Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Toward Optimal ANC: Establishing Mutual Information Lower Bound

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that any active noise cancellation algorithm has its normalized residual error bounded below by the maximum of an information-theoretic term and a spectral-support term.

desk verdict The central Lemma 1 is algebraically invalid and is violated by a simple Gaussian example, so the paper's unified lower bound does not hold. read the letter →

arxiv 2505.17877 v1 pith:GWDFQSR4 submitted 2025-05-23 cs.IT cs.AIcs.LGcs.SDeess.ASmath.IT

classification cs.ITcs.AIcs.LGcs.SDeess.ASmath.IT MSC 94A1794A12
keywords activenoisecancellationmutualinformationlowerboundnormalizedmeansquarederrorrate-distortiontheoryspectralsupportreverberationdeeplearningANC
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 sets out to find a theoretical floor for the performance of active noise cancellation (ANC) systems. It proposes a unified lower bound on the normalized mean squared error (NMSE) that combines two independent constraints: how much information the anti-noise signal carries about the disturbance, and how much disturbance power lies in frequency bands the cancellation path cannot touch. If the bound is correct, it gives every ANC algorithm—classical or deep learning—a concrete benchmark to measure against, and tells engineers whether to improve the algorithm or the physical hardware. The authors test the bound on the NOISEX dataset with three deep ANC models over several reverberation times and find their results lie above the claimed floor.

What carries the argument

The engine of the paper is Lemma 1, an information-theoretic inequality: σ_e² ≥ σ_d²(1 - I(y;d)/H(d)). It is obtained by starting from the Shannon lower bound on the rate-distortion function, expressing distortion in terms of the entropy H(d) and mutual information I(y;d), then linearizing the resulting exponential expression with a first-order Taylor expansion. The second component is the spectral-support ratio, the fraction of disturbance energy in the frequency set supp(P) \ supp(S), computed directly from path transfer functions. The unified bound is the maximum of these two separate floors.

What would settle it

For a scalar Gaussian channel with d ~ N(0,1) and y = d + n where n ~ N(0,1) is independent, compute the mutual information I(y;d) = 0.5 ln 2 and the entropy H(d) = 0.5 ln(2πe). The claimed bound gives NMSE ≥ 1 - I/H ≈ 0.756, whereas the true minimum mean-squared error estimate of d from y, the conditional mean, has NMSE = 0.5. A direct calculation or a small simulation that estimates NMSE from samples will show the bound is violated, contradicting Lemma 1.

Watch

Extended reading notes

Core claim

The central claim is Theorem 2: for any ANC system with a primary path P(z) and secondary path S(z), the NMSE in decibels must satisfy NMSEdB ≥ max{10 log10(1 - I(y;d)/H(d)), 10 log10(∫_{supp(P)\supp(S)} S_dd($e^{{jω}}$) dω / ∫ S_dd($e^{{jω}}$) dω)}, where I(y;d) is the mutual information rate between the synthesized signal and the disturbance and H(d) is the disturbance's differential entropy rate. The first term, derived in Lemma 1, states that the residual error power is at least σ_d²(1 - I(y;d)/H(d)); the second term is the portion of disturbance power concentrated in frequencies the secondary path cannot reproduce. Taking the maximum of the two yields a single, model-independent ceiling that the paper argues applies to every cancellation algorithm, including deep networks trained end-to-end.

Load-bearing premise

The proof assumes that the first-order Taylor approximation of an exponential rate-distortion bound, turned into the linear inequality σ_e² ≥ σ_d²(1 - I/H), is a valid lower bound on the residual error; if that linearization direction is wrong, the information-theoretic half of the unified bound collapses.

