Pith. sign in

REVIEW 3 major objections 6 minor 50 references

NABEATs: Noise-Aware Audio Representation Learning

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that conditioning a frozen BEATs audio encoder on a reference noise signal—instead of on the target sound—lets it recover clean representations from noisy mixtures, improving downstream classification and anomaly detection

desk verdict A practical, clearly evaluated method for noise-aware audio SSL; the main untested assumption is that the reference noise matches the actual noise, so the generalization claim only holds when a good reference is available. read the letter →

arxiv 2607.16688 v1 pith:2BWUG7VH submitted 2026-07-18 eess.AS

Takuya Fujimura , Yoshiki Masuyama , Gordon Wichern , Christoph Boeddeker , Julius Richter , Jonathan Le Roux This is my paper
classification eess.AS
keywords noise-awareaudioSSLBEATsreferencenoiseconditioningconditionaldenoisingcross-attentionFiLMrobustnessself-supervisedrepresentationlearning
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 proposes that general-purpose audio self-supervised models can be made robust to noise by giving them access to a reference noise signal at inference time. It introduces NABEATs, a version of the BEATs encoder augmented with trainable noise-aware layers that take both the noisy mixture and a reference noise segment and are trained to reproduce the clean BEATs representation. The authors claim this conditional denoising substantially improves linear-probe accuracy across six classification tasks in environmental sound, speech, and music under noisy conditions, and generalizes to unseen noise types better than either the original BEATs or a no-reference denoising baseline (DBEATs). They also report gains on the DCASE 2025 Task 2 anomaly detection benchmark. The central idea is that conditioning on noise, rather than on the target sound, is practical because noise reference segments are often available in real-world deployments.

What carries the argument

The load-bearing mechanism is conditional denoising in representation space: the model never outputs a waveform or a mask; instead it predicts the clean feature sequence of a frozen BEATs encoder. The conditioning is implemented through two lightweight adapter variants—cross-attention (NABEATs-CA), where the noisy representation attends to the reference-noise representation frame by frame, and FiLM (NABEATs-FiLM), where the sequence-averaged reference representation produces per-channel scale and shift parameters. Both variants reuse the frozen BEATs components for both inputs and add only a small number of trainable parameters, trained with an MSE loss against the clean BEATs target.

What would settle it

Evaluate NABEATs-CA against DBEATs in a controlled setup where the reference noise n' is intentionally mismatched—taken from a different recording, a different device, or a time segment far removed from the mixture—and show that downstream accuracy falls below the no-reference baseline. Alternatively, a synthetic test with rapidly changing non-stationary noise (e.g., a new noise type every 0.5 s) where the reference only matches the first segment should expose whether cross-attention truly tracks time-varying noise or merely uses the reference as a global hint.

Watch

Extended reading notes

Core claim

NABEATs is trained by distillation to estimate the representation sequence r = BEATs(s) of the clean target sound s from a noisy mixture x = s + n, using an auxiliary reference noise n' that shares characteristics with n. The architecture keeps the original BEATs frozen and inserts additional trainable 'noise-aware' layers after each encoder layer; two variants use cross-attention (per-frame reference integration) or FiLM (global reference conditioning). The paper demonstrates that on six downstream classification tasks, NABEATs-CA outperforms both BEATs and DBEATs under WHAM! and CHiME-3 noise, and—unlike DBEATs—maintains an advantage even under MUSDB18 music noise, which is far outside the

Load-bearing premise

The approach assumes a reference noise signal is available at inference time and that it resembles the noise present in the mixture; the paper explicitly restricts to a reference obtained from a different time segment of the same recording, so performance under large temporal, spatial, or acoustic reference-noise mismatch is untested.

Editorial extensions

If this is right

  • If NABEATs is right, any existing frozen audio SSL encoder can be made noise-robust by adding noise-aware adapters and training on (mixture, reference-noise, clean-representation) triples, without retraining the base model.
  • Since the reference is noise rather than target sound, the approach is applicable to tasks where the target class is unknown, such as audio tagging and anomaly detection.
  • The cross-attention variant's success on speech and music tasks suggests that time-varying noise suppression benefits from per-frame reference information rather than global statistics.
  • The DCASE result indicates that the reference need not be temporally adjacent; any sufficiently similar noise recording can be selected, widening the practical deployment space.
  • Because NABEATs operates in representation space, it avoids cascaded separation pipelines and their computational overhead.

