Pith. sign in

REVIEW 4 major objections 5 minor 37 references

A 2.1-thousand-parameter VAD matches MarbleNet's accuracy at 43x fewer parameters under strictly causal, 200 ms evaluation.

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 ↗

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.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A solid tiny-VAD engineering paper with a genuinely new QAT loss and honest failure reporting; just don't read the abstract's SOTA claim without the n=8 caveat. the 4 major comments →

arxiv 2607.25870 v1 pith:6MJHJOAQ submitted 2026-07-28 eess.AS cs.LG

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

Stephen Bauer , Sheila Seidel , Shanza Iftikhar , Scott Veidenheimer , Gorkem Ulkar This is my paper
classification eess.AS cs.LG
keywords voice activity detectionedge deploymentstructured pruningquantization-aware trainingself-distillationconvolutional neural networksMel spectrogramlow-latency inference
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 argues that a voice activity detector built strictly from standard, widely supported components—fixed Mel spectrogram input, convolutional layers only, and a 200 ms causal evaluation window—can match the accuracy of much larger models. It claims kiloVAD, at 2.1 thousand parameters, achieves 0.850 AUC on AVA-Speech, the same score as the 91-thousand-parameter MarbleNet but with 43x fewer parameters and 3x lower latency. The paper also introduces two compression techniques: per-layer structured pruning guided by multi-objective search plus self-distillation, and an angle-aware self-distilling quantization-aware training loss that improves INT4 accuracy by 1–4% over standard QAT. If these claims hold, always-on edge devices can run a VAD on microcontrollers using only operators that are already supported by embedded runtimes.

Core claim

In the paper's own terms, the central discovery is that extreme compression and deployment compatibility don't have to trade away VAD accuracy when the architecture is co-designed for pruning and quantization. The design separates Mel resolution from channel width via a 1x1 adapter, uses global average pooling to decouple parameter count from input length, and normalizes each frequency bin per-frame. The resulting CNN-only model, unpruned at 81.1k parameters, scores 0.862 AUC; pruning to 2.1k parameters with per-layer ratios and self-distillation retains 0.850 AUC, within 1.3% of the baseline while cutting MACs from 1.7M to 44k. Under the same causal protocol, this matches MarbleNet's 0.850

What carries the argument

The load-bearing mechanism is a compression pipeline: (1) per-layer structured pruning via dependency-graph analysis with layer-specific ratios found by multi-objective search (minimizing false positives at fixed true-positive rate and parameter count), followed by fine-tuning with self-distillation from the unpruned model; (2) an angle-aware self-distilling QAT loss (Eqs. 1–2) that, with a frozen full-precision classifier, pushes each penultimate feature toward its own class weight vector in cosine space and pushes it away from the other class's vector, explicitly countering the angular distortion introduced by low-bit quantization. The architecture itself—a 1x1 adapter, depthwise separable

Load-bearing premise

The claim that 2.1k parameters is a reliable operating point rests on the assumption that the two training runs that collapsed under the transferred pruning ratios are rare, avoidable failures rather than a signal that this configuration sits at the edge of a stability cliff.

What would settle it

Retrain the 2.1k-parameter configuration on, say, 30 random seeds with the same pruning ratios and protocol; if more than ~10% of runs collapse to unusable AUC, or if the mean AUC including those failures drops below 0.850, the central 'matches MarbleNet' claim would not hold as a system-level guarantee. Alternatively, run the same per-layer pruning + self-distillation pipeline on a different VAD dataset (e.g., a far-field noisy set) and check whether the 2.1k configuration stays within 1.3% of the unpruned baseline.

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

If this is right

  • A causal, 200 ms-context VAD with only 2.1k parameters can be deployed on microcontroller-class hardware using standard Mel frontends and CNN ops, a capability current compact VADs lack.
  • The per-layer pruning plus self-distillation recipe is claimed to generalize across compression levels and seeds (10 independent models, 8 surviving), offering a path to hit arbitrary parameter budgets.
  • Angle-aware QAT gives a usable INT4 regime: the 2.1k model goes from 0.693 AUC (standard QAT) to 0.719 AUC (angle-aware), and INT8 stays lossless at 0.851.
  • Extending context to 360 ms yields 0.872 AUC, which exceeds AtomicVAD's causal 0.869 with 630 ms context—suggesting latency can be traded for accuracy without leaving standard operators.

Where Pith is reading between the lines

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

  • Because the angle-aware QAT loss only requires a frozen full-precision classifier, the same training recipe could be lifted onto other tiny classification models, such as keyword spotting, where angular error dominates at low bit-widths—a testable extension the paper does not run.
  • The pruning-ratio transfer protocol implies a single multi-objective search per architecture might serve many deployments; if the two collapsed seeds instead reflect run-level instability, a production guardrail would be to verify any pruned configuration on multiple initializations before shipping.
  • The 200 ms performance plateau aligns with the 4–5 Hz syllable rate of speech, suggesting an adaptive context length could trade latency against accuracy in very noisy domains—an idea the paper leaves implicit.
  • A direct follow-up could test whether per-frame normalization plus pruning generalizes to far-field or multilingual speech, where the frozen teacher's class prototypes may be less stable.
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 / 5 minor

