Pith. sign in

REVIEW 3 major objections 6 minor 33 references

A tuned dynamic-range compressor, causal and low-latency, is the most effective tested method for softening sounds that trigger neurodivergent distress, outperforming equalization and a neural auto-encoder in a 180-person listening test.

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

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

Among DSP and ML audio enhancement approaches evaluated on trigger-sound mixtures, Dynamic Range Compression (DRC) attenuates distressing sounds most effectively in both objective metrics and a neurodivergent listening test.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection DRC lowers self-reported triggerability in a purpose-built listening test, but the 'selective' claim is unverified without a volume-matched baseline and neutral-preservation measurement. the 3 major comments →

arxiv 2509.10202 v1 pith:KJV6GEVK submitted 2025-09-12 eess.AS

Low-latency Assistive Audio Enhancement for Neurodivergent People

Alexander Popescu , Rosie Frost , Milos Cernak This is my paper
classification eess.AS
keywords assistive audiodynamic range compressiondecreased sound toleranceneurodivergencetrigger soundsselective transparency modeaudio enhancementlistening test
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper asks whether a headphone transparency mode can selectively soften sounds that trigger distress in neurodivergent people without muting the rest of the world. It builds a trigger-sound library from online community reports, maps those reports to public sound-event datasets, and benchmarks five DSP methods plus a neural auto-encoder on synthetic mixtures of trigger, neutral, and background sounds. The central result is that a causal Dynamic Range Compressor, with parameters tuned on a validation set (threshold −35 dB, ratio 30:1, attack 0.01 ms, release 100 ms), attenuates trigger sounds and lowers self-reported triggerability from 66.71 to 38.22 for neurodivergent listeners, and from 63.09 to 33.13 for controls. The paper positions DRC as a low-latency building block for selective transparency in ANC headphones, with the neural model as a viable but more distortion-prone alternative.

Core claim

On its own terms, the paper's discovery is that a well-tuned compressor—not a learned separator—is the best low-latency assistive filter for trigger sounds. In a listening test with 133 neurodivergent and 47 neurotypical participants, the DRC-processed mixtures received the lowest triggerability ratings on a 0–100 scale for nearly every one of the ten trigger categories, and the overall mean dropped from 66.71 (unprocessed) to 38.22 for neurodivergent listeners. The auto-encoder and equalization also helped but trailed DRC, and listeners penalized the neural network for audible distortions. The paper attributes DRC's edge to its ability to catch transient, percussive triggers quickly (0.01 m

What carries the argument

Dynamic Range Compression (DRC): a real-time, causal audio effect that reduces the gain of a signal when its level exceeds a threshold, with a fast attack for transient peaks and a slower release to avoid pumping. Here it is configured at threshold −35 dB, ratio 30:1, attack 0.01 ms, and release 100 ms, and it is the algorithmic workhorse: it carries the paper's main result because it is simple, causal, and low-latency enough for a selective transparency mode on top of active noise cancellation. The neural baseline is an auto-encoder with causal dilated convolutions in the encoder and self/cross-attention in the decoder, trained with a negative SI-SNR objective; the comparison shows what a l

Load-bearing premise

The whole result rests on treating a 5-second synthetic mixture, with one repeated trigger and a 0–100 rating, as a faithful proxy for the real-world distress that triggers cause; if that proxy does not transfer, the measured reduction in triggerability may not mean much in daily life.

