Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

SincQDR-VAD claims that replacing fixed filterbanks with learnable sinc filters and training with a ranking loss makes a compact voice-activity detector robust in noise.

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 ↗

A lightweight VAD model with a learnable sinc filterbank and a squared-margin ranking loss reports higher AUROC and F2 scores on AVA-Speech and ACAM using only 8.0k parameters.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A compact VAD that combines sinc filters and a ranking loss shows real gains, but the training-label construction and missing error bars keep the headline claims from being fully sealed. the 4 major comments →

arxiv 2508.20885 v1 pith:E4EUI62R submitted 2025-08-28 cs.SD

SincQDR-VAD: A Noise-Robust Voice Activity Detection Framework Leveraging Learnable Filters and Ranking-Aware Optimization

Chien-Chun Wang , En-Lun Yu , Jeih-Weih Hung , Shih-Chieh Huang , Berlin Chen This is my paper
classification cs.SD
keywords voice activity detectionlearnable sinc filtersquadratic disparity ranking lossAUROC optimizationnoise robustnesslightweight neural networkedge deployment
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

This paper proposes SincQDR-VAD, a compact voice activity detection framework intended for noisy and resource-limited settings. Its central claim is that two design changes—using learnable sinc bandpass filters in place of a fixed mel filterbank, and adding a pairwise quadratic-disparity ranking loss on top of binary cross-entropy—make a small model markedly more robust to background noise. On the paper's experiments, the model reaches AUROC 0.914 and F2-Score 0.911 on AVA-Speech, averages AUROC 0.815 on a noisy AVA-Speech variant spanning 10 dB to -10 dB SNR, and reaches AUROC 0.97 / F2-Score 0.92 on ACAM, with only 8.0k parameters (31% fewer than the leading lightweight baseline). A practical reader would care because VAD is a first processing stage for speech recognizers and hearing devices, and gains that come with fewer parameters rather than more are directly usable on edge hardware.

Core claim

The central discovery is that the combination of a learnable sinc-filter front-end and a ranking loss aligns a very small VAD with the AUROC objective and with noisy spectral structure. Each filter is a bandpass sinc function with learnable low/high cutoff frequencies and a learnable gain, applied to raw waveform frames to produce sub-band log-energies; the ranking loss penalizes, with a squared margin, every speech/non-speech pair whose scores are separated by less than m=1. Trained with this loss plus BCE on one-second speech clips with added environmental noise, the model achieves the reported gains; ablations attribute the larger share of the gain to the ranking loss (AUROC drops to 0.87

What carries the argument

The Sinc-extractor front-end: each filter is a truncated, delayed difference of two sinc functions parameterized by learnable lower and upper cutoff frequencies, multiplied by a learnable band gain and a Hamming window; the bank converts raw waveform frames into log sub-band energies. The quadratic disparity ranking (QDR) loss is $$L_{QDR} = \frac{1}{|P||N|}\sum_{i \in P}\sum_{j \in N} \left(\max(0, m - (score_i - score_j))\right)^2$$, with margin $m=1$, combined with BCE via weight $\lambda=0.25$. The front-end is meant to provide explicit, noise-resistant spectral control; the ranking loss is meant to optimize pairwise ordering of speech vs non-speech frames, i.e., the AUROC objective.

Load-bearing premise

The load-bearing premise is that speech labels generated by the rule 'central 0.2–0.83 seconds of each one-second training clip is speech, the rest is background' transfer to real continuous audio; if the model instead learns that speech appears only in a predictable middle window, the reported test gains could come from that positional prior rather than from noise robustness.

What would settle it

Train SincQDR-VAD on the same SCF data but with manually annotated speech boundaries (not the fixed central-segment rule), then evaluate on ACAM and noisy AVA-Speech. If the AUROC and F2 advantages over TinyVAD largely disappear, the reported noise-robustness gains are artifacts of the synthetic label prior rather than the sinc filters or ranking loss.