Reading between the lines

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

  • A natural extension the paper leaves open is explicit robustness to reference mismatch: if the reference segment is short, distant in time, or from a different device, does NABEATs degrade gracefully? The current experiments use same-recording references, so reference-mismatch stress tests would clarify the practical envelope.
  • The same conditional-denoising idea could be transferred to other SSL targets than BEATs, such as token-based speech models, potentially unifying noise-aware speech and audio SSL under one mechanism.
  • The SNR-vs-downstream discrepancy noted for DBEATs on CRM-D hints that MSE in representation space is not perfectly aligned with task performance; a task-oriented or contrastive auxiliary loss might make the representation cleaner for downstream heads.
  • One could test the generalization claim more aggressively by training only on stationary noise and evaluating on highly non-stationary noise with a time-varying reference, isolating whether cross-attention actually tracks noise dynamics.
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

3 major / 6 minor

Summary. The paper introduces noise-aware audio self-supervised learning and its instantiation NABEATs, which augments a frozen BEATs backbone with trainable adapter layers that take a noisy audio input and a separately recorded reference noise signal. The model is trained by MSE distillation to predict the clean BEATs representation from the noisy mixture. The authors also propose DBEATs, a no-reference denoising baseline trained with the same objective. Two conditionings are investigated: cross-attention (NABEATs-CA) and FiLM (NABEATs-FiLM). The method is evaluated on six downstream classification tasks using WHAM! (seen noise), CHiME-3 (unseen environmental noise), and MUSDB18 (unseen music-noise), plus the DCASE 2025 Task 2 anomalous sound detection benchmark. The core empirical claim is that NABEATs significantly improves over BEATs and DBEATs under noisy conditions and generalizes to unseen noise types when the reference noise is obtained from a different time segment of the same recording.

Significance. If the claimed results hold, this is a useful contribution to general-purpose audio representation learning. The idea of conditioning a frozen SSL model on a reference noise signal is novel for non-speech audio and practically motivated, since noise-only segments or separate microphones are often available. The integrated adapter approach avoids cascaded target-sound-extraction systems. Strengths of the empirical study include six downstream tasks covering speech, music, and environmental sound; multiple seeds with 95% confidence intervals in Section 4; held-out noise types (CHiME-3, MUSDB18); and a publicly defined benchmark in Section 5. The main weakness is that all experiments assume a well-matched reference noise (same recording, adjacent or nearest selection), while the abstract and conclusion make the stronger claim of generalization to unseen noise types without testing reference mismatch. The SNR results in Table 3 are partly circular with the training objective, and the DCASE evaluation lacks confidence intervals or significance testing. No code or checkpoints are provided, limiting reproducibility.