Summary. kiloVAD is a CNN-only voice activity detector designed for edge deployment on standard Mel spectrogram features with a 200 ms input context. The paper contributes three main elements: (1) a per-layer structured pruning strategy with a multi-objective Optuna search and self-distillation from the unpruned model, (2) an angle-aware self-distilling quantization-aware training (QAT) method that freezes full-precision classifier weight vectors as angular prototypes, and (3) an evaluation protocol that scores each frame independently under strictly causal conditions on AVA-Speech. The central empirical claims are that the pruned 2.1 k-parameter model reaches 0.850 AUC (F1 0.783), matching MarbleNet at 43× fewer parameters and 3× lower latency; the full 81 k model reaches 0.862 AUC, and with 360 ms context reaches 0.872 AUC; INT8 post-training quantization is lossless; and angle-aware QAT improves INT4 AUC over standard STE-based QAT by 1–4% (0.811 vs. 0.800 for the 10 k model and 0.719 vs. 0.693 for the 2.1 k model). The authors emphasize deployability: no learnable filterbanks, no recurrent layers, no non-standard activations, and compatibility with TensorFlow Lite for Microcontrollers. Code, pretrained weights, and a web demo are promised.

Significance. If the results are robust, this is a practically valuable contribution: it demonstrates that a standard-Mel, CNN-only, TFLM-compatible model can match the accuracy of much larger or architecturally specialized compact VADs, and the compression pipeline (multi-objective per-layer pruning plus self-distillation plus angle-based QAT) is reusable for other tiny audio models. The paper's strengths include the explicit choice of strictly causal per-frame evaluation to avoid the inflated numbers produced by non-causal sliding-window protocols, a clear comparison table that separates causal from non-causal results, and a commendable disclosure that 2 of 10 runs collapsed when the pruning configuration was transferred. However, the central 0.850 AUC headline is an n=8 conditional estimate, and the QAT comparison appears to rest on single runs; both issues need to be resolved before the claims as written can be accepted.

major comments (4)
  1. [§5 'Structured pruning' and Table 1] The headline result 0.850±0.007 is computed over n=8 runs because 2 of 10 independent training runs exhibited layer collapse after transferring pruning ratios that were optimized on a single seed. This makes the central claim an estimate of performance conditional on successful training, not the expected outcome of the published recipe. A practitioner has a 20% chance of not obtaining the reported model on a single run. The paper discloses this, which is good, but the abstract and Table 1 do not carry the caveat. Please report all-seed statistics with collapsed runs scored as failures (or a clear deterministic mitigation that avoids collapse), give the failure rate in the abstract, and report the selected per-layer pruning ratios so the transfer protocol is reproducible.
  2. [§5 'Quantization' and Figure 4] The QAT comparison (0.811 vs. 0.800 for 10 k; 0.719 vs. 0.693 for 2.1 k) appears to be based on a single training run for each configuration. No confidence intervals or significance tests are provided. The claimed 1–4% relative improvement is a central method contribution, and its magnitude is comparable to the seed-to-seed variability seen elsewhere in the paper (e.g., ±0.007 AUC for the pruned models). Please provide multi-seed results with error bars, or explicitly state that the improvement was observed in a single run and temper the claim accordingly.
  3. [Abstract and §5 'Comparison to prior work'] The abstract states that the 0.850 AUC result with 2.1 k parameters establishes 'a new state of the art for causal, deployment-ready VAD,' but Table 1 shows AtomicVAD's causal AUC of 0.869, and the paper's own 0.872 AUC result is for the 81 k full model at 360 ms context, not for the 2.1 k model. If 'deployment-ready' is intended to exclude AtomicVAD because of its GGCU activation or raw-audio frontend, that comparison class must be defined explicitly and the claim should be aligned with the actual 0.872 result or relaxed to, e.g., 'state of the art among standard-operator causal VADs.' As written, the abstract's claim is not supported by the table.
  4. [§3.1–3.2 and §5] The paper does not report the actual per-layer pruning ratios selected by the Optuna search, the repel weight λ in Eq. (1), or the soft-to-hard annealing schedule used for the reported QAT results. These are free parameters tuned on validation data and are necessary for reproducibility of the 0.850 and 0.872 results. Please include them in a table or the supplementary material.
minor comments (5)
  1. [Figure 4] The caption says 'open vs. filled markers indicate 2.1 k vs. 10 k parameters,' but the legend labels methods. In grayscale printing the marker shape is the only cue; please add an explicit legend entry for parameter count.
  2. [Table 1] R1–R4 are defined in the header but not in the caption. Add a caption sentence such as 'R1: standard Mel frontend; R2: portable ops; R3: low latency; R4: causal evaluation.'
  3. [§2.1] The text says normalization is computed over '21 time steps for 64 mel bins.' Please specify the frame shift/hop length and window size so a reader can reproduce the 200 ms context yielding 21 time steps.
  4. [§5 'Quantization'] The 1–4% improvement is reported as 'relative'; state this explicitly in the text and in the abstract to avoid confusion with absolute AUC differences.
  5. [Eq. (2)] The hinge-style repulsion term has no margin parameter; it activates as soon as any non-target cosine similarity exceeds zero. This may be intentional, but the choice of zero margin and any sensitivity to it should be mentioned in the text.

Circularity Check

0 steps flagged

No circularity: the central claims are empirical evaluations against an external benchmark, with no self-citations and no fitted inputs renamed as predictions.

full rationale

The paper's derivation chain is self-contained in the relevant sense. kiloVAD is trained on LibriSpeech with noise augmentation and evaluated on AVA-Speech, a held-out external benchmark; the central accuracy claims are therefore not constructed from the evaluation data. The pruning ratios are tuned on validation data (via Optuna multi-objective search) and then transferred to independent seeