Enforcing Speech Content Privacy in Environmental Sound Recordings using Segment-wise Waveform Reversal

Mathieu Lagrange; Modan Tailleur; Pierre Aumond; Vincent Tourre

REVIEW 3 major objections 5 minor 25 references

Enforcing Speech Content Privacy in Environmental Sound Recordings using Segment-wise Waveform Reversal

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A deterministic method that reverses waveform fragments only within speech regions makes spoken content 97.9% unintelligible in environmental recordings while leaving the acoustic scene and perceptual quality nearly intact.

desk verdict The intelligibility result is strong and the method is simple; the perceptual-quality and scene-preservation claims are weaker than advertised. read the letter →

arxiv 2507.08412 v2 pith:W3CV6EGJ submitted 2025-07-11 cs.SD eess.AS

Modan Tailleur , Mathieu Lagrange , Pierre Aumond , Vincent Tourre This is my paper · ORCID

classification cs.SDeess.AS

keywords speechcontentprivacyenvironmentalaudiowaveformreversalvoiceactivitydetectionsourceseparationworderrorrateFréchetdistanceurbansoundrecordings

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

This paper argues that a deterministic signal-processing recipe can strip intelligible speech out of environmental recordings without damaging the acoustic scene or the audio quality that analysts need. The recipe detects speech regions with a voice-activity classifier, separates the speech track with a source-separation model, and reverses each waveform fragment inside that track so words become backwards and unrecognizable. On a simulated dataset mixing LibriSpeech speech with urban sounds from SONYC-UST, the authors report a word error rate of 97.9%, a drop in sound-source detection accuracy of only 2.7%, and a Fréchet Audio Distance of 1.40, closing the gap between the best previously published method and the white-noise lower bound. The practical hope is that field recordings can be shared, annotated, and reused internally without exposing the words people said.

What carries the argument

The engine of the method is Segment-wise Waveform Reversal (SWR): the speech-only track is fragmented by cutting at low-energy regions (with a 6 dBFS threshold on 2-second texture frames), and each fragment's samples are played backward, with a 5% overlap-add at segment and frame boundaries to suppress clicks. This reversal is applied only to the speech channel after a BEATs-based voice activity detector selects speech windows and a Hybrid Demucs source-separation model isolates the speech from the background. The optional random-splicing variant reuses the same low-energy fragmentation but shuffles the reversed segments, making the transformation stochastic and therefore harder to invert by reapplying the pipeline.

What would settle it

Play the released audio samples to a panel of listeners and ask them to rate artifacts and overall quality against the unprocessed mixtures; if the PANN-Wavegram-Logmel-based FAD ordering (1.40 vs 1.54) does not match human ordering, the 'high perceptual quality' claim loses its support.

Watch

Extended reading notes

Core claim

The central claim is that segment-wise waveform reversal, applied only to speech that has been isolated by voice activity detection and source separation, simultaneously solves three otherwise conflicting objectives: it makes the spoken content unintelligible to automatic speech recognizers (averaged WER 97.9%, close to the 98.5% of pure white noise), it leaves the background scene almost untouched (a 2.7% drop in sound-source classification accuracy, tied with the best baseline), and it yields the lowest Fréchet Audio Distance (1.40) among all tested methods, indicating the best perceptual quality. The method is deterministic, requires no training, and can be tuned by adding a stochastic segment-reordering step (random splicing) that lifts adversarial robustness from 54.9% to 93.2% WER under a re-reversal attack, at the cost of a modest FAD increase to 1.54.

Load-bearing premise

The quality claim rests on the assumption that the Fréchet Audio Distance computed with PANN-Wavegram-Logmel embeddings tracks how human listeners perceive the artifacts of segment-wise reversal in environmental mixtures.

Editorial extensions

If this is right

Field-recording archives can be processed once and then shared internally for annotation, listening-based analysis, and quality control without exposing the words spoken by recorded individuals.
The averaged WER of 97.9% is nearly as strong as replacing the whole audio with white noise (98.5%), so ASR-based extraction of spoken content is effectively blocked.
The preservation of the background scene is near-total: source detection accuracy drops by only 2.7%, matching the fragmentation-only baseline of Burkhardt et al.
Adding random segment reordering changes the privacy guarantee from deterministic to stochastic, raising adversarial WER under a second-pass re-reversal attack from 54.9% to 93.2%, with FAD rising from 1.40 to 1.54.
The ablation shows that voice activity detection contributes most to preserving non-speech content: removing it raises SCAD by roughly 1.5%, while removing source separation mainly raises FAD.