Editorial extensions

If this is right

  • If the bound holds, any ANC model whose NMSE sits far above the ceiling still has headroom for algorithmic improvement, while a model near the bound is limited by information or hardware constraints.
  • The identity of the dominant term directs engineering effort: a large information term signals the algorithm is not extracting enough about the disturbance, while a large support term signals the actuator or secondary path must be redesigned.
  • The support-based term is entirely model-independent, so it can be computed before training any neural controller and used as a sanity-check lower bound.
  • The information-theoretic term grows with reverberation time, providing a quantitative explanation for why ANC is harder in more reverberant rooms.
  • The bound offers a unified reporting standard for the ANC literature, allowing different methods to be compared against a common theoretical reference rather than only against each other.

Reading between the lines

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

  • The critical step is the Taylor linearization from an exponential bound to 1 - I/H; because exp(-x) ≥ 1 - x, the linear form is a weaker bound than the exponential, but the paper's specific replacement of exp(-2I) with exp(-I/H) is where a reader should check the direction of the inequality.
  • A simple scalar Gaussian test—d ~ N(0,1), y = d + n with n ~ N(0,1)—gives I/H ≈ 0.244 and a claimed floor of about 0.756, while the true minimum mean-squared error is 0.5, indicating the linear bound may not hold as stated for Gaussian sources.
  • The paper's numerical estimation of mutual information relies on kernel density estimates from finite samples, which the authors acknowledge as a limitation; the bound's practical reliability therefore depends on sample size and bandwidth choices.
  • The support-based term is robust and likely to survive scrutiny, but the full unified bound inherits whatever flaws the information-theoretic term possesses.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a unified theoretical lower bound on the normalized mean-squared error (NMSE) of active noise cancellation systems. The bound in Theorem 2 is the maximum of two terms: an information-theoretic term sigma_d^2 (1 - I(y;d)/H(d)) from Lemma 1, and a spectral-support term from Lemma 2 measuring the disturbance power in frequencies where the secondary path has no gain. The authors validate the bound on the NOISEX dataset with three deep-learning ANC baselines over several reverberation times. The support-based Lemma 2 is a correct elementary observation, but the central information-theoretic Lemma 1 is false, and the dB conversion in Eq. (8) is inconsistent. Because Theorem 2 takes the maximum of these two terms, the paper's main claim is not established.

Significance. A valid, algorithm-independent lower bound on ANC performance would be a useful benchmark for the deep-learning ANC literature, and the empirical comparison across reverberation times is a sensible design. The support-based bound (Lemma 2) is a sound and simple frequency-domain constraint, and the authors are right that it is model-independent in its idealized form. However, the information-theoretic component, which is the paper's main novelty, is invalid: Lemma 1 is contradicted by a Gaussian counterexample, and the proof rests on an incorrect Taylor truncation. The paper also ships no code or reproducibility artifacts, and the experimental validation is weakened by the data-dependent estimation of I(y;d) from the very signals used to measure NMSE. The contribution therefore does not meet the standard for a theoretical result in information theory.