major comments (3)
  1. [§2, §3, §4.1, §5] The core advantage of NABEATs over DBEATs is the reference noise input n'. The paper explicitly states in §2 that it focuses on 'the case where the reference noise is obtained from a different time segment from the same recording', and §3 confirms that downstream evaluations use a noise-only segment 'immediately preceding n'. The DCASE experiment (§5) selects the nearest reference from the same machine's noise-only files. No experiment varies reference quality or match — e.g., non-adjacent segments, different recording locations, different channels/devices, or estimated/reconstructed noise. Thus the abstract's claim that NABEATs 'generalizes well to unseen noise types' is only demonstrated when a well-matched reference is available. This is load-bearing: if the reference is uninformative or mismatched, conditioning could degrade rather than improve performance. I recommend adding a refer
  2. [Eq. (1) and Table 3] The SNR in Table 3 is computed directly between the clean representation r and the estimated representation \hat{r}, i.e., the same quantity minimized by the MSE training loss in Eq. (1). Improvements on this metric are therefore partly by construction, not an independent validation of representation quality. The CRM-D row makes this clear: DBEATs improves SNR over BEATs while worsening downstream accuracy. I suggest re-positioning Table 3 as a diagnostic of the training objective and relying on downstream task performance as the primary evidence, rather than presenting SNR as a major result.
  3. [§5] The DCASE 2025 Task 2 scores (BEATs 54.57, DBEATs 55.76, NABEATs-CA 56.15) are reported as a single run without confidence intervals, per-machine breakdown, or significance testing. The NABEATs-CA gain over DBEATs (0.39 points) is small and may be within run-to-run variability. In addition, the evaluation is restricted to 9 of the 15 machine types; the selection criterion is not stated. Please report per-machine scores and multiple runs or a significance test, and justify the subset of machine types.
minor comments (6)
  1. [§4.2] The text contains internal inconsistencies with the numbers in Figure 2. The sentence 'NABEATs-CA and DBEATs improve the performance in the WHAM! and CHiME-3 noise conditions, except for DBEATs on CRM-D' is contradicted by Figure 2: DBEATs improves on CREMA-D under both WHAM! and CHiME-3, whereas under US8K DBEATs is worse than BEATs(N) for all three noise types. Similarly, the statement that NABEATs-CA beats DBEATs 'in all tasks except SPCV2' is not supported, since NABEATs-CA outperforms DBEATs on SPCV2 in all three noise conditions in Figure 2. Please correct these descriptions.
  2. [Fig. 1] The annotation 'Shared weights' in Figure 1 is ambiguous. Clarify whether the original BEATs layers are shared between the noisy input and reference paths, and whether the additional NA layers are shared across encoder layers or separate per layer.
  3. [Eq. (5)–(7)] The left-arrow notation z ← z + ... is nonstandard. Use explicit assignment statements for clarity.
  4. [§5] The reference selection by 'minimum distance to the input noisy machine sound in the original BEATs representation' should specify the distance metric (cosine/Euclidean) and whether the averaged or sequence representation is used.
  5. [Table 3] Table 3 omits CHiME-3 despite the downstream evaluation using it. State why (e.g., space constraints) or include it for completeness.
  6. [General] No code, checkpoints, or data-preprocessing scripts are made available. Given that the method depends on details such as reference-segment extraction, SNR sampling, and dataset splits, releasing these artifacts would substantially improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central claims are validated on held-out downstream labels, not by construction.

full rationale

NABEATs is an empirical system rather than a derivation from first principles. Equation (1) defines a distillation objective in which the frozen BEATs encoder supplies the target r = BEATs(s), and the trainable layers regress r from the noisy input x and reference n'. The paper's central claims are tested on held-out task labels: Figure 2 reports mAP/accuracy on FSD50K, US8K, SPCV2, CREMA-D, NSynth, and Surge under WHAM!, CHiME-3, and MUSDB18 noise. These labels are external to the training loss, and CHiME-3 and MUSDB18 are explicitly treated as unseen noise types. Table 3's SNR is closely tied to the MSE training objective, but the authors themselves caution that SNR can diverge from downstream performance (DBEATs improves SNR on CRM-D but degrades accuracy), so the table is not presented as independent validation. The acknowledged scope limitation in Section 2—reference noise taken from a different time segment of the same recording, with other reference conditions left to future work—limits the breadth of the generalization claim but is not a circular step. No load-bearing self-citation was found: the cited prior work supports implementation choices and scoring details, not the core generalization claim. There are no fitted constants, uniqueness theorems, or ansatz-equivalent inputs that force the result. The evaluation is self-contained against external benchmarks, so the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No new physical entities or fitted constants are introduced. The method's load-bearing assumptions are the availability and validity of the reference noise and the transferability of BEATs-distillation; the only hand-set numeric elements are the mixture SNR range and the reference-segment duration.