Reading between the lines

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

We infer that the method's practical reach depends on the threat model: it is designed against casual or accidental intelligibility, not against a determined attacker who knows the segmentation algorithm, as the 54.9% Adv-WER demonstrates; the reordering variant only helps if the random seed stays secret.
We infer that the quality metric may be the weak link: PANN-Wavegram-Logmel embeddings were validated for environmental audio synthesis, not for time-reversal discontinuities, so a human listening study on exactly this transformation is the natural next test.
We infer that the method could be extended to continuous streaming field audio without training, because all stages (VAD, separation, reversal) operate on fixed windows; the main unknown is how often speech windows are misclassified under real-world reverberation.
We infer that the authors' deliberate choice not to mask speaker identity means the method is best viewed as a content-privacy layer, complementary to speaker-identity anonymizers rather than a replacement.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Basically, this is a useful, honest paper. The method is simple: detect speech with BEATs, separate with HDemucs, fragment low-energy regions, reverse each segment, mix back. The new bit is reversal rather than reordering, and it works: WER 97.9% across four ASR models, essentially at the white-noise ceiling, while SCAD drops only 2.7% and FAD is 1.40, better than the references. The ablation is sensible, the robustness section acknowledges the determinism weakness and tests a reordering variant, and the code and data are shared. That is real work, and the main intelligibility claim is solid.

The soft spots are in the other two pillars. FAD is the weakest. The paper calls FAD 1.40 'high perceptual quality,' but the embedding choice (PANN-Wavegram-Logmel) is justified by the authors' own EUSIPCO study on environmental audio synthesis, not on privacy transformations. Time-reversed speech preserves spectral content and introduces discontinuities that PANN may not penalize. And in Section VI they report that increasing FAD from 1.40 to 1.54 produces a 'clearly perceptible' effect on informal listening, which tells you the FAD scale is not well calibrated for these artifacts. So the perceptual-quality claim is undersupported; a small human listening test or at least a bootstrap on FAD would fix it.

The SCAD metric is entangled with the method: the same BEATs model is used for VAD in the pipeline and for the detectability evaluation. It is not strictly circular because the VAD threshold and the ranking use different outputs, but it is a shared representation that could favor the method. The paper does not address this. Also, the evaluation is on synthetic single-source mixtures; real-world field recordings are messier. That is a generalization caveat, not a fatal flaw.

The robustness attack is a bit weak: re-applying the pipeline is not the most obvious attack for a deterministic reversal method. An attacker who knows the low-energy segmentation could just reverse each segment back. The paper's own VAD and separation errors protect against this, but they do not analyze that reconstruction, only a second pass. Worth a sentence or two in revision.

Overall: the intelligibility result is credible, the method is simple and reproducible, and the paper would benefit from a more careful treatment of the quality metric and a less circular scene-preservation evaluation. I would send it to review; with revisions it could be a solid contribution to the audio privacy literature.

Referee Report

3 major / 5 minor

Summary. The paper proposes a deterministic speech-content privacy method for environmental sound recordings. The pipeline first detects speech with a BEATs-based voice activity detector, separates speech from background with HDemucs, applies segment-wise waveform reversal (SWR) to the speech component, and remixes the modified speech with the background. The authors evaluate on simulated mixtures of LibriSpeech speech and SONYC-UST urban sound scenes, reporting a Word Error Rate of 97.9%, a Sound source Classification Accuracy-Drop of 2.7%, and a Fréchet Audio Distance of 1.40, together with an ablation study and a robustness analysis against re-applying the pipeline as an adversarial attack. The manuscript claims that the method renders speech unintelligible while preserving the acoustic scene and overall audio quality, and that adding random splicing improves robustness at a small quality cost.

Significance. If the claims are validated, this is a practical, simple, and computationally lightweight contribution to an important privacy problem: sharing field recordings without exposing intelligible speech. The paper's strengths include a reproducible dataset and code release, evaluation with four external ASR models, comparison against three reference methods, an ablation study, and a first attempt at modeling an adversarial recovery attack. The central intelligibility result is well supported by external ASR models. However, the two non-speech evaluation pillars are less secure: the perceptual-quality claim rests on a single FAD embedding with no uncertainty quantification or human validation, and the source-preservation metric uses the same model (BEATs) that drives the VAD stage, creating a potential entanglement. These gaps affect the advertised three-part claim and should be addressed before publication.

major comments (3)