What would settle it

A reader could falsify the central claim by running the same DRC parameters on recordings of real kitchens or offices with a trigger sound in the background and a speech foreground, and measuring whether listeners still rate the processed audio as less triggering than unprocessed audio. If self-reported triggerability does not drop, or if speech intelligibility collapses, the paper's selective-transparency premise fails. A simpler check: the paper's own listening test has no condition where trigger and neutral sounds vary in SNR; adding a low-SNR trigger condition would show whether DRC's adva

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

If this is right

  • A selective transparency mode built on DRC could run inside low-latency ANC headphones, since the algorithm is causal and requires no trained model inference.
  • DRC's benefit is tied to transient, foreground triggers; the paper notes its performance depends on the SNR of the trigger relative to other sounds, so background or sustained triggers are less cleanly attenuated.
  • The neural auto-encoder, while competitive objectively, introduces audible distortions that reduce its subjective benefit; improving training data and architecture could close the gap.
  • The community-derived trigger list and released dataset give other researchers a standardized test bed for assistive audio algorithms targeting decreased sound tolerance.
  • Equalization and automatic gain control are not sufficient alone: EQ had modest effects, and AGC tended to silence the mixture rather than selectively suppress triggers.

Where Pith is reading between the lines

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

  • Editorial inference: the same DRC parameters may transfer poorly to real-world acoustic scenes, where trigger sounds arrive at varying SNRs and overlap with speech; a listening test with naturalistic audio and clinically assessed participants would test this directly.
  • Editorial inference: because DRC is blind to sound identity, it will also attenuate non-triggering loud transients (e.g., a door slam during conversation), so a hybrid system that uses a lightweight classifier to modulate the compressor's threshold could preserve more neutral content.
  • Editorial inference: the 0–100 triggerability scale in a forced web-based comparison may inflate the apparent benefit; pairing with physiological measures (heart rate, skin conductance) or real-world ecological momentary assessment would triangulate the self-report.
  • Editorial inference: the dataset construction pipeline—community text mining plus LLM labeling—could be extended to other sensory triggers (visual, tactile) or to personalized trigger profiles per listener, which the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes low-latency assistive audio enhancement for selective transparency in hearing devices for neurodivergent people. It constructs a trigger-sound list from Reddit via LLM-based extraction, maps it to AudioSet labels, compiles trigger/neutral sound sets, creates synthetic 10-second mixtures, and evaluates five DSP algorithms (DRC, EQ, AGC, MCTR, LPF) and an auto-encoder NN using ΔSI-SNR. A listening test with 133 neurodivergent and 47 control participants rates triggerability of processed stimuli. The central claim is that DRC, with optimized parameters (threshold -35 dB, ratio 30:1, attack 0.01 ms, release 100 ms), is the most effective method, lowering mean triggerability from 66.71 (unprocessed) to 38.22 (anc-drc). The paper acknowledges several limitations, including lack of neutral-sound recognition tests and SNR dependence.

Significance. The topic is timely and practically relevant, and the dataset-construction pipeline plus the large neurodivergent listening panel (N=133, with a neurotypical control group) are substantive strengths. The paper also applies multiple baseline algorithms and statistical corrections. If the selective-transparency claim were fully supported, the result would be a useful step toward hearable/ANC assistive features. However, the current evidence does not distinguish selective attenuation of trigger sounds from global attenuation, the objective metric is scale-invariant, and the subjective test lacks the necessary volume-matched and neutral-preservation controls. The planned data release after the conference decision is positive but not yet verifiable.