free parameters (2)
  • Mixture SNR range = -5 to 10 dB
    Hand-selected range for mixing target and noise in training and all downstream evaluations; the deployment SNR distribution is not evaluated.
  • Reference noise segment duration = Same duration as the noise segment n
    Design choice; a shorter or longer reference could change conditioning quality.
assumptions (5)
  • domain assumption A reference noise signal from a different time segment of the same recording is sufficiently similar to the interfering noise for conditional denoising.
    Invoked in Section 2 and used in all NABEATs training and evaluation reference construction.
  • domain assumption Clean BEATs representations are a suitable supervision target; reducing MSE to BEATs(s) transfers to downstream task accuracy.
    Eqs. (1)-(3) define training; the paper itself notes a mismatch between SNR and downstream performance for CRM-D in Section 4.2, so transfer is imperfect.
  • domain assumption Reference noise is available at inference in target deployments.
    Section 2 states intended scenarios: noise-only segments, pre-recorded noise, or a separately placed microphone.
  • standard math Standard Transformer components (MHSA, MHCA, RMSNorm, SwiGLU, FiLM) behave as in cited prior work.
    Eqs. (5)-(11) rely on these definitions; no proof is needed or given.
  • domain assumption Training on FSD50K target sounds generalizes to speech and music downstream tasks.
    Section 3 uses FSD50K as the only target-sound training data despite downstream tasks in speech and music.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NABEATs: Noise-Aware Audio Representation Learning." pith.science (2026). https://pith.science/paper/2BWUG7VH

@misc{pith2026260716688,
  author       = {Pith},
  title        = {Pith review of: NABEATs: Noise-Aware Audio Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2BWUG7VH}},
  note         = {Machine review of arXiv:2607.16688}
}
read the original abstract

We propose the concept of noise-aware audio self-supervised learning (SSL), whose goal is to encode audio mixtures while suppressing undesired noise, and present Noise-Aware BEATs (NABEATs) as a BEATs-based realization of this framework. Audio SSL models are designed to handle a wide range of audio signals. Consequently, under noisy conditions, they cannot effectively focus on the target sounds relevant to a downstream task, resulting in degraded performance. To address this issue, NABEATs is trained to estimate clean BEATs representations from a noisy audio signal with an auxiliary reference noise input. This reference noise enables the model to account for specific noise characteristics at inference time, thereby achieving better generalization across operating environments. Our experimental evaluations demonstrate that NABEATs significantly improves performance of various downstream tasks under noisy conditions and also generalizes well to unseen noise types.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 3 linked inside Pith

  1. [1]

    These models learn general- purpose representations from large amounts of unlabeled audio data, enabling transfer to various downstream tasks

    INTRODUCTION Self-supervised learning (SSL) models have achieved remarkable success in speech and audio processing. These models learn general- purpose representations from large amounts of unlabeled audio data, enabling transfer to various downstream tasks. For example, speech SSL models [1, 2] have demonstrated impressive performance in downstream tasks...

  2. [2]

    In this paper, we focus arXiv:2607.16688v1 [eess.AS] 18 Jul 2026 on the case where the reference noise is obtained from a different time segment from the same recording

    PROPOSED METHODS The intended usage scenario of NABEATs is similar to that of tra- ditional spectral subtraction [17], where a reference noise signal is obtained from noise-only segments, pre-recorded noise at the target location, or a separately placed microphone. In this paper, we focus arXiv:2607.16688v1 [eess.AS] 18 Jul 2026 on the case where the refe...

    arXiv 2026

  3. [3]

    First, in Section 4, we evalu- ate performance on various downstream tasks under simulated noisy conditions using a custom-designed setup

    EXPERIMENTAL SETUP We conduct two types of evaluations. First, in Section 4, we evalu- ate performance on various downstream tasks under simulated noisy conditions using a custom-designed setup. For the reference noisen′ in NABEATs, we use a noise-only segment immediately precedingn from the same recording, with the same duration, for both NABEATs trainin...

    2025

  4. [4]

    Setups We conducted evaluation on various downstream tasks according to [5, 6]

    EV ALUA TION ON V ARIOUS DOWNSTREAM TASKS 4.1. Setups We conducted evaluation on various downstream tasks according to [5, 6]. We trained only a single linear layer on top of the frozen SSL model for downstream classification tasks. We adopted six down- stream tasks covering environmental sound, speech, and music do- mains. For environmental sound classif...

  5. [5]

    This dataset contains 15 machine types, each with 1000 normal training samples and 200 test samples including normal and anomalous samples

    EV ALUA TION ON DCASE 2025 TASK 2 We also conducted evaluation using the DCASE 2025 Task 2 dataset [24], compiled from [38–40]. This dataset contains 15 machine types, each with 1000 normal training samples and 200 test samples including normal and anomalous samples. Nine of the 15 machine types include 100 noise-only signals as supplementary data, and we...