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

If this is right

  • Noise-robust frame-level VAD can be made smaller, not larger: an 8.0k-parameter model outperforms baselines that use more parameters, which helps real-time edge processing.
  • Pairwise rank optimization is a workable complement to BCE for imbalanced speech/non-speech labels, since it targets AUROC directly rather than threshold-dependent accuracy.
  • Learnable front-ends can replace fixed mel filterbanks without making the model heavier, shifting spectral adaptation into the training process.
  • The ACAM results indicate the training recipe transfers from synthetic one-second clips to continuous real-world recordings with very different acoustics.

Where Pith is reading between the lines

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

  • Editorial: the QDR loss is a general ranking objective; the same squared-margin pairwise loss should transfer to other imbalanced frame-level audio detection tasks, such as sound event detection or wake-word spotting, though the paper does not test those.
  • Editorial: the training label rule—speech only in the central 0.2–0.83 s of each one-second clip—may teach a positional prior; if the model is using temporal position, the ACAM F2 gain could partially reflect that prior, and retraining on fully annotated boundaries would be the clean check.
  • Editorial: the ablation suggests ranking loss, not the sinc front-end, is the larger contributor; a cross-dataset decomposition analysis would clarify which component carries low-SNR robustness.
  • Editorial: a promising extension is pairing this front-end and loss with a state-space temporal encoder to reduce residual false alarms, but stability at very low SNR would need validating.
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

4 major / 6 minor

Summary. The paper proposes SincQDR-VAD, a lightweight VAD system combining a learnable Sinc-filter front-end with a hybrid training objective that mixes BCE with a quadratic disparity ranking (QDR) loss. The model is trained on one-second Google Speech Commands V2 clips mixed with Freesound noise, using a fixed central window (0.2–0.83 s) as the active-speech region, and is evaluated on AVA-Speech, a noisy variant of AVA-Speech, and ACAM. The authors report AUROC/F2 improvements over MarbleNet and TinyVAD, an ACAM F2 gain from 0.65 to 0.92, and a parameter count of 8.0k, with code and checkpoints made available.

Significance. If the reported gains are genuine, the paper would make a useful contribution to lightweight noise-robust VAD: a sub-10k-parameter model with a learnable filterbank and a ranking-aware loss that improves both discrimination and recall-oriented metrics is practically attractive. The QDR loss is a plausible AUROC surrogate, the ablation study supports both main components, and the availability of code/checkpoints is a strength. However, the central evaluation claims currently hinge on an under-specified training/evaluation protocol and on point estimates without variability measures, so the significance cannot be fully assessed from the manuscript as written.