major comments (3)
  1. [Sec. 3.2 / Table 1] The listening test compares DRC, NN1, EQ, and unprocessed mixtures but does not include a loudness-matched or static-gain/peak-limiter control. The optimized DRC settings (threshold -35 dB, ratio 30:1, attack 0.01 ms; Sec. 2.2.1) act essentially as a hard limiter on transients, and Sec. 3.1 concedes that DRC and NN 'often affect the rest of the mixture.' Without loudness normalization or a gain-only condition, the large reduction in triggerability (38.22 vs 66.71) could be explained by global attenuation rather than selective suppression of trigger sounds. This is load-bearing for the paper's 'selective transparency' claim; the paper itself defers neutral-sound recognition tests to Sec. 4. Please add a volume-matched control and/or a neutral-preservation rating.
  2. [Sec. 3.1 / Fig. 2, Sec. 2.2.2] The objective evaluation uses ΔSI-SNR, which is invariant to global scaling (a property the paper notes for SI-SNR in Sec. 2.2.2). Therefore global level reduction alone is invisible to the metric, so the positive ΔSI-SNR values cannot by themselves establish that DRC's benefit is due to selective attenuation. Moreover, DRC/EQ/AGC parameters were optimized on the validation set to maximize SI-SNR (Sec. 2.2.1), and the objective ranking in Fig. 2 uses the same metric. The objective results are thus not independent evidence for the selective-enhancement claim. Please report component-wise or loudness-normalized metrics (e.g., SI-SNR/STOI of the neutral source) so that transparency can be assessed.
  3. [Sec. 3.1-3.2, Test Set 3] Test Set 3 is used both for objective evaluation and for selecting which algorithms enter the subjective test: Sec. 3.2 states that DRC, NN1, and EQ were included because they 'had demonstrated strong objective performance' on Test Set 3. The subjective comparison is therefore conditional on performance on the same 10 stimuli, and the claim that DRC is 'the most effective' among the candidate systems is not supported by a clean selection/evaluation split. Please report subjective results for all candidate systems or use a disjoint set for algorithm selection and final evaluation.
minor comments (6)
  1. [Abstract/Introduction] Missing period after 'enhancement technologies' before 'In this paper'.
  2. [Sec. 2.2.1] The EQ description '-2.75 dB Hz (high-shelf)' appears to omit the center frequency; please clarify.
  3. [Table 1] The first condition is labeled 'mix'; it would be clearer to label it 'Unprocessed'. The ground-truth condition shown to listeners is also not reported in the table, though it would provide a useful anchor.
  4. [Sec. 3.2] The subjective study involves 180 human listeners but no ethics approval or informed-consent statement is reported. Please add this information.
  5. [Sec. 4] The data-release statement ('open after conference decision') lacks a concrete repository, license, or timeline; this is important for reproducibility claims.
  6. [Fig. 2] The reference signal used to compute ΔSI-SNR should be stated explicitly in the caption or text; the current description ('ground truth') is ambiguous between the trigger-free mixture and the original target component.

Circularity Check

0 steps flagged

No significant circularity: DRC's efficacy is supported by an independent listening test and held-out objective evaluation; parameter selection on a validation set does not make the test outcomes equivalent to the fitted inputs.

full rationale

The paper's derivation chain is self-contained and non-circular. DRC parameters (threshold, ratio, attack, release) were optimized on the validation set of Dataset 2 to maximize SI-SNR, but the reported objective evaluation uses a separately constructed Test Set 3, and the central subjective claim rests on listening-test ratings that were not used in parameter fitting. DRC's behavior is not defined in terms of triggerability ratings or SI-SNR improvement; it is a standard DSP operation. The paper does not rely on self-citations: the cited prior work (e.g., Veluri et al., Keshavarzi et al.) is external and not from the same authors. The appended limitation that neutral-sound preservation was not tested ('Listening tests should be conducted to determine to what extent neutral or non-triggering sounds remain recognizable...') is a genuine gap in support for the selectivity claim, and the lack of a volume-matched baseline is a correctness risk, but neither reduces the central result to its own inputs by construction. Score 0 reflects the absence of circularity, not the absence of validity concerns.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper has no invented physical entities. Its central empirical contributions rest on a chain of domain assumptions: the validity of the triggerability rating scale, the representativeness of synthetic mixtures, the validity of the Reddit-derived trigger list, and the assumption that the unaided auto-encoder learns trigger features. The fitted parameters (DRC, EQ, AGC, model hyperparameters) are not a fundamental contribution, but they shape the reported results and are not all reported.