    2025

  6. [6]

    CONCLUSION In this paper, we explored the concept of noise-aware audio SSL and proposed NABEATs as a realization of this framework based on BEATs. While the standard denoising baseline DBEATs is simply trained to estimate clean representations from noisy signals, NABEATs additionally takes a reference noise as input and is trained to suppress the correspo...

  7. [7]

    HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,”IEEE/ACM TASLP, vol. 29, pp. 3451–3460, 2021

    2021

  8. [8]

    Wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “Wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Proc. NeurIPS, vol. 33, 2020

    2020

Show all 50 references
  1. [9]

    SUPERB: Speech Processing Universal PERfor- mance Benchmark,

    S.-w. Yang et al., “SUPERB: Speech Processing Universal PERfor- mance Benchmark,” inProc. Interspeech, 2021

    2021

  2. [10]

    BEATs: Audio pre-training with acoustic tokeniz- ers,

    S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, W. Che, X. Yu, and F. Wei, “BEATs: Audio pre-training with acoustic tokeniz- ers,” inProc. ICML, 2023

    2023

  3. [11]

    BYOL for Audio: Exploring pre-trained general-purpose audio representations,

    D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, and K. Kashino, “BYOL for Audio: Exploring pre-trained general-purpose audio representations,”IEEE/ACM TASLP, vol. 31, pp. 137–151, 2022

    2022

  4. [12]

    Masked Modeling Duo: Towards a universal audio pre-training framework,

    D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, and K. Kashino, “Masked Modeling Duo: Towards a universal audio pre-training framework,”IEEE/ACM TASLP, vol. 32, pp. 2391–2406, 2024

    2024

  5. [13]

    SSAST: Self- supervised audio spectrogram transformer,

    Y . Gong, C.-I. Lai, Y .-A. Chung, and J. Glass, “SSAST: Self- supervised audio spectrogram transformer,” inProc. AAAI, vol. 36, 2022

    2022

  6. [14]

    Single channel target speaker extraction and recognition with speaker beam,

    M. Delcroix, K. Zmolikova, K. Kinoshita, A. Ogawa, and T. Nakatani, “Single channel target speaker extraction and recognition with speaker beam,” inProc. ICASSP, 2018

    2018

  7. [15]

    Separate anything you describe,

    X. Liu, Q. Kong, Y . Zhao, H. Liu, Y . Yuan, Y . Liu, R. Xia, Y . Wang, M. D. Plumbley, and W. Wang, “Separate anything you describe,” IEEE/ACM TASLP, vol. 33, pp. 458–471, 2024

    2024

  8. [16]

    Self-guided target sound extraction and classification through universal sound separa- tion model and multiple clues,

    Y . Kwon, D. Lee, D. Kim, and J.-W. Choi, “Self-guided target sound extraction and classification through universal sound separa- tion model and multiple clues,” inProc. DCASE, 2025

    2025

  9. [17]

    WavLM: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, et al., “WavLM: Large-scale self-supervised pre-training for full stack speech processing,”IEEE J-STSP, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  10. [18]

    Cocktail HuBERT: Generalized self-supervised pre-training for mixture and single-source speech,

    M. Fazel-Zarandi and W.-N. Hsu, “Cocktail HuBERT: Generalized self-supervised pre-training for mixture and single-source speech,” in Proc. ICASSP, 2023

    2023

  11. [19]

    An adapter based multi-label pre-training for speech separation and enhancement,

    T. Wang, X. Chen, Z. Chen, S. Yu, and W. Zhu, “An adapter based multi-label pre-training for speech separation and enhancement,” in Proc. ICASSP, 2023

    2023

  12. [20]

    Weakly-Supervised Speech Pre-training: A Case Study on Target Speech Recognition,

    W. Zhang and Y . Qian, “Weakly-Supervised Speech Pre-training: A Case Study on Target Speech Recognition,” inProc. Interspeech, 2023

    2023

  13. [21]

    Adapting self- supervised models to multi-talker speech recognition using speaker embeddings,

    Z. Huang, D. Raj, P. Garc ´ıa, and S. Khudanpur, “Adapting self- supervised models to multi-talker speech recognition using speaker embeddings,” inProc. ICASSP, 2023