major comments (4)
  1. [Section 4, Lemma 1 and Eq. (2)] Lemma 1 is false. Take d ~ N(0,1) and y = d + n with n ~ N(0,1) independent. The minimum mean-squared error estimator is E[d|y] = y/2, giving NMSE = 0.5. For this Gaussian pair, H(d) = (1/2) log(2*pi*e) approx 1.419 nats and I(y;d) = (1/2) log 2 approx 0.347 nats, so the claimed right-hand side sigma_d^2 (1 - I/H) approx 0.756. Since 0.5 < 0.756, Eq. (2) is violated. The proof's transition from Eq. (5) to Eq. (6) is invalid: the Shannon lower bound yields D >= sigma_d^2 exp(-2I), and replacing exp(-2I) by 1 - I/H is neither a valid Taylor truncation nor the correct normalized exponent; in fact exp(-x) >= 1 - x would give a different linear term. Theorem 2 inherits this false first argument through its maximum.
  2. [Section 4, Eq. (8)] The conversion to dB double-counts the primary path norm. With NMSE = E[|e|^2]/E[|d|^2], Eq. (2) immediately gives NMSE >= 1 - I(y;d)/H(d) in linear units; there is no additional 10 log10(||P||_2^2) term in the decibel expression. Equation (8) appends +10 log10(||P||_2^2) after the variance ratio has already been normalized by sigma_d^2 = (1/2*pi) integral |P|^2 S_xx d*omega. The experiments in Section 8 report boundaries computed from Eq. (8), so the information-theoretic curves are offset by an unjustified constant that is not part of the stated NMSE definition.
  3. [Sections 4, 7, and A.1] The information-theoretic bound is not a model-independent ceiling. Lemma 1 bounds sigma_e^2 using I(y;d), which is a functional of the algorithm's output y(n); in the experiments I(y;d) is estimated from the same d(n) and y(n) that define the measured NMSE (Sections 7 and A.1). The manuscript itself states in Section 4 that the bound is 'inherently dependent on the specific algorithm.' A lower bound that is a function of the algorithm's own statistics cannot support the claimed 'theoretical ceiling on the NMSE attainable by any ANC algorithm' (Abstract and Theorem 2) unless the mutual information is extremized over the admissible class of y(n) or replaced by a source/channel quantity independent of the implementation. The current derivation gives no such extremization.
  4. [Section 5, Lemma 2 and Appendix A.2] Lemma 2 is sound as stated for exact spectral supports, but the numerical implementation does not evaluate this quantity. Appendix A.2 defines support by a magnitude threshold (e.g., 45 dB below the peak) and treats all bins below the threshold as having zero gain; with finite-length room impulse responses the true support is generically the full frequency band, so supp(P)\supp(S) is empty and Lemma 2 is vacuous. The reported support-based boundary is therefore an artifact of the threshold choice, and no sensitivity analysis over the threshold is provided. Consequently the experimental 'support-based bound' in Figures 2 and 4 is not the model-independent bound of Lemma 2.
minor comments (5)
  1. [Section 4, proof of Lemma 1] The sentence 'Recall, that for gaussian noise, the entropy is maximazied...' contains typos and an imprecise statement; the equality sigma_d^2 = (1/(2*pi*e)) e^{2H(d)} holds only for Gaussian sources, not for general noise.
  2. [Section 4, after Eq. (5)] The phrase 'for negative exponent it is upper bound' is unclear and, as written, incorrect; the Taylor step needs to be re-derived with the exponent I/H rather than 2I.
  3. [Section 5, Lemma 2] The proof would benefit from explicitly stating that supp(P)\supp(S) is interpreted as the set of frequencies where S has exactly zero gain; otherwise for typical finite-length impulse responses the bound is vacuous.
  4. [Section 9] The limitation paragraph admits that insufficient sampling 'distorts the bound calculations,' but no convergence analysis or bias correction for the KDE/MI estimates is provided; this should be quantified if the empirical validation is to be meaningful.
  5. [Section 7] 'Ressources' should be 'Resources'; the figure captions should state whether error bars or repeated trials are included.

Circularity Check

2 steps flagged · score 7.0 of 10

The information-theoretic bound is not an independent ceiling: Lemma 1 defines the bound as 1 minus the algorithm's own normalized mutual information, and the experiments compute I(y;d) from the same signals whose NMSE is being bounded.

  1. self definitional [Lemma 1 proof, Eqs. (5)-(7), Section 4]
    "expanding the second exponent as taylor series, and taking the first order yeilds, σ2 e ≥ σ2 d (1 − I(d(n); y(n)) H(d(n)) ) ... We thus propose a linear lower bound on the distortion based on this ratio: σ2 e ≥ σ2 d(1 − α)"

    Starting from the Shannon lower bound D ≥ (1/2πe)exp(2(H(d)−I)), the proof replaces exp(−2I) by exp(−I/H) and then Taylor-truncates to 1 − I/H. Because α := I/H is defined as the ratio of the same signals' mutual information and entropy, Eq. (2) is not derived from rate-distortion theory; it is the chosen linear function of α. The claimed lower bound is false (e.g., d∼N(0,1), y=d+n, I=0.3466, H=1.4189 gives claimed bound 0.7557 while the true MMSE is 0.5), confirming that the substitution is an imposed definition rather than a consequence of an external bound.