[§IV-B, Eq. (1), Table 2a] The claim of 'high perceptual quality' is load-bearing for the paper's central message, but it is supported only by a single FAD value (1.40) computed with PANN-Wavegram-Logmel embeddings. The embedding choice is justified by the authors' own prior study [18], which correlated FAD with human perception for environmental audio synthesis, not for segment-wise waveform reversal. Time-reversed speech segments introduce fine temporal and phase discontinuities that may not be captured by an embedding dominated by spectral content, so the FAD could be insensitive to exactly the artifacts the method creates. The paper reports no confidence intervals, no per-item variability, and no listening test for the main result. To support the quality claim, please provide bootstrap CIs or multiple embeddings, and ideally a small perceptual rating experiment.
[§III-A vs. §IV-B] The sound-source detectability metric (SCAD) uses BEATs, which is also the model that performs voice activity detection in the privacy pipeline. Because the same model decides where processing is applied and then evaluates whether sources remain detectable, the SCAD estimate may be favorably biased: VAD errors and classifier errors are not independent. This is especially relevant for the claim of 'minimal degradation of the sound sources detectability'. I recommend recomputing SCAD with an independent source classifier (e.g., a SONYC-UST-trained tagger or a different pre-trained audio tagger) or, at minimum, providing a clear argument why the shared representation does not inflate the preservation result.
[§VI] The adversarial-robustness evaluation considers only the attack of re-applying the privacy pipeline to an already-processed signal. Because the segmentation procedure is deterministic and based on RMS energy, an attacker who knows the algorithm can compute the segment boundaries directly and invert each reversed segment without needing VAD or source separation. The reported Adv-WER of 54.9% may therefore overestimate robustness against a simple, algorithm-aware adversary. Please specify the threat model more precisely and test a direct inversion attack, since the robustness claim is an explicit part of the paper's contribution.

minor comments (5)

[§V and Fig. 2] The text refers to 'Table 2a' and 'Table 2b', but the results are presented in Figure 2; please make the cross-references consistent.
[§V and Fig. 2] The table header uses 'WER (%) (↑)' and 'SCAD (%) (↓)' with arrows that may confuse readers; please clarify in the caption that higher WER indicates better intelligibility reduction and lower SCAD indicates better scene preservation.
[Table 2a] The row 'Ours (Sp.Brain SepFormer)' reports a lower SCAD (1.0%) than the HDemucs-based method (2.7%), but the text does not discuss this difference; a sentence explaining it would be helpful.
[§IV-B] The FAD reference is described as 'our reference dataset that contains unaltered speech'; please clarify whether the reference distribution is the original mixtures or a separate set, since this affects interpretation of the FAD values.
[Abstract] The abstract contains the typo 'effectivness'; please proofread the manuscript for minor spelling and grammar issues.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the WER and SCAD claims are measured with external pretrained models, and the FAD value is computed from real embeddings; the FAD embedding choice relies in part on a self-cited perceptual-correlation study, but that study is externally falsifiable and is not the source of the reported numbers.

full rationale

The paper's central claims are empirical measurements rather than derivations. The intelligibility claim (WER 97.9%) is obtained from four external ASR systems (Wav2Vec2, Whisper large-v3, Fairseq S2T, CRDNN), none of which are fitted or defined by the proposed SWR pipeline. The source-detectability claim (SCAD 2.7%) uses the pretrained BEATs model, which is also used for VAD in Section III-A; this creates a mild evaluator-model overlap, but BEATs is a fixed public pretrained model and no parameter is fitted to the evaluation data, so the SCAD value is not forced by construction. The audio-quality claim (FAD 1.40) is computed from actual PANN-Wavegram-Logmel embeddings of the original and processed datasets. The choice of this embedding follows the authors' own prior study [18], a self-citation; however, [18] is an externally falsifiable perceptual-correlation study, and the paper also cites independent FAD-perception work [20], so the self-citation is not load-bearing in a way that would reduce the result to an input. The remaining risk is external validity: transferring [18] from environmental-audio synthesis to segment-wise time-reversal artifacts is untested, and the perceptual significance of FAD 1.40 is not independently validated for this transformation. That is a correctness or validity concern, not circularity. No equation or fitted parameter in the paper reduces any reported metric to its own input, and the ablation study varies pipeline components rather than reusing the evaluation metric as a fitting target. Therefore no significant circularity is found.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The method inherits several pre-trained models as unverified inputs (BEATs, HDemucs, PANN, ASR systems) and introduces hand-chosen thresholds and block sizes (VAD 0.3, 6 dBFS, 2 s frame, 5% overlap). No new entities are postulated. The dataset construction uses a 6 dB background boost and RMSE normalization, which set the evaluation regime. The reliance on the authors' own FAD-embedding study [18] is a mild self-referential element.