    2023

  14. [22]

    SA-WavLM: Speaker-aware self-supervised pre-training for mixture speech,

    J. Lin, M. Ge, J. Ao, L. Deng, and H. Li, “SA-WavLM: Speaker-aware self-supervised pre-training for mixture speech,” inProc. Interspeech, 2024

    2024

  15. [23]

    Suppression of acoustic noise in speech using spectral sub- traction,

    S. Boll, “Suppression of acoustic noise in speech using spectral sub- traction,”IEEE TASSP, vol. 27, no. 2, pp. 113–120, 1979

    1979

  16. [24]

    Whisper-Flamingo: Integrating visual features into whisper for audio-visual speech recognition and translation,

    A. Rouditchenko, Y . Gong, S. Thomas, L. Karlinsky, H. Kuehne, R. Feris, and J. Glass, “Whisper-Flamingo: Integrating visual features into whisper for audio-visual speech recognition and translation,” in Proc. Interspeech, 2024

    2024

  17. [25]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Proc. NeurIPS, vol. 30, 2017

    2017

  18. [26]

    Root mean square layer normalization,

    B. Zhang and R. Sennrich, “Root mean square layer normalization,” inProc. NeurIPS, vol. 32, 2019

    2019

  19. [27]

    GLU variants improve transformer,

    N. Shazeer, “GLU variants improve transformer,”arXiv preprint arXiv:2002.05202, 2020

    2002 arXiv

  20. [28]

    FiLM: Visual reasoning with a general conditioning layer,

    E. Perez, F. Strub, H. De Vries, V . Dumoulin, and A. Courville, “FiLM: Visual reasoning with a general conditioning layer,” inProc. AAAI, vol. 32, 2018

    2018

  21. [29]

    SUNAC: Source-aware unified neural audio codec,

    R. Aihara, Y . Masuyama, F. Paissan, F. G. Germain, G. Wichern, and J. Le Roux, “SUNAC: Source-aware unified neural audio codec,” in Proc. ICASSP, 2026

    2026

  22. [30]

    Description and discussion on DCASE 2025 chal- lenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,

    T. Nishida et al., “Description and discussion on DCASE 2025 chal- lenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,” inProc. DCASE, 2025

    2025

  23. [31]

    FSD50K: An open dataset of human-labeled sound events,

    E. Fonseca, X. Favory, J. Pons, F. Font, and X. Serra, “FSD50K: An open dataset of human-labeled sound events,”IEEE/ACM TASLP, vol. 30, pp. 829–852, 2022

    2022

  24. [32]

    Audio set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmeke, D. P. W. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” inProc. ICASSP, 2017

    2017

  25. [33]

    WHAM!: Extending speech separation to noisy environments,

    G. Wichern, J. Antognini, M. Flynn, L. R. Zhu, E. McQuinn, D. Crow, E. Manilow, and J. Le Roux, “WHAM!: Extending speech separation to noisy environments,” inProc. Interspeech, 2019

    2019

  26. [34]

    The Diverse Environments Multi-channel Acoustic Noise Database (DEMAND): A database of multichannel environmental noise recordings,

    J. Thiemann, N. Ito, and E. Vincent, “The Diverse Environments Multi-channel Acoustic Noise Database (DEMAND): A database of multichannel environmental noise recordings,” inPOMA, vol. 19, 2013

    2013

  27. [35]

    The QUT-NOISE- TIMIT corpus for the evaluation of voice activity detection algo- rithms,

    D. Dean, S. Sridharan, R. V ogt, and M. Mason, “The QUT-NOISE- TIMIT corpus for the evaluation of voice activity detection algo- rithms,” inProc. Interspeech, 2010

    2010

  28. [36]

    A dataset and taxonomy for urban sound research,

    J. Salamon, C. Jacoby, and J. P. Bello, “A dataset and taxonomy for urban sound research,” inProc. ACM MM, 2014

    2014

  29. [37]

    Speech Commands: A dataset for limited-vocabulary speech recognition,

    P. Warden, “Speech Commands: A dataset for limited-vocabulary speech recognition,”arXiv preprint arXiv:1804.03209, 2018

    2018 arXiv

  30. [38]