  2. fitted input called prediction [Section 7-8 empirical validation; Appendix A.1]
    "This information-theoretic bound is inherently dependent on the specific algorithm (as it determines the statistics of y(n) and thus I(y; d))."

    The paper validates the bound by estimating I(y;d) from the same algorithm output y(n) and disturbance d(n) whose NMSE is plotted (Appendix A.1 estimates marginal and joint KDE PDFs from these sequences). Thus the 'theoretical ceiling' is a per-algorithm re-encoding of the algorithm's own statistics; the empirical agreement is a consistency check between two estimates made on identical data, not a test of an independent prediction. The support-based term is independent, but the information-theoretic half of the unified bound is fitted input presented as a predicted lower bound.

full rationale

The paper's support-based Lemma 2 and the max construction are self-contained and non-circular, and no load-bearing self-citation appears. The circularity is concentrated in Lemma 1 and Theorem 2: the information-theoretic bound is constructed as 1 − I(y;d)/H(d), with I(y;d) estimated from the same d(n), y(n) whose NMSE it claims to bound. The proof's replacement of the Shannon-lower-bound exponent exp(−2I) by exp(−I/H), followed by a first-order Taylor truncation, is the specific reduction that manufactures this form; the paper even says 'we thus propose' the linear bound. In the experiments, the 'boundary' is computed per algorithm from its own output, so the empirical 'tightness' is a tautological comparison rather than validation of an independent prediction. Because the information-theoretic term is the paper's central contribution, the partial circularity warrants score 7. A mathematical counterexample (d∼N(0,1), y=d+n: claimed bound 0.7557 versus true MMSE 0.5) further shows the substituted expression is not a valid lower bound, but that failure is a correctness issue rather than the circularity itself.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim depends on a chain of distributional assumptions, a mathematically invalid Taylor truncation, and two unreported numerical parameters (KDE bandwidth and support threshold). The support bound alone is simple and correct, but the info bound introduces the main unsupported content.

free parameters (3)
  • spectral support threshold = 45 dB below the peak magnitude
    Appendix A.2 defines primary-path support as FFT bins above a 45 dB threshold; lower bins are treated as exactly uncancelable, and the support bound changes with this threshold.
  • KDE bandwidth/bin_count = not reported
    Appendix A.1 sets bandwidth inversely proportional to bin_count, but the actual bin_count and bandwidth values are not stated; MI and entropy estimates, and therefore the info bound, depend on this choice.
  • direct path scaling coefficient = not reported
    Appendix A.1 says the MI-derived coefficient scales the energy of the direct path in the room impulse response to yield the bound; the precise definition and fitting procedure are omitted.
assumptions (6)
  • domain assumption d(n) and y(n) are jointly WSS and ergodic with finite differential entropy rate H(d).
    Invoked at the start of Section 4 and in Lemma 1 to justify spectral power formulas and the use of entropy rates.
  • domain assumption The disturbance is generated by an LTI primary path P(z), and the cancellation signal is generated by an LTI secondary path S(z).
    Used in Section 5 to define spectral support; in Section 4 f is allowed to be nonlinear, so the two sections assume different structures for f.
  • standard math Shannon lower bound R(D) >= H(d) - (1/2) log(2 pi e D) applies to the source with MSE distortion.
    Basis of Eq (4) in Lemma 1; valid for stationary ergodic sources under stated conditions but not enough to justify the later algebra.
  • ad hoc to paper The inequality I(d;y) <= H(d) holds for differential entropy and alpha = I/H lies in [0,1].
    Used to normalize alpha in Lemma 1; false for differential entropies, which can be negative, and mutual information can exceed them. This is one source of the invalid bound.
  • ad hoc to paper First-order Taylor expansion exp(-x) approx 1-x is a valid lower bound on the exponential term.
    Used in Lemma 1, Eqs (6)-(7). The direction is wrong: exp(-x) >= 1-x, so the truncation is a lower bound on exp(-x), not on the distortion D. This reverses the inequality chain.
  • ad hoc to paper Spectral support can be identified by a magnitude threshold (45 dB below peak) and frequencies below threshold have exactly zero cancellation gain.
    Used in Appendix A.2 to compute the support-based bound; introduces a free threshold that changes the bound's value.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Toward Optimal ANC: Establishing Mutual Information Lower Bound." pith.science (2026). https://pith.science/paper/GWDFQSR4