free parameters (5)

VAD speech-class threshold = 0.3
Threshold on BEATs speech logits (Section III.A); chosen for inclusive speech detection, affects which segments are reversed.
Low-energy fragmentation threshold = 6 dBFS
Threshold for identifying quiet regions for segment boundaries (Section III.C); chosen to minimize audible artifacts.
Texture frame size = 2 s
Length of analysis frames for fragmentation (Section III.C); chosen for the segmentation algorithm.
Overlap-add ratio = 5%
Overlap applied at segment and frame boundaries (Section III.C) to reduce discontinuities; hand-chosen.
Background-to-speech gain in evaluation dataset = 6 dB
SONYC-UST signal is boosted by 6 dB before mixing (Section IV.A) to simulate realistic noise; this affects the evaluation conditions.

assumptions (6)

domain assumption BEATs pre-trained model reliably identifies speech segments for the VAD.
The pipeline relies on BEATs speech-class logits with a 0.3 threshold (Section III.A) to locate speech; if this detector fails, speech may remain intelligible.
domain assumption HDemucs, trained on music data (MUSDB-HQ), can separate speech from environmental background sound at 44.1 kHz.
The method uses HDemucs weights from Torchaudio (Section III.B) to isolate speech; its effectiveness on urban environmental audio is assumed by transfer.
domain assumption FAD with PANN-Wavegram-Logmel embeddings correlates with human perception of this type of processed audio.
The audio quality metric is justified by [18], a self-cited study on generative environmental audio, not on privacy transformations (Section IV.B).
domain assumption Word Error Rate from four ASR models is a valid proxy for speech intelligibility.
WER measures ASR transcription, not human perception; this is standard in privacy research, but it is an assumption about the metric.
domain assumption The original unaltered mixtures are the correct reference set for FAD.
FAD is computed against a reference dataset containing unaltered speech (abstract); the paper does not discuss whether using a clean non-speech reference would change conclusions.
ad hoc to paper Reapplying the privacy pipeline is a representative adversarial recovery attack.
The robustness test in Section VI simulates an attacker who runs the same pipeline a second time; a more informed attacker could try to invert segmentation and overlap-add operations directly.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Enforcing Speech Content Privacy in Environmental Sound Recordings using Segment-wise Waveform Reversal." pith.science (2026). https://pith.science/paper/W3CV6EGJ

@misc{pith2026250708412,
  author       = {Pith},
  title        = {Pith review of: Enforcing Speech Content Privacy in Environmental Sound Recordings using Segment-wise Waveform Reversal},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3CV6EGJ}},
  note         = {Machine review of arXiv:2507.08412}
}

read the original abstract

Environmental sound recordings often contain intelligible speech, raising privacy concerns that limit analysis, sharing and reuse of data. In this paper, we introduce a method that renders speech unintelligible while preserving both the integrity of the acoustic scene, and the overall audio quality. Our approach involves reversing waveform segments to distort speech content. This process is enhanced through a voice activity detection and speech separation pipeline, which allows for more precise targeting of speech. In order to demonstrate the effectivness of the proposed approach, we consider a three-part evaluation protocol that assesses: 1) speech intelligibility using Word Error Rate (WER), 2) sound sources detectability using Sound source Classification Accuracy-Drop (SCAD) from a widely used pre-trained model, and 3) audio quality using the Fr\'echet Audio Distance (FAD), computed with our reference dataset that contains unaltered speech. Experiments on this simulated evaluation dataset, which consists of linear mixtures of speech and environmental sound scenes, show that our method achieves satisfactory speech intelligibility reduction (97.9% WER), minimal degradation of the sound sources detectability (2.7% SCAD), and high perceptual quality (FAD of 1.40). An ablation study further highlights the contribution of each component of the pipeline. We also show that incorporating random splicing to our speech content privacy enforcement method can enhance the algorithm's robustness to attempt to recover the clean speech, at a slight cost of audio quality.

Figures

Figures reproduced from arXiv: 2507.08412 by the authors.