major comments (4)
  1. [§III-A, §III-C] The training labels are defined by a fixed rule: for each one-second GSC-V2 clip, the central 0.2–0.83 s is positive and the rest is negative. The paper does not state how the AVA-Speech and ACAM test segments are labeled for the frame/segment-level evaluation. If test segments are also scored positive only when the center of the window contains speech, then the temporal prior is not an artifact; if test labels are boundary-based, the model could satisfy the training objective largely by predicting “speech near the center of the window,” and the reported gains, especially the ACAM F2 jump (0.65→0.92), would be inflated by this prior. This is load-bearing and must be resolved explicitly. Please state the exact rule used to assign labels to test windows and provide a control experiment, e.g., training with randomized speech positions inside the clip or with labels derived from actual word
  2. [Tables I–IV] All reported results are point estimates without error bars, confidence intervals, or significance tests. The ACAM F2-Score difference between TinyVAD (0.65) and SincQDR-VAD (0.92) is unusually large relative to the AUROC difference (0.96 vs 0.97), which suggests that F2 at a fixed 0.5 threshold may be sensitive to score calibration, post-processing, or a few boundary decisions. At minimum, report multiple seeds with standard deviations, and provide F2 as a function of threshold or a precision-recall summary for the ACAM results. Without this, the headline 41.5% relative F2 gain is not yet substantiated.
  3. [§III-C] The post-processing step (“median smoothing filter using an 87.5% overlap between adjacent segments”) is described only in one sentence. The smoothing window length, whether it is causal, and its effect on AUROC/F2 are not reported. Since the claimed improvements are partly in recall-oriented metrics on noisy data, the interaction between median smoothing and the 0.5 threshold should be quantified, and the same post-processing should be shown to be applied identically to all baselines.
  4. [§I, §II-D] The claim of being “the first to formalize this notion for use on the VAD task” is overstated in view of reference [42], which already applies AUC optimization to VAD. The QDR loss in Eq. (5) is a squared-hinge pairwise ranking loss, a known surrogate for AUROC. The novelty should be framed as an adaptation and empirical demonstration for lightweight VAD, not as the first ranking loss for the task.
minor comments (6)
  1. [Eq. (5)] The normalization “1/|P| 1/|N|” can be ambiguous when either set is empty in a mini-batch; specify the handling of empty positive/negative sets. Also define the value of sinc(0) explicitly in Eq. (2).
  2. [§III-A] The sentence “For testing, a 0.15-second stride was applied…” is unclear about whether it refers only to the SCF test set or also to AVA-Speech/ACAM. Please clarify the segmentation and labeling pipeline for each benchmark.
  3. [Table II] The “Avg.” column is not defined; state whether it is the unweighted mean over the five SNR levels and report the standard deviation across SNR conditions.
  4. [Fig. 4] The prediction plots lack axis labels, time scale, and ground-truth boundaries. Adding these would make the qualitative robustness claim more interpretable.
  5. [§IV-E] The analysis of learned Sinc filters is descriptive. A quantitative comparison, e.g., overlap between learned filter bands and speech-relevant frequency ranges, would strengthen the claim that the front-end is noise-robust.
  6. [General] The URL in the abstract/code statement contains a space (“SincQDR-V AD”); please replace with “SincQDR-VAD” consistently. Some figure color schemes (red/green) may not be accessible; consider adding markers or a more colorblind-safe palette.

Circularity Check

0 steps flagged

No material circularity; only minor non-load-bearing self-citations in the sinc front-end.

full rationale

After walking the derivation chain, I find no circular step that reduces a prediction to an input or relies on a load-bearing self-citation. The QDR loss (Eq. 5) is a squared-hinge pairwise function defined in the paper; it is optimized on the SCF training set whose labels are assigned by the fixed temporal rule in Section III-A, while all reported AUROC and F2-Score numbers are computed on independently annotated test sets (AVA-Speech, noisy AVA-Speech, ACAM). Thus the evaluation is not equivalent by construction to the training objective. The sinc front-end is presented as an adaptation of the standard SincNet filterbank [37]; the only self-citations ([39], [40], with overlapping authors) are background references for learnable sinc-convolution filters and are not used to justify any theorem, uniqueness claim, or predicted value. The margin m=1.0 and mixing weight lambda=0.25 are hyperparameters chosen by validation, not fitted parameters renamed as predictions. The synthetic center-anchored training labels could create a temporal-prior artifact, but that is an external-validity risk, not a circularity: the test labels are not defined by the model's training rule.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim rests on two tuned hyperparameters (m, lambda), one hand-chosen training-label rule, and three domain assumptions about the ranking-loss surrogate, label fidelity, and sinc filter robustness. No new physical entities are introduced. The claims are evaluated on external benchmarks, but the training signal itself is synthetic.