@misc{pith2026250517877,
  author       = {Pith},
  title        = {Pith review of: Toward Optimal ANC: Establishing Mutual Information Lower Bound},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GWDFQSR4}},
  note         = {Machine review of arXiv:2505.17877}
}
read the original abstract

Active Noise Cancellation (ANC) algorithms aim to suppress unwanted acoustic disturbances by generating anti-noise signals that destructively interfere with the original noise in real time. Although recent deep learning-based ANC algorithms have set new performance benchmarks, there remains a shortage of theoretical limits to rigorously assess their improvements. To address this, we derive a unified lower bound on cancellation performance composed of two components. The first component is information-theoretic: it links residual error power to the fraction of disturbance entropy captured by the anti-noise signal, thereby quantifying limits imposed by information-processing capacity. The second component is support-based: it measures the irreducible error arising in frequency bands that the cancellation path cannot address, reflecting fundamental physical constraints. By taking the maximum of these two terms, our bound establishes a theoretical ceiling on the Normalized Mean Squared Error (NMSE) attainable by any ANC algorithm. We validate its tightness empirically on the NOISEX dataset under varying reverberation times, demonstrating robustness across diverse acoustic conditions.

Figures

Figures reproduced from arXiv: 2505.17877 by the authors.

Figure 1
Figure 1. High-level schematic of a feedfor￾ward ANC system with lower-bound calcula￾tion. To establish fundamental benchmarks against which practical algorithms can be evaluated, we derive a lower bound on the achievable cancellation performance rooted in information theory. This bound quantifies the irreducible error stemming from the information processing limitations inherent in generating y(n) based on available observat… view at source ↗
Figure 2
Figure 2. Model performances for various reverberation times [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Time-Frequency analysis of the different noise [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: shows the performances and the discussed boundaries but this time for t60 = 0.2s. As previously mentioned, the unified boundary—defined as the maximum of the Information-Theoretic and Support-based boundaries—encloses the performance achieved by the different methods a…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 43 canonical work pages

  1. [1]

    Process of silencing sound oscillations

    Paul Lueg. Process of silencing sound oscillations. US patent 2043416, 1936

  2. [2]

    Active control of sound

    Philip Arthur Nelson and Stephen J Elliott. Active control of sound. Academic press, 1991

  3. [3]

    Active control of vibration

    Christopher C Fuller, Sharon Elliott, and Philip Arthur Nelson. Active control of vibration. Academic press, 1996

  4. [4]

    Active control of noise and vibration

    Colin H Hansen, Scott D Snyder, Xiaojun Qiu, Laura A Brooks, and Danielle J Moreau. Active control of noise and vibration. E & Fn Spon London, 1997

  5. [5]

    Active noise control: a tutorial review

    Sen M Kuo and Dennis R Morgan. Active noise control: a tutorial review. Proceedings of the IEEE, 87(6):943–973, 1999

  6. [6]

    Effect of errors in the plant model on the performance of algorithms for adaptive feedforward control

    CC Boucher, SJ Elliott, and PA Nelson. Effect of errors in the plant model on the performance of algorithms for adaptive feedforward control. In IEE Proceedings F (Radar and Signal Processing), volume 138, pages 313–319. IET, 1991