**Figure 2.** Evaluation of speech content privacy enforcement methods. Results not statistically different from the best-performing [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 22 canonical work pages

[18]

Correlation of fr ´echet audio distance with human per- ception of environmental audio is embedding dependent,

M. Tailleur, J. Lee, M. Lagrange, K. Choi, L. M. Heller, K. Imoto, and Y . Okamoto, “Correlation of fr ´echet audio distance with human per- ception of environmental audio is embedding dependent,” in EUSIPCO. IEEE, 2024

work page 2024
[1]

SONYC: a system for monitoring, analyzing, and mitigating urban noise pollution,

J. P. Bello, C. Silva, O. Nov, R. L. Dubois, A. Arora, J. Salamon, C. Mydlarz, and H. Doraiswamy, “SONYC: a system for monitoring, analyzing, and mitigating urban noise pollution,” CACM, 2019

work page 2019
[2]

An Innovative Approach for long term environmental noise measurement: RUMEUR Network in the Paris Region,

F. Mietlicki, C. Mietlicki, and M. Sineau, “An Innovative Approach for long term environmental noise measurement: RUMEUR Network in the Paris Region,” in EuroNoise, 2015

work page 2015
[3]

A strongly-labelled polyphonic dataset of urban sounds with spatiotem- poral context,

K. Ooi, K. N. Watcharasupat, S. Peksi, F. A. Karnapi, Z.-T. Ong, D. Chua, H.-W. Leow, L.-L. Kwok, X.-L. Ng, and Z.-A. Loh, “A strongly-labelled polyphonic dataset of urban sounds with spatiotem- poral context,” in APSIPA ASC. IEEE, 2021

work page 2021
[4]

Librispeech: an asr corpus based on public domain audio books,

V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: an asr corpus based on public domain audio books,” in ICASSP. IEEE, 2015

work page 2015
[5]

SONYC-UST-V2: An Urban Sound Tagging Dataset with Spatiotemporal Context,

M. Cartwright, J. Cramer, A. E. M. Mendez, Y . Wang, H.-H. Wu, V . Lostanlen, M. Fuentes, G. Dove, C. Mydlarz, J. Salamon, O. Nov, and J. P. Bello, “SONYC-UST-V2: An Urban Sound Tagging Dataset with Spatiotemporal Context,” 2020, arXiv:2009.05188 [cs]

arXiv 2020
[6]

Masking speech contents by random splicing: is emotional expression preserved?

F. Burkhardt, A. Derington, M. Kahlau, K. Scherer, F. Eyben, and B. Schuller, “Masking speech contents by random splicing: is emotional expression preserved?” in ICASSP. IEEE, 2023

work page 2023
[7]

V oice anonymization in urban sound recordings,

A. Cohen-Hadria, M. Cartwright, B. McFee, and J. P. Bello, “V oice anonymization in urban sound recordings,” in MLSP. IEEE, 2019

work page 2019

Show all 25 references

[8]

BEATs: Audio pre-training with acoustic tokenizers,

S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chen, and F. Wei, “BEATs: Audio pre-training with acoustic tokenizers,” in ICML, 2023

2023
[9]

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

J. F. Gemmeke, D. P. 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,” in ICASSP. IEEE, 2017

2017
[10]

Hybrid spectrogram and waveform source separation,

A. D ´efossez, “Hybrid spectrogram and waveform source separation,” in MDX Workshop, ISMIR , 2021

2021
[11]

MUSDB18-HQ-an uncompressed version of MUSDB18,

Z. Rafii, A. Liutkus, F.-R. St ¨oter, S. I. Mimilakis, and R. Bittner, “MUSDB18-HQ-an uncompressed version of MUSDB18,” 2019

2019
[12]

Librivox: Free public domain audiobooks,

J. Kearns, “Librivox: Free public domain audiobooks,” Reference Re- views, 2014

2014
[13]

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,” NeurIPS, 2020

2020
[14]

Robust speech recognition via large-scale weak supervi- sion,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” in International conference on machine learning . PMLR, 2023

2023
[15]

Fairseq S2T: Fast Speech-to-Text Modeling with Fairseq,

C. Wang, Y . Tang, X. Ma, A. Wu, D. Okhonko, and J. Pino, “Fairseq S2T: Fast Speech-to-Text Modeling with Fairseq,” in AACL-IJCNLP, 2020

2020
[16]

SpeechBrain: A General- Purpose Speech Toolkit,

M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell, L. Lu- gosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong, J.-C. Chou, S.-L. Yeh, S.-W. Fu, C.-F. Liao, E. Rastorgueva, F. Grondin, W. Aris, H. Na, Y . Gao, R. De Mori, and Y . Bengio, “SpeechBrain: A General- P...