free parameters (5)
  • DRC parameters (threshold, ratio, attack, release) = threshold -35 dB, ratio 30:1, attack 0.01 ms, release 100 ms
    Optimized on the validation split of Dataset 2 to maximize SI-SNR; a free-parameter fit that shapes the central positive result for DRC.
  • EQ gains and shelf frequencies = gain -8 dB at 200 Hz low-shelf, -2.75 dB (frequency unspecified, likely around 2-4 kHz), +1.6 dB at 5 kHz, -3 dB at 10 k
    Optimized on the same validation set for SI-SNR; only reported as the resulting configuration.
  • AGC parameters (attack/release, target level, max gain) = not reported, tuning 'resulted in poor SI-SNR performance'
    Tuned on the validation set but not reported in the paper.
  • Neural network hyperparameters (epochs, learning rate, latent dim) = NN1 150 epochs, NN2 50 epochs, Adam lr 5e-4, latent dim 512
    Chosen by the authors without a reported search; these choices affect model performance and the comparison with DRC.
  • Test Set 3 construction parameters = trigger at 0 dB, neutral at -10 dB, traffic at -35 dB, 5 s clips
    The specific SNR levels and durations are hand-picked and are not varied; the paper acknowledges that DRC performance depends on SNR and foreground/background placement.
axioms (4)
  • domain assumption Self-reported triggerability on a 0-100 continuous scale measures auditory distress (or 'triggerability') in a valid way.
    Used for the main subjective outcome; no validation study or clinical instrument is cited for this specific scale.
  • domain assumption A 5-10 second synthetic mixture of trigger + neutral + background noise is representative of real trigger-sound scenarios.
    Used in both datasets and Test Set 3; the conclusion treats performance on these mixtures as evidence for real-world usefulness.
  • domain assumption LLM-based extraction from Reddit posts, mapped with GPT-4 to AudioSet labels, gives a valid and representative list of neurodivergent trigger sounds.
    The trigger list is the foundation of the dataset; no human curation or validation of the extraction process is reported.
  • domain assumption The Semantic Hearing / Waveformer auto-encoder architecture can be retrained without input labels and still learn trigger-sound features.
    The authors removed the label input from Veluri et al. [16] and assume the decoder-side self/cross-attention can discover trigger characteristics from SI-SNR loss alone.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Low-latency Assistive Audio Enhancement for Neurodivergent People." pith.science (2026). https://pith.science/paper/KJV6GEVK

@misc{pith2026250910202,
  author       = {Pith},
  title        = {Pith review of: Low-latency Assistive Audio Enhancement for Neurodivergent People},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KJV6GEVK}},
  note         = {Machine review of arXiv:2509.10202}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Neurodivergent people frequently experience decreased sound tolerance, with estimates suggesting it affects 50-70% of this population. This heightened sensitivity can provoke reactions ranging from mild discomfort to severe distress, highlighting the critical need for assistive audio enhancement technologies In this paper, we propose several assistive audio enhancement algorithms designed to selectively filter distressing sounds. To address this, we curated a list of potential trigger sounds by analyzing neurodivergent-focused communities on platforms such as Reddit. Using this list, a dataset of trigger sound samples was compiled from publicly available sources, including FSD50K and ESC50. These samples were then used to train and evaluate various Digital Signal Processing (DSP) and Machine Learning (ML) audio enhancement algorithms. Among the approaches explored, Dynamic Range Compression (DRC) proved the most effective, successfully attenuating trigger sounds and reducing auditory distress for neurodivergent listeners.

discussion (0)

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

Reference graph

Works this paper leans on

33 extracted references · 7 linked inside Pith

  1. [1]

    In some in- dividuals – particularly those who are neurodivergent – these sounds may cause distress, anxiety, or other negative reactions that can diminish overall quality of life

    INTRODUCTION From daily activities like attending school or going to work to more recreational pursuits such as watching movies or playing video games, people are routinely exposed to various sounds. In some in- dividuals – particularly those who are neurodivergent – these sounds may cause distress, anxiety, or other negative reactions that can diminish o...

