Probing the information encoded in x-vectors,

    D. Raj, D. Snyder, D. Povey, and S. Khudanpur, “Probing the information encoded in x-vectors,” inIEEE Automatic DAFx.7 Proceedings of the 29th International Conference on Digital Audio Effects (DAFx26), Cambridge, MA, USA, 1–4 September 2026 Speech Recognition and Understandin...

    2026 doi

  12. [20]

    Audiolens: A closer look at auditory attribute perception of large audio-language models,

    C.-K. Yang, N. Ho, Y .-J. Lee, and H. yi Lee, “Audiolens: A closer look at auditory attribute perception of large audio-language models,” 2025. [Online]. Available: https: //arxiv.org/abs/2506.05140

    2025 arXiv

  13. [21]

    Causal tracing of audio-text fusion in large audio language models,

    W.-C. Chen, C. yu Huang, and H. yi Lee, “Causal tracing of audio-text fusion in large audio language models,” 2026. [Online]. Available: https://arxiv.org/abs/2603.13768

    2026

  14. [22]

    AND: Audio network dissection for interpreting deep acoustic models,

    T.-Y . Wu, Y .-X. Lin, and T.-W. Weng, “AND: Audio network dissection for interpreting deep acoustic models,” inInter- national Conference on Machine Learning (ICML). PMLR, 2024, pp. 53 656–53 680

    2024

  15. [23]

    Do joint language- audio embeddings encode perceptual timbre semantics?

    Q. Deng, B. Pardo, and T. N. Pappas, “Do joint language- audio embeddings encode perceptual timbre semantics?” arXiv preprint arXiv:2510.14249, 2025

    2025

  16. [24]

    Evaluating foundation models on timbre-related cognitive tasks,

    G. Velissaridis, R. Athwal, M. Musharaf, G. Fazekas, and C. Saitis, “Evaluating foundation models on timbre-related cognitive tasks,” in1st Workshop on Large Language Models for Music & Audio (LLM4MA), 2025

    2025

  17. [25]

    Investigat- ing the sensitivity of pre-trained audio embeddings to com- mon effects,

    V . Deng, C. Wang, G. Richard, and B. McFee, “Investigat- ing the sensitivity of pre-trained audio embeddings to com- mon effects,” inIEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–5

    2025

  18. [26]

    Fx-encoder++: Extracting instrument-wise audio effects representations from mix- tures,

    Y .-T. Yeh, J. Koo, M. Martínez-Ramírez, W.-H. Liao, Y .- H. Yang, and Y . Mitsufuji, “Fx-encoder++: Extracting instrument-wise audio effects representations from mix- tures,” inProceedings of the 26th International Society for Music Information Retrieval Conference (ISMIR), 2025

    2025

  19. [27]

    Large-scale contrastive language-audio pre- training with feature fusion and keyword-to-caption aug- mentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language-audio pre- training with feature fusion and keyword-to-caption aug- mentation,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023

    2023

  20. [28]

    Hts-at: A hierarchical token-semantic audio transformer for sound classification and detection,

    K. Chen, X. Du, B. Zhu, Z. Ma, T. Berg-Kirkpatrick, and S. Dubnov, “Hts-at: A hierarchical token-semantic audio transformer for sound classification and detection,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022

    2022

  21. [29]

    New method of measuring reverberation time,

    M. R. Schroeder, “New method of measuring reverberation time,”The Journal of the Acoustical Society of America, vol. 37, no. 3, pp. 409–412, 1965

    1965

  22. [30]

    Algorithms to measure audio programme loudness and true-peak audio level,

    ITU-R, “Algorithms to measure audio programme loudness and true-peak audio level,” International Telecommunication Union, Recommendation ITU-R BS.1770-5, 11 2023. [Online]. Available: https://www.itu.int/rec/R-REC-BS. 1770-5-202311-I/en

    2023

  23. [31]

    Müller,Fundamentals of Music Processing: Audio, Anal- ysis, Algorithms, Applications

    M. Müller,Fundamentals of Music Processing: Audio, Anal- ysis, Algorithms, Applications. Springer, 2015, vol. 5

    2015

  24. [32]

    J. O. S. III,Spectral Audio Signal Processing. W3K Publishing, 2011. [Online]. Available: https://ccrma. stanford.edu/~jos/sasp/

    2011

  25. [33]

    Schölkopf and A

    B. Schölkopf and A. J. Smola,Learning with kernels: sup- port vector machines, regularization, optimization, and be- yond. MIT press, 2002

    2002

  26. [34]

    Neural audio synthesis of musi- cal notes with wavenet autoencoders,

    J. Engel, C. Resnick, A. Roberts, S. Dieleman, M. Norouzi, D. Eck, and K. Simonyan, “Neural audio synthesis of musi- cal notes with wavenet autoencoders,” inInternational Con- ference on Machine Learning (ICML). PMLR, 2017, pp. 1068–1077

    2017

  27. [35]

    CSTR VCTK corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),

    J. Yamagishi, C. Veaux, and K. MacDonald, “CSTR VCTK corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),” 2019

    2019

  28. [36]

    MUSDB18-HQ - an uncompressed version of musdb18,

    Z. Rafii, A. Liutkus, F.-R. Stöter, S. I. Mimilakis, and R. Bittner, “MUSDB18-HQ - an uncompressed version of musdb18,” Dec. 2019. [Online]. Available: https://doi.org/10.5281/zenodo.3338373

    2019 doi

  29. [37]

    SonicMaster: Towards controllable all-in- one music restoration and mastering,

    J. Melechovsky, A. Mehrish, A. Roy, and D. Herre- mans, “SonicMaster: Towards controllable all-in- one music restoration and mastering,”arXiv preprint arXiv:2508.03448, 2025

    2025 arXiv

  30. [38]

    gpurir: A python library for room impulse response simulation with gpu acceleration,

    D. Diaz-Guerra, A. Miguel, and J. R. Beltran, “gpurir: A python library for room impulse response simulation with gpu acceleration,”Multimedia Tools and Applications, vol. 80, no. 4, pp. 5653–5671, 2021

    2021

  31. [39]

    Mert: Acoustic music understanding model with large-scale self-supervised train- ing,

    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-scale self-supervised train- ing,” inInternational Conference on Learning Representa- tions (ICLR), 2024, pp. 12 181–12 204

    2024

  32. [40]

    wav2vec 2.0: A framework for self-supervised learning of speech rep- resentations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech rep- resentations,”Advances in Neural Information Processing Systems (NeurIPS), vol. 33, pp. 12 449–12 460, 2020

    2020

  33. [41]

    Wavlm: Large- scale self-supervised pre-training for full stack speech pro- cessing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiaoet al., “Wavlm: Large- scale self-supervised pre-training for full stack speech pro- cessing,”IEEE Journal of Selected Topics in Signal Process- ing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  34. [42]

    Robust speech recognition via large-scale weak supervision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational Conference on Machine Learning (ICML). PMLR, 2023, pp. 28 492–28 518

    2023

  35. [43]

    Cnn architectures for large-scale audio classification,

    S. Hershey, S. Chaudhuri, D. P. Ellis, J. F. Gemmeke, A. Jansen, R. C. Moore, M. Plakal, D. Platt, R. A. Saurous, B. Seyboldet al., “Cnn architectures for large-scale audio classification,” inIEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). ...

    2017

  36. [44]

    Adapting Fréchet audio distance for generative music eval- uation,

    A. Gui, H. Gamper, S. Braun, and D. Emmanouilidou, “Adapting Fréchet audio distance for generative music eval- uation,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 1331–1335. DAFx.8

    2024

Pith tools

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