2021 arXiv
[17]

Fr ´echet audio distance: A reference-free metric for evaluating music enhancement algorithms,

K. Kilgour, M. Zuluaga, D. Roblek, and M. Sharifi, “Fr ´echet audio distance: A reference-free metric for evaluating music enhancement algorithms,” in INTERSPEECH, 2019

2019
[19]

Foley Sound Synthesis at the DCASE 2023 Challenge,

K. Choi, J. Im, L. Heller, B. McFee, K. Imoto, Y . Okamoto, M. La- grange, and S. Takamichi, “Foley Sound Synthesis at the DCASE 2023 Challenge,” 2023, arXiv:2304.12521 [cs]

2023 arXiv
[20]

Adapting frechet audio distance for generative music evaluation,

A. Gui, H. Gamper, S. Braun, and D. Emmanouilidou, “Adapting frechet audio distance for generative music evaluation,” in ICASSP. IEEE, 2024

2024
[21]

Gans trained by a two time-scale update rule converge to a local nash equilibrium,

M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” NeurIPS, 2017

2017
[22]

Panns: Large-scale pretrained audio neural networks for audio pattern recognition,

Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumbley, “Panns: Large-scale pretrained audio neural networks for audio pattern recognition,” IEEE/ACM TASLP, 2020

2020
[23]

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, and B. Seybold, “CNN architectures for large-scale audio classification,” in ICASSP. IEEE, 2017

2017
[24]

Attention is all you need in speech separation,

C. Subakan, M. Ravanelli, S. Cornell, M. Bronzi, and J. Zhong, “Attention is all you need in speech separation,” in ICASSP. IEEE, 2021

2021
[25]

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,” Interspeech, 2019

2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [18]

Correlation of fr ´echet audio distance with human per- ception of environmental audio is embedding dependent,

M. Tailleur, J. Lee, M. Lagrange, K. Choi, L. M. Heller, K. Imoto, and Y . Okamoto, “Correlation of fr ´echet audio distance with human per- ception of environmental audio is embedding dependent,” in EUSIPCO. IEEE, 2024

work page 2024

[2] [1]

SONYC: a system for monitoring, analyzing, and mitigating urban noise pollution,

J. P. Bello, C. Silva, O. Nov, R. L. Dubois, A. Arora, J. Salamon, C. Mydlarz, and H. Doraiswamy, “SONYC: a system for monitoring, analyzing, and mitigating urban noise pollution,” CACM, 2019

work page 2019

[3] [2]

An Innovative Approach for long term environmental noise measurement: RUMEUR Network in the Paris Region,

F. Mietlicki, C. Mietlicki, and M. Sineau, “An Innovative Approach for long term environmental noise measurement: RUMEUR Network in the Paris Region,” in EuroNoise, 2015

work page 2015

[4] [3]

A strongly-labelled polyphonic dataset of urban sounds with spatiotem- poral context,

K. Ooi, K. N. Watcharasupat, S. Peksi, F. A. Karnapi, Z.-T. Ong, D. Chua, H.-W. Leow, L.-L. Kwok, X.-L. Ng, and Z.-A. Loh, “A strongly-labelled polyphonic dataset of urban sounds with spatiotem- poral context,” in APSIPA ASC. IEEE, 2021

work page 2021

[5] [4]

Librispeech: an asr corpus based on public domain audio books,

V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: an asr corpus based on public domain audio books,” in ICASSP. IEEE, 2015

work page 2015

[6] [5]

SONYC-UST-V2: An Urban Sound Tagging Dataset with Spatiotemporal Context,

M. Cartwright, J. Cramer, A. E. M. Mendez, Y . Wang, H.-H. Wu, V . Lostanlen, M. Fuentes, G. Dove, C. Mydlarz, J. Salamon, O. Nov, and J. P. Bello, “SONYC-UST-V2: An Urban Sound Tagging Dataset with Spatiotemporal Context,” 2020, arXiv:2009.05188 [cs]

arXiv 2020

[7] [6]

Masking speech contents by random splicing: is emotional expression preserved?

F. Burkhardt, A. Derington, M. Kahlau, K. Scherer, F. Eyben, and B. Schuller, “Masking speech contents by random splicing: is emotional expression preserved?” in ICASSP. IEEE, 2023

work page 2023

[8] [7]

V oice anonymization in urban sound recordings,

A. Cohen-Hadria, M. Cartwright, B. McFee, and J. P. Bello, “V oice anonymization in urban sound recordings,” in MLSP. IEEE, 2019

work page 2019