    Pith/arXiv arXiv 2025

  2. [2]

    METHODS AND EXPERIMENTAL SETUP The methodology consists of three major steps: i) design and con- struction of novel neurodivergent-trigger and neutral sound collec- tions, ii) mixing trigger sounds and non-triggering neutral sounds for training and assessment of different audio processing algorithms, and iii) training and optimizing existing (baseline) al...

  3. [3]

    very weak

    EV ALUA TION AND RESULTS To ensure that ourNN1andNN2models are not biased on the re- spective test sets, we designed a newTest Set 3used for both ob- jective and subjective assessment. The test contains 10 different 5-second stimuli, each with a different trigger and neutral sound pairing. The audio mixtures were constructed by combining three sounds over...

  4. [4]

    A key aspect of this research was the creation of a trigger sound dataset, which enabled the training and evaluation of both DSP and ML audio enhancement algorithms

    CONCLUSION AND FUTURE WORK This study has shown the potential of low-latency assistive audio en- hancement in reducing auditory distress for neurodivergent individ- uals by selectively attenuating trigger sounds and its potential use in a selective transparency mode. A key aspect of this research was the creation of a trigger sound dataset, which enabled ...

  5. [5]

    ACKNOWLEDGMENTS We would like to express our heartfelt gratitude to the pool of neu- rodivergent listening test participants for their time and effort

  6. [6]

    A review of decreased sound tol- erance in autism: Definitions, phenomenology, and potential mechanisms,

    Zachary J. Williams, Jason L. He, Carissa J. Cascio, and Tiffany G. Woynaroski, “A review of decreased sound tol- erance in autism: Definitions, phenomenology, and potential mechanisms,”Neuroscience & Biobehavioral Reviews, vol. 121, pp. 1–17, Feb. 2021

    2021

  7. [7]

    Assessment of Reduced Tolerance to Sound (Hyperacusis) in University Students,

    Sule Yilmaz, Memduha Tas ¸, Erdo˘gan Bulut, and Elc ¸in Nurc ¸in, “Assessment of Reduced Tolerance to Sound (Hyperacusis) in University Students,”Noise & Health, vol. 19, no. 87, pp. 73– 78, 2017

    2017

  8. [8]

    Prevalence of Hyperacusis in the General and Special Populations: A Scoping Review,

    Jing Ren, Tao Xu, Tao Xiang, Jun-mei Pu, Lu Liu, Yan Xiao, and Dan Lai, “Prevalence of Hyperacusis in the General and Special Populations: A Scoping Review,”Frontiers in Neurol- ogy, vol. 12, pp. 706555, Sept. 2021

    2021

  9. [9]

    Prevalence of Decreased Sound Tolerance (Hypera- cusis) in Individuals With Autism Spectrum Disorder: A Meta- Analysis,

    Zachary J. Williams, Evan Suzman, and Tiffany G. Woy- naroski, “Prevalence of Decreased Sound Tolerance (Hypera- cusis) in Individuals With Autism Spectrum Disorder: A Meta- Analysis,”Ear and Hearing, vol. 42, no. 5, pp. 1137–1150, 2021

    2021

  10. [10]

    A neuropsychological study of misophonia,

    Amitai Abramovitch, Tanya A. Herrera, and Joseph L. Ether- ton, “A neuropsychological study of misophonia,”Journal of Behavior Therapy and Experimental Psychiatry, vol. 82, pp. 101897, Mar. 2024

    2024

  11. [11]

    Autistic traits, emotion regulation, and sensory sensitivities in children and adults with Misophonia,

    L. J. Rinaldi, J. Simner, S. Koursarou, and J. Ward, “Autistic traits, emotion regulation, and sensory sensitivities in children and adults with Misophonia,”Journal of Autism and Develop- mental Disorders, vol. 53, no. 3, pp. 1162–1174, Mar. 2023

    2023

  12. [12]

    A phenomenological cartography of miso- phonia and other forms of sound intolerance,