free parameters (3)
  • QDR margin m (Eq. 5) = 1.0
    Set to 1.0 for all experiments; controls the minimum score gap between speech and non-speech frames in the ranking loss. Hand-chosen, not derived.
  • Loss weight lambda (Eq. 6) = 0.25
    Tuned 'based on empirical validation' (Section III-B); balances QDR and BCE terms. Directly controls the contribution of the novel loss to the final objective.
  • Active speech window in SCF training clips = 0.2 s to 0.83 s within a 1 s clip
    Hand-chosen labeling rule (Section III-A) that defines frame-level ground truth for all training. Determines what the model learns as speech vs non-speech boundaries.
axioms (4)
  • domain assumption The QDR loss, a squared hinge on pairwise score differences, is a valid surrogate for AUROC and improves AUROC when optimized on training data.
    Invoked in Section II-D as the design premise; the paper relies on the standard AUC surrogate literature ([41], [42]) without proving the connection for the VAD setting.
  • domain assumption Frame-level speech/non-speech labels in AVA-Speech and ACAM are accurate, consistent, and comparable across all compared models.
    Used in Section III-C and Section IV to adjudicate model quality; no label-noise or human-annotation uncertainty analysis is provided.
  • domain assumption A truncated, Hamming-windowed sinc filterbank with learnable cutoffs and gains (Eqs. 2-4) extracts features that are more noise-robust than mel filterbank features for VAD.
    Central architectural premise of Section II-B; supported only by the in-paper ablation (Table IV), not by an external independent result specific to VAD.
  • standard math The sinc function, convolution, and log-energy operations are standard signal-processing identities.
    Standard definitions assumed without proof (Eqs. 1-3).

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SincQDR-VAD: A Noise-Robust Voice Activity Detection Framework Leveraging Learnable Filters and Ranking-Aware Optimization." pith.science (2026). https://pith.science/paper/E4EUI62R

@misc{pith2026250820885,
  author       = {Pith},
  title        = {Pith review of: SincQDR-VAD: A Noise-Robust Voice Activity Detection Framework Leveraging Learnable Filters and Ranking-Aware Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E4EUI62R}},
  note         = {Machine review of arXiv:2508.20885}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Voice activity detection (VAD) is essential for speech-driven applications, but remains far from perfect in noisy and resource-limited environments. Existing methods often lack robustness to noise, and their frame-wise classification losses are only loosely coupled with the evaluation metric of VAD. To address these challenges, we propose SincQDR-VAD, a compact and robust framework that combines a Sinc-extractor front-end with a novel quadratic disparity ranking loss. The Sinc-extractor uses learnable bandpass filters to capture noise-resistant spectral features, while the ranking loss optimizes the pairwise score order between speech and non-speech frames to improve the area under the receiver operating characteristic curve (AUROC). A series of experiments conducted on representative benchmark datasets show that our framework considerably improves both AUROC and F2-Score, while using only 69% of the parameters compared to prior arts, confirming its efficiency and practical viability.

Figures

Figures reproduced from arXiv: 2508.20885 by Berlin Chen, Chien-Chun Wang, En-Lun Yu, Jeih-Weih Hung, Shih-Chieh Huang.

Figure 1
Figure 1. Figure 1: The proposed SincQDR-VAD framework consists of a feature extractor with learnable sinc filters parameterized by low/high cutoff frequencies and a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of the learned cutoff frequency ranges and corresponding [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Output plots of SincQDR-VAD and TinyVAD on the noisy variant of [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment

    eess.AS 2026-07 conditional novelty 6.0

    kiloVAD, a 2.1k-parameter causal CNN VAD on standard Mel features, reaches 0.850 AUC on AVA-Speech and beats standard QAT by 1–4% at INT4 via angle-based self-distillation.

Reference graph

Works this paper leans on

59 extracted references · 52 canonical work pages · cited by 1 Pith paper · 9 internal anchors

  1. [1]

    Speaker diarization with plda i-vector scoring and unsupervised calibration,

    G. Sell and D. Garcia-Romero, “Speaker diarization with plda i-vector scoring and unsupervised calibration,” in Proc. IEEE SLT , 2014

    work page 2014

  2. [2]

    Recurrent neural networks for voice activity detection,

    T. Hughes and K. Mierle, “Recurrent neural networks for voice activity detection,” in Proc. ICASSP, 2013

    work page 2013

  3. [3]

    Deep belief networks based voice activity detection,

    X.-L. Zhang and J. Wu, “Deep belief networks based voice activity detection,” IEEE Transactions on Audio, Speech, and Language Pro- cessing, vol. 21, no. 4, pp. 697–710, 2013

    work page 2013

  4. [4]