  7. [7]

    Active mitigation of nonlinear noise processes using a novel filtered-s lms algorithm

    Debi Prasad Das and Ganapati Panda. Active mitigation of nonlinear noise processes using a novel filtered-s lms algorithm. IEEE Transactions on Speech and Audio Processing, 12(3): 313–322, 2004

  8. [8]

    Identification of nonlinear dynamic systems using functional link artificial neural networks

    Jagdish Chandra Patra, Ranendra N Pal, BN Chatterji, and Ganapati Panda. Identification of nonlinear dynamic systems using functional link artificial neural networks. IEEE transactions on systems, man, and cybernetics, part b (cybernetics), 29(2):254–262, 1999

Show all 44 references
  1. [9]

    Adaptive volterra filters for active control of nonlinear noise processes

    Li Tan and Jean Jiang. Adaptive volterra filters for active control of nonlinear noise processes. IEEE Transactions on signal processing, 49(8):1667–1676, 2001

  2. [10]

    Nonlinear adaptive bilinear filters for active noise control systems

    Sen M Kuo and Hsien-Tsai Wu. Nonlinear adaptive bilinear filters for active noise control systems. IEEE Transactions on Circuits and Systems I: Regular Papers, 52(3):617–624, 2005

  3. [11]

    Leaky-fxlms algorithm: Stochastic analysis for gaussian data and secondary path modeling error

    Orlando José Tobias and Rui Seara. Leaky-fxlms algorithm: Stochastic analysis for gaussian data and secondary path modeling error. IEEE Transactions on speech and audio processing, 13(6):1217–1230, 2005

  4. [12]

    Deep anc: A deep learning approach to active noise control

    Hao Zhang and DeLiang Wang. Deep anc: A deep learning approach to active noise control. Neural Networks, 141:1–10, 2021

  5. [13]

    Self-attending rnn for speech enhancement to improve cross-corpus generalization

    Ashutosh Pandey and DeLiang Wang. Self-attending rnn for speech enhancement to improve cross-corpus generalization. IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, 30:1374–1385, 2022

  6. [14]

    Deep active speech cancellation with multi- band mamba network, 2025

    Yehuda Mishaly, Lior Wolf, and Eliya Nachmani. Deep active speech cancellation with multi- band mamba network, 2025. URL https://arxiv.org/abs/2502.01185

  7. [15]

    Had-anc: A hy- brid system comprising an adaptive filter and deep neural networks for active noise control

    JungPhil Park, Jeong-Hwan Choi, Yungyeo Kim, and Joon-Hyuk Chang. Had-anc: A hy- brid system comprising an adaptive filter and deep neural networks for active noise control. In INTERSPEECH, volume 2023, pages 2513–2517. International Speech Communication Association, 2023

  8. [16]

    Deep learning-based active noise control on construction sites

    Alireza Mostafavi and Young-Jin Cha. Deep learning-based active noise control on construction sites. Automation in Construction, 151:104885, 2023

  9. [17]

    Enhancing active noise control through stacked autoencoders: Training strategies, comparative analysis, and evaluation with practical setup

    Deepali Singh, Rinki Gupta, Arun Kumar, and Rajendar Bahl. Enhancing active noise control through stacked autoencoders: Training strategies, comparative analysis, and evaluation with practical setup. Engineering Applications of Artificial Intelligence, 135:108811, 2024

  10. [18]

    Generalized performance of neural network controllers for feedforward active control of nonlinear systems

    Alexander Pike and Jordan Cheer. Generalized performance of neural network controllers for feedforward active control of nonlinear systems. 2023

  11. [19]

    Feedforward selective fixed-filter active noise control: Algorithm and implementation

    Dongyuan Shi, Woon-Seng Gan, Bhan Lam, and Shulin Wen. Feedforward selective fixed-filter active noise control: Algorithm and implementation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:1479–1492, 2020. 10