    Nora Andermane, Mathilde Bauer, Ediz Sohoglu, Julia Simner, and Jamie Ward, “A phenomenological cartography of miso- phonia and other forms of sound intolerance,”iScience, vol. 26, no. 4, pp. 106299, Feb. 2023

    2023

  13. [13]

    What sound sources trigger misophonia? Not just chewing and breathing,

    Heather A. Hansen, Andrew B. Leber, and Zeynep M. Saygin, “What sound sources trigger misophonia? Not just chewing and breathing,”Journal of Clinical Psychology, vol. 77, no. 11, pp. 2609–2625, Nov. 2021

    2021

  14. [14]

    Prevalence and Char- acteristics of Patients with Severe Hyperacusis among Patients Seen in a Tinnitus and Hyperacusis Clinic,

    Hashir Aazh and Brian C. J. Moore, “Prevalence and Char- acteristics of Patients with Severe Hyperacusis among Patients Seen in a Tinnitus and Hyperacusis Clinic,”Journal of the American Academy of Audiology, vol. 29, no. 7, pp. 626–633, 2018

    2018

  15. [15]

    Audiometric Characteristics of Hyperacusis Patients,

    Jacqueline Sheldrake, Peter U. Diehl, and Roland Schaette, “Audiometric Characteristics of Hyperacusis Patients,”Fron- tiers in Neurology, vol. 6, pp. 105, May 2015

    2015

  16. [16]

    Effectiveness of Noise-Attenuating Head- phones on Physiological Responses for Children With Autism Spectrum Disorders,

    Beth Pfeiffer, Leah Stein Duker, AnnMarie Murphy, and Chengshi Shui, “Effectiveness of Noise-Attenuating Head- phones on Physiological Responses for Children With Autism Spectrum Disorders,”Frontiers in Integrative Neuroscience, vol. 13, pp. 65, Nov. 2019

    2019

  17. [17]

    Knowledge and Awareness of Ear Protection Devices for Sound Sensitivity by Individuals With Autism Spectrum Dis- orders,

    DiToro Dorothy Neave, Akiko Fuse, and Michael Bergen, “Knowledge and Awareness of Ear Protection Devices for Sound Sensitivity by Individuals With Autism Spectrum Dis- orders,”Language, Speech, and Hearing Services in Schools, vol. 52, no. 1, pp. 409–425, Jan. 2021, Publisher: American Speech-Language-Hearing Association

    2021

  18. [18]

    Stft-domain neural speech enhancement with very low algorithmic latency,

    Zhong-Qiu Wang, Gordon Wichern, Shinji Watanabe, and Jonathan Le Roux, “Stft-domain neural speech enhancement with very low algorithmic latency,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 397– 410, 2022

    2022

  19. [19]

    Ultra-low latency speech enhancement-a comprehensive study,

    Haibin Wu and Sebastian Braun, “Ultra-low latency speech enhancement-a comprehensive study,”arXiv preprint arXiv:2409.10358, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Towards sub- millisecond latency real-time speech enhancement models on hearables,

    Artem Dementyev, Chandan KA Reddy, Scott Wisdom, Navin Chatlani, John R Hershey, and Richard F Lyon, “Towards sub- millisecond latency real-time speech enhancement models on hearables,”arXiv preprint arXiv:2409.18239, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Semantic Hearing: Programming Acoustic Scenes with Binaural Hearables,

    Bandhav Veluri, Malek Itani, Justin Chan, Takuya Yoshioka, and Shyamnath Gollakota, “Semantic Hearing: Programming Acoustic Scenes with Binaural Hearables,” inProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, San Francisco CA USA, Oct. 2023, pp. 1–15, ACM

    2023

  22. [22]