    A statistical model-based voice activity detection,

    J. Sohn, N. S. Kim, and W. Sung, “A statistical model-based voice activity detection,” IEEE Signal Processing Letters , vol. 6, no. 1, pp. 1–3, 1999

    work page 1999

  5. [5]

    Real-life voice activity detection with LSTM Recurrent neural networks and an appli- cation to Hollywood movies,

    F. Eyben, F. Weninger, S. Squartini, and B. Schuller, “Real-life voice activity detection with LSTM Recurrent neural networks and an appli- cation to Hollywood movies,” in Proc. ICASSP, 2013

    work page 2013

  6. [6]

    Improvements to deep convolutional neural networks for LVCSR

    T. N. Sainath, B. Kingsbury, A.-r. Mohamed, G. E. Dahl, G. Saon, H. Soltau, T. Beran, A. Y . Aravkin, and B. Ramabhadran, “Improvements to deep convolutional neural networks for LVCSR,” in Arxiv preprint arXiv:1309.1501, 2013

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  7. [7]

    V oice activity detection for transient noisy environment based on diffusion nets,

    A. Ivry, B. Berdugo, and I. Cohen, “V oice activity detection for transient noisy environment based on diffusion nets,” IEEE Journal of Selected Topics in Signal Processing , vol. 13, no. 2, pp. 254–264, 2019

    work page 2019

  8. [8]

    Voice Activity Detection in presence of background noise using EEG

    G. Krishna, C. Tran, M. Carnahan, Y . Han, and A. H. Tewfik, “V oice activity detection in presence of background noise using EEG,” in Arxiv preprint arXiv:1911.04261, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1911

  9. [9]

    A bin encoding training of a spiking neural network based voice activity detection,

    G. Dellaferrera, F. Martinelli, and M. Cernak, “A bin encoding training of a spiking neural network based voice activity detection,” in Proc. ICASSP, 2020

    work page 2020

  10. [10]

    V oice activity detection in the wild via weakly supervised sound event detection,

    H. Dinkel, Y . Chen, M. Wu, and K. Yu, “V oice activity detection in the wild via weakly supervised sound event detection,” in Proc. Interspeech, 2020

    work page 2020

  11. [11]

    Spiking neural networks trained with backpropagation for low power neuromorphic implementation of voice activity detection,

    F. Martinelli, G. Dellaferrera, P. Mainar, and M. Cernak, “Spiking neural networks trained with backpropagation for low power neuromorphic implementation of voice activity detection,” in Proc. ICASSP, 2020

    work page 2020

  12. [12]

    End-to-end domain-adversarial voice activity detection,

    M. Lavechin, M.-P. Gill, R. Bousbib, H. Bredin, and L. P. Garcia- Perera, “End-to-end domain-adversarial voice activity detection,” in Proc. Interspeech, 2020

    work page 2020

  13. [13]

    End-to-end automatic speech recognition integrated with CTC-based voice activity detection,

    T. Yoshimura, T. Hayashi, K. Takeda, and S. Watanabe, “End-to-end automatic speech recognition integrated with CTC-based voice activity detection,” in Proc. ICASSP, 2020

    work page 2020

  14. [14]

    V oice activity detection in the wild: A data-driven approach using teacher-student training,

    H. Dinkel, S. Wang, X. Xu, M. Wu, and K. Yu, “V oice activity detection in the wild: A data-driven approach using teacher-student training,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 29, pp. 1542–1555, 2021

    work page 2021

  15. [15]

    Improvement of noise-robust single-channel voice activity detection with spatial pre-processing,

    M. Væhrens, A. J. Fuglsig, A. P. Jacobsen, N. A. Rasmussen, V . M. Nissen, J. R. Hejslet, and Z.-H. Tan, “Improvement of noise-robust single-channel voice activity detection with spatial pre-processing,” in Proc. Interspeech, 2021

    work page 2021

  16. [16]