  12. [20]

    Integrated 3d active noise cancellation simulation and synthesis platform using tcl

    Seunghyun Park and Daejin Park. Integrated 3d active noise cancellation simulation and synthesis platform using tcl. In 2023 IEEE 16th International Symposium on Embedded MCSoC, pages 111–116. IEEE, 2023

  13. [21]

    Gfanc-kalman: Generative fixed-filter active noise control with cnn-kalman filtering

    Zhengding Luo, Dongyuan Shi, Xiaoyi Shen, Junwei Ji, and Woon-Seng Gan. Gfanc-kalman: Generative fixed-filter active noise control with cnn-kalman filtering. IEEE Signal Processing Letters, 2023

  14. [22]

    Low-latency active noise control using attentive recurrent network

    Hao Zhang, Ashutosh Pandey, and DeLiang Wang. Low-latency active noise control using attentive recurrent network. IEEE/ACM transactions on audio, speech, and language processing, 31:1114–1123, 2023

  15. [23]

    Active adaptive sound control in a duct: A computer simulation

    John C Burgess. Active adaptive sound control in a duct: A computer simulation. The Journal of the Acoustical Society of America, 70(3):715–726, 1981

  16. [24]

    Nonlinear thf-fxlms algorithm for active noise control with loudspeaker nonlinearity

    Sepehr Ghasemi, Raja Kamil, and Mohammad Hamiruce Marhaban. Nonlinear thf-fxlms algorithm for active noise control with loudspeaker nonlinearity. Asian Journal of Control, 18 (2):502–513, 2016

  17. [25]

    Dnoisenet: Deep learning-based feedback active noise control in various noisy environments

    Young-Jin Cha, Alireza Mostafavi, and Sukhpreet S Benipal. Dnoisenet: Deep learning-based feedback active noise control in various noisy environments. Engineering Applications of Artificial Intelligence, 121:105971, 2023

  18. [26]

    Selective fixed- filter active noise control based on convolutional neural network

    Dongyuan Shi, Bhan Lam, Kenneth Ooi, Xiaoyi Shen, and Woon-Seng Gan. Selective fixed- filter active noise control based on convolutional neural network. Signal Processing, 190: 108317, 2022

  19. [27]

    Transferable latent of cnn-based selective fixed-filter active noise control

    Dongyuan Shi, Woon-Seng Gan, Bhan Lam, Zhengding Luo, and Xiaoyi Shen. Transferable latent of cnn-based selective fixed-filter active noise control. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2910–2921, 2023

  20. [28]

    A hybrid sfanc-fxnlms algorithm for active noise control based on deep learning

    Zhengding Luo, Dongyuan Shi, and Woon-Seng Gan. A hybrid sfanc-fxnlms algorithm for active noise control based on deep learning. IEEE Signal Processing Letters, 29:1102–1106, 2022

  21. [29]

    Deep generative fixed-filter active noise control

    Zhengding Luo, Dongyuan Shi, Xiaoyi Shen, Junwei Ji, and Woon-Seng Gan. Deep generative fixed-filter active noise control. In ICASSP, pages 1–5. IEEE, 2023

  22. [30]

    Delayless generative fixed- filter active noise control based on deep learning and bayesian filter

    Zhengding Luo, Dongyuan Shi, Woon-Seng Gan, and Qirui Huang. Delayless generative fixed- filter active noise control based on deep learning and bayesian filter. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023

  23. [31]

    Unsupervised learning based end-to-end delayless generative fixed-filter active noise control

    Zhengding Luo, Dongyuan Shi, Xiaoyi Shen, and Woon-Seng Gan. Unsupervised learning based end-to-end delayless generative fixed-filter active noise control. In ICASSP, pages 441–445. IEEE, 2024

  24. [32]

    A new dropout leaky control strategy for multi-channel narrowband active noise cancellation in irregular reverberation room