    FSD50K: An Open Dataset of Human- Labeled Sound Events,

    Eduardo Fonseca, Xavier Favory, Jordi Pons, Frederic Font, and Xavier Serra, “FSD50K: An Open Dataset of Human- Labeled Sound Events,” Apr. 2022, arXiv:2010.00475 [cs] ver- sion: 2

    Pith/arXiv arXiv 2022

  23. [23]

    DISCO-10M: A Large-Scale Music Dataset,

    Luca A. Lanzend ¨orfer, Florian Gr ¨otschla, Emil Funke, and Roger Wattenhofer, “DISCO-10M: A Large-Scale Music Dataset,” Oct. 2023, arXiv:2306.13512 [cs]

    Pith/arXiv arXiv 2023

  24. [24]

    ESC: Dataset for Environmental Sound Clas- sification,

    Karol J. Piczak, “ESC: Dataset for Environmental Sound Clas- sification,” inProceedings of the 23rd ACM international con- ference on Multimedia, New York, NY , USA, Oct. 2015, MM ’15, pp. 1015–1018, Association for Computing Machinery

    2015

  25. [25]

    Optuna: A hyperparameter optimization framework — Op- tuna 4.2.0 documentation,

    “Optuna: A hyperparameter optimization framework — Op- tuna 4.2.0 documentation,”

  26. [26]

    Tree-Structured Parzen Estimator: Under- standing Its Algorithm Components and Their Roles for Better Empirical Performance,

    Shuhei Watanabe, “Tree-Structured Parzen Estimator: Under- standing Its Algorithm Components and Their Roles for Better Empirical Performance,” May 2023, arXiv:2304.11127 [cs]

    Pith/arXiv arXiv 2023

  27. [27]

    Equalization (audio),

    “Equalization (audio),” Feb. 2025, Page Version ID: 1274442372

    2025

  28. [28]

    All About Audio Equal- ization: Solutions and Frontiers,

    Vesa V ¨alim¨aki and Joshua D. Reiss, “All About Audio Equal- ization: Solutions and Frontiers,”Applied Sciences, vol. 6, no. 5, pp. 129, May 2016, Number: 5 Publisher: Multidisciplinary Digital Publishing Institute

    2016

  29. [29]

    Transient Noise Reduction Using a Deep Recur- rent Neural Network: Effects on Subjective Speech Intelligi- bility and Listening Comfort,

    Mahmoud Keshavarzi, Tobias Reichenbach, and Brian C. J. Moore, “Transient Noise Reduction Using a Deep Recur- rent Neural Network: Effects on Subjective Speech Intelligi- bility and Listening Comfort,”Trends in Hearing, vol. 25, pp. 23312165211041475, Oct. 2021

    2021

  30. [30]

    Evaluation of a multi-channel algorithm for reducing transient sounds,

    Mahmoud Keshavarzi, Thomas Baer, and Brian C. J. Moore, “Evaluation of a multi-channel algorithm for reducing transient sounds,”International Journal of Audiology, vol. 57, no. 8, pp. 624–631, Aug. 2018

    2018

  31. [31]

    Real-Time Tar- get Sound Extraction,

    Bandhav Veluri, Justin Chan, Malek Itani, Tuochao Chen, Takuya Yoshioka, and Shyamnath Gollakota, “Real-Time Tar- get Sound Extraction,” Apr. 2023, arXiv:2211.02250 [cs]

    Pith/arXiv arXiv 2023

  32. [32]

    SDR – Half-baked or Well Done?,

    Jonathan Le Roux, Scott Wisdom, Hakan Erdogan, and John R. Hershey, “SDR – Half-baked or Well Done?,” inICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Brighton, United Kingdom, May 2019, pp. 626–630, IEEE

    2019

  33. [33]

    Separation of over- lapping audio signals: A review on current trends and evolving approaches,

    Kakali Nath and Kandarpa Kumar Sarma, “Separation of over- lapping audio signals: A review on current trends and evolving approaches,”Signal Process., vol. 221, no. C, Aug. 2024

    2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.