    A lightweight framework for online voice activity detection in the wild,

    X. Xu, H. Dinkel, M. Wu, and K. Yu, “A lightweight framework for online voice activity detection in the wild,” in Proc. Interspeech, 2021

    work page 2021

  17. [17]

    Cross-domain Voice Activity Detection with Self-Supervised Representations

    S. Alisamir, F. Ringeval, and F. Portet, “Cross-domain voice activ- ity detection with self-supervised representations,” in Arxiv preprint arXiv:2209.11061, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  18. [18]

    Adversarial Multi-Task Deep Learning for Noise-Robust Voice Activity Detection with Low Algorithmic Delay

    C. M. Larsen, P. Koch, and Z.-H. Tan, “Adversarial multi-task deep learning for noise-robust voice activity detection with low algorithmic delay,” in Arxiv preprint arXiv:2207.01691 , 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  19. [19]

    BC-VAD: A Robust Bone Conduction Voice Activity Detection

    N. Polvani, D. Ronssin, and M. Cernak, “BC-V AD: A robust bone conduction voice activity detection,” inArxiv preprint arXiv:2212.02996, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [20]

    Unsupervised voice activity detection by modeling source and system information using zero frequency filtering,

    E. Sarkar, R. Prasad, and M. Magimai. Doss, “Unsupervised voice activity detection by modeling source and system information using zero frequency filtering,” in Proc. Interspeech, 2022

    work page 2022

  21. [21]

    CNN self-attention voice activity detector

    A. Sofer and S. E. Chazan, “CNN self-attention voice activity detector,” in Arxiv preprint arXiv:2203.02944 , 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  22. [22]

    V oice activity detection (V AD) in noisy environments,

    J. Ball, “V oice activity detection (V AD) in noisy environments,” inArxiv preprint arXiv:2312.05815, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Semantic V AD: Low-latency voice activity detection for speech interaction,

    M. Shi, Y . Shu, L. Zuo, Q. Chen, S. Zhang, J. Zhang, and L.-R. Dai, “Semantic V AD: Low-latency voice activity detection for speech interaction,” in Proc. Interspeech, 2023

    work page 2023

  24. [24]

    Real-time causal spectro-temporal voice activity detection based on convolutional encoding and residual decoding,

    J. Wang, J. Zhang, and L.-R. Dai, “Real-time causal spectro-temporal voice activity detection based on convolutional encoding and residual decoding,” in Proc. Interspeech, 2023

    work page 2023

  25. [25]

    CLIP-VAD: Exploiting Vision-Language Models for Voice Activity Detection

    A. Appiani and C. Beyan, “CLIP-V AD: Exploiting vision-language models for voice activity detection,” in Arxiv preprint arXiv:2410.14509, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    A Real-Time Voice Activity Detection Based On Lightweight Neural

    J. Jia, P. Zhao, and D. Wang, “A real-time voice activity detection based on lightweight neural,” in Arxiv preprint arXiv:2405.16797 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [27]

    A Transformer-based voice activity detector,

    B. Karan, J. Jansen Van V ¨uren, F. De Wet, and T. Niesler, “A Transformer-based voice activity detector,” in Proc. Interspeech, 2024

    work page 2024

  28. [28]

    Robust voice activity detection using locality- sensitive hashing and residual frequency-temporal attention,

    S. Li, P. Zhang, and Y . Li, “Robust voice activity detection using locality- sensitive hashing and residual frequency-temporal attention,” in Proc. Interspeech, 2024

    work page 2024

  29. [29]

    Channel- combination algorithms for robust distant voice activity and overlapped speech detection,

    T. Mariotte, A. Larcher, S. Montr ´esor, and J.-H. Thomas, “Channel- combination algorithms for robust distant voice activity and overlapped speech detection,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 1859–1872, 2024

    work page 2024

  30. [30]

    sV AD: A robust, low-power, and light-weight voice activity detection with spiking neural networks,