    CREMA-D: Crowd-sourced emotional multimodal actors dataset,

    H. Cao, D. G. Cooper, M. K. Keutmann, R. C. Gur, A. Nenkova, and R. Verma, “CREMA-D: Crowd-sourced emotional multimodal actors dataset,”IEEE Trans. Affect. Comput., vol. 5, no. 4, pp. 377– 390, 2014

    2014

  31. [39]

    Neural audio synthesis of musical notes with wavenet autoencoders,

    J. Engel, C. Resnick, A. Roberts, S. Dieleman, M. Norouzi, D. Eck, and K. Simonyan, “Neural audio synthesis of musical notes with wavenet autoencoders,” inProc. ICML, 2017

    2017

  32. [40]

    One billion audio sounds from GPU-enabled modular synthesis,

    J. Turian, J. Shier, G. Tzanetakis, K. McNally, and M. Henry, “One billion audio sounds from GPU-enabled modular synthesis,” inProc. DAFx, 2021

    2021

  33. [41]

    The third ‘CHiME’ speech separation and recognition challenge: Dataset, task and baselines,

    J. Barker, R. Marxer, E. Vincent, and S. Watanabe, “The third ‘CHiME’ speech separation and recognition challenge: Dataset, task and baselines,” inProc. ASRU, 2015

    2015

  34. [42]

    Rafii, A

    Z. Rafii, A. Liutkus, F.-R. St ¨oter, S. I. Mimilakis, and R. Bittner,The MUSDB18 corpus for music separation, 2017,https : / / doi . org/10.5281/zenodo.1117372

    2017 doi

  35. [43]

    Visualizing data using t-SNE,

    L. Van der Maaten and G. Hinton, “Visualizing data using t-SNE,” JMLR, vol. 9, no. 11, 2008

    2008

  36. [44]

    ToyADMOS2: Another dataset of miniature-machine operat- ing sounds for anomalous sound detection under domain shift condi- tions,

    N. Harada, D. Niizumi, D. Takeuchi, Y . Ohishi, M. Yasuda, and S. Saito, “ToyADMOS2: Another dataset of miniature-machine operat- ing sounds for anomalous sound detection under domain shift condi- tions,” inProc. DCASE, 2021

    2021

  37. [45]

    MIMII DG: Sound dataset for mal- functioning industrial machine investigation and inspection for do- main generalization task,

    K. Dohi, T. Nishida, H. Purohit, R. Tanabe, T. Endo, M. Yamamoto, Y . Nikaido, and Y . Kawaguchi, “MIMII DG: Sound dataset for mal- functioning industrial machine investigation and inspection for do- main generalization task,” inProc. DCASE, 2022

    2022

  38. [46]

    Toy- ADMOS2025: The evaluation dataset for the DCASE2025T2 first- shot unsupervised anomalous sound detection for machine condition monitoring,

    N. Harada, D. Niizumi, Y . Ohishi, D. Takeuchi, and M. Yasuda, “Toy- ADMOS2025: The evaluation dataset for the DCASE2025T2 first- shot unsupervised anomalous sound detection for machine condition monitoring,” inProc. DCASE, 2025

    2025

  39. [47]

    Deep generic representations for domain-generalized anomalous sound detection,

    P. Saengthong and T. Shinozaki, “Deep generic representations for domain-generalized anomalous sound detection,” inProc. ICASSP, 2025

    2025

  40. [48]

    SMOTE: Synthetic minority over-sampling technique,

    N. V . Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: Synthetic minority over-sampling technique,”JAIR, vol. 16, pp. 321–357, 2002

    2002

  41. [49]

    ASDKit: A toolkit for comprehensive evaluation of anomalous sound detection methods,

    T. Fujimura, K. Wilkinghoff, K. Imoto, and T. Toda, “ASDKit: A toolkit for comprehensive evaluation of anomalous sound detection methods,” inProc. DCASE, 2025

    2025

  42. [50]

    Adaptive prototype learning for anomalous sound detection with partially known attributes,

    A. Jiang, X. Zheng, B. Han, Y . Qiu, P. Fan, W.-Q. Zhang, C. Lu, and J. Liu, “Adaptive prototype learning for anomalous sound detection with partially known attributes,” inProc. ICASSP, 2025

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.