    Wenzhao Zhu, Bo Xu, Zong Meng, and Lei Luo. A new dropout leaky control strategy for multi-channel narrowband active noise cancellation in irregular reverberation room. In 2021 7th International Conference on Computer and Communications (ICCC), pages 1773–1777. IEEE, 2021

  25. [33]

    Deep mcanc: A deep learning approach to multi-channel active noise control

    Hao Zhang and DeLiang Wang. Deep mcanc: A deep learning approach to multi-channel active noise control. Neural Networks, 158:318–327, 2023

  26. [34]

    Remote micro- phone technique for active noise control over distributed networks

    Christian Antoñanzas, Miguel Ferrer, Maria De Diego, and Alberto Gonzalez. Remote micro- phone technique for active noise control over distributed networks. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:1522–1535, 2023

  27. [35]

    Spatially selective active noise control systems

    Tong Xiao, Buye Xu, and Chuming Zhao. Spatially selective active noise control systems. The Journal of the Acoustical Society of America, 153(5):2733–2733, 2023

  28. [36]

    A time- domain multi-channel directional active noise control system

    Huawei Zhang, Jihui Zhang, Fei Ma, Prasanga N Samarasinghe, and Huiyuan Sun. A time- domain multi-channel directional active noise control system. In 2023 31st EUSIPCO, pages 376–380. IEEE, 2023. 11

  29. [37]

    Multichannel two-gradient direction filtered reference least mean square algorithm for output-constrained multichannel active noise control

    Dongyuan Shi, Bhan Lam, Xiaoyi Shen, and Woon-Seng Gan. Multichannel two-gradient direction filtered reference least mean square algorithm for output-constrained multichannel active noise control. Signal Processing, 207:108938, 2023

  30. [38]

    What is behind the meta-learning initialization of adaptive filter?—a naive method for accelerating convergence of adaptive multichannel active noise control

    Dongyuan Shi, Woon-seng Gan, Xiaoyi Shen, Zhengding Luo, and Junwei Ji. What is behind the meta-learning initialization of adaptive filter?—a naive method for accelerating convergence of adaptive multichannel active noise control. Neural Networks, 172:106145, 2024

  31. [39]

    Attentive recurrent network for low-latency active noise control

    Hao Zhang, Ashutosh Pandey, and DeLiang Wang. Attentive recurrent network for low-latency active noise control. In INTERSPEECH, pages 956–960, 2022

  32. [40]

    Genetic algorithm-based adaptive active noise control without secondary path identification

    Yang Zhou, Haiquan Zhao, and Dongxu Liu. Genetic algorithm-based adaptive active noise control without secondary path identification. IEEE Transactions on Instrumentation and Measurement, 2023

  33. [41]

    An improved artificial bee colony algorithm for model-free active noise control: algorithm and implementation

    Xing Ren and Hongwei Zhang. An improved artificial bee colony algorithm for model-free active noise control: algorithm and implementation. IEEE Transactions on Instrumentation and Measurement, 71:1–11, 2022

  34. [42]

    Assessment for automatic speech recognition: Ii

    Andrew Varga and Herman JM Steeneken. Assessment for automatic speech recognition: Ii. noisex-92: A database and an experiment to study the effect of additive noise on speech recognition systems. Speech communication, 12(3):247–251, 1993

  35. [43]

    Image method for efficiently simulating small-room acoustics

    Jont B Allen and David A Berkley. Image method for efficiently simulating small-room acoustics. The Journal of the Acoustical Society of America, 65(4):943–950, 1979

  36. [44]

    On the lms algorithm with constant and variable leakage factor in a nonlinear environment

    Orlando José Tobias and Rui Seara. On the lms algorithm with constant and variable leakage factor in a nonlinear environment. IEEE transactions on signal processing, 54(9):3448–3458, 2006. 12 A Technical Appendices and Supplementary Material A.1 Numerical calculation of the In...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.