    Q. Yang, Q. Liu, N. Li, M. Ge, Z. Song, and H. Li, “sV AD: A robust, low-power, and light-weight voice activity detection with spiking neural networks,” in Proc. ICASSP, 2024

    work page 2024

  31. [31]

    Robust speech activity detection in movie audio: Data resources and experimental evaluation,

    R. Hebbar, K. Somandepalli, and S. Narayanan, “Robust speech activity detection in movie audio: Data resources and experimental evaluation,” in Proc. ICASSP, 2019

    work page 2019

  32. [32]

    MarbleNet: Deep 1D time- channel separable convolutional neural network for voice activity de- tection,

    F. Jia, S. Majumdar, and B. Ginsburg, “MarbleNet: Deep 1D time- channel separable convolutional neural network for voice activity de- tection,” in Proc. ICASSP, 2021

    work page 2021

  33. [33]

    SG-V AD: Stochastic gates based speech activity detection,

    J. Svirsky and O. Lindenbaum, “SG-V AD: Stochastic gates based speech activity detection,” in Proc. ICASSP, 2023

    work page 2023

  34. [34]

    ResectNet: An efficient architecture for voice activity detection on mobile devices,

    O. K ¨op¨ukl¨u and M. Taseska, “ResectNet: An efficient architecture for voice activity detection on mobile devices,” in Proc. Interspeech, 2022

    work page 2022

  35. [35]

    Small-footprint convolutional neural network with reduced feature map for voice activity detection,

    H. Chae and S. Lee, “Small-footprint convolutional neural network with reduced feature map for voice activity detection,” inProc. ICASSP, 2024

    work page 2024

  36. [36]

    A tutorial on the cross-entropy method,

    P.-T. de Boer, D. P. Kroese, S. Mannor, and R. Y . Rubinstein, “A tutorial on the cross-entropy method,” Annals of Operations Research, vol. 134, no. 1, pp. 19–67, 2005

    work page 2005

  37. [37]

    Speaker recognition from raw waveform with SincNet,

    M. Ravanelli and Y . Bengio, “Speaker recognition from raw waveform with SincNet,” in Proc. IEEE SLT , 2018

    work page 2018

  38. [38]

    Learning filterbanks from raw speech for phone recogni- tion,

    N. Zeghidour, N. Usunier, I. Kokkinos, T. Schatz, G. Synnaeve, and E. Dupoux, “Learning filterbanks from raw speech for phone recogni- tion,” in Proc. ICASSP, 2018

    work page 2018

  39. [39]

    What do neural networks listen to? Exploring the crucial bands in speech enhancement using sinc- convolution,

    K.-H. Ho, J.-w. Hung, and B. Chen, “What do neural networks listen to? Exploring the crucial bands in speech enhancement using sinc- convolution,” in Proc. ICASSP, 2024

    work page 2024

  40. [40]

    Speaker conditional sinc-extractor for personal V AD,

    E.-L. Yu, K.-H. Ho, J.-w. Hung, S.-C. Huang, and B. Chen, “Speaker conditional sinc-extractor for personal V AD,” inProc. Interspeech, 2024

    work page 2024

  41. [41]

    Benchmarking deep AUROC opti- mization: Loss functions and algorithmic choices,

    D. Zhu, X. Wu, and T. Yang, “Benchmarking deep AUROC opti- mization: Loss functions and algorithmic choices,” in Arxiv preprint arXiv:2203.14177, 2022

    Pith/arXiv arXiv 2022

  42. [42]

    AUC optimization for deep learning-based voice activity detection,

    X.-L. Zhang and M. Xu, “AUC optimization for deep learning-based voice activity detection,”EURASIP Journal on Audio, Speech, and Music Processing, vol. 2022, no. 1, p. 27, 2022

    work page 2022

  43. [43]

    A V A- speech: A densely labeled dataset of speech activity in movies,

    S. Chaudhuri, J. Roth, D. P. W. Ellis, A. Gallagher, L. Kaver, R. Marvin, C. Pantofaru, N. Reale, L. G. Reid, K. Wilson, and Z. Xi, “A V A- speech: A densely labeled dataset of speech activity in movies,” in Proc. Interspeech, 2018

    work page 2018

  44. [44]

    V oice activity detection using an adaptive context attention model,

    J. Kim and M. Hahn, “V oice activity detection using an adaptive context attention model,” IEEE Signal Processing Letters , vol. 25, no. 8, pp. 1181–1185, 2018

    work page 2018

  45. [45]

    Some windows with very good sidelobe behavior,

    A. Nuttall, “Some windows with very good sidelobe behavior,” IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 29, no. 1, pp. 84–91, 1981

    work page 1981

  46. [46]

    Improved regularization of convolutional neural networks with cutout,

    T. DeVries and G. W. Taylor, “Improved regularization of convolutional neural networks with cutout,” in Arxiv preprint arXiv:1708.04552, 2017

    Pith/arXiv arXiv 2017

  47. [47]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” in Arxiv preprint arXiv:1803.01271, 2018

    Pith/arXiv arXiv 2018

  48. [48]

    CSPNet: A new backbone that can enhance learning capability of CNN,

    C.-Y . Wang, H.-Y . M. Liao, I.-H. Yeh, Y .-H. Wu, P.-Y . Chen, and J.-W. Hsieh, “CSPNet: A new backbone that can enhance learning capability of CNN,” in Proc. CVPR, 2020

    work page 2020

  49. [49]

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

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

    Pith/arXiv arXiv 2018

  50. [50]

    Freesound technical demo,

    F. Font, G. Roma, and X. Serra, “Freesound technical demo,” in Proceedings of the 21st ACM international conference on Multimedia , 2013, pp. 411–412

    work page 2013

  51. [51]

    ESC: Dataset for environmental sound classification,

    K. J. Piczak, “ESC: Dataset for environmental sound classification,” in Proceedings of the 23rd ACM international conference on Multimedia , 2015

    work page 2015

  52. [52]

    An overview of gradient descent optimization algorithms,

    S. Ruder, “An overview of gradient descent optimization algorithms,” in Arxiv preprint arXiv:1609.04747 , 2016

    Pith/arXiv arXiv 2016

  53. [53]

    Bag of tricks for image classification with convolutional neural networks,

    T. He, Z. Zhang, H. Zhang, Z. Zhang, J. Xie, and M. Li, “Bag of tricks for image classification with convolutional neural networks,” in Proc. CVPR, 2019

    work page 2019

  54. [54]

    The use of the area under the ROC curve in the evaluation of machine learning algorithms,

    A. P. Bradley, “The use of the area under the ROC curve in the evaluation of machine learning algorithms,” Pattern Recognition, vol. 30, no. 7, pp. 1145–1159, 1997

    work page 1997

  55. [55]

    A hybrid CNN-BiLSTM voice activity detector,

    N. Wilkinson and T. Niesler, “A hybrid CNN-BiLSTM voice activity detector,” in Proc. ICASSP, 2021

    work page 2021

  56. [56]

    ADA-V AD: Unpaired adversarial domain adaptation for noise-robust voice activity detection,

    T. Kim, J. Chang, and J. H. Ko, “ADA-V AD: Unpaired adversarial domain adaptation for noise-robust voice activity detection,” in Proc. ICASSP, 2022

    work page 2022

  57. [57]

    NAS-V AD: Neural architecture search for voice activity detection,

    D. Rho, J. Park, and J. H. Ko, “NAS-V AD: Neural architecture search for voice activity detection,” in Proc. Interspeech, 2022

    work page 2022

  58. [58]

    On training targets for noise-robust voice activity detection

    S. Braun and I. Tashev, “On training targets for noise-robust voice activity detection,” in Arxiv preprint arXiv:2102.07445 , 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  59. [59]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” in Arxiv preprint arXiv:2312.00752 , 2023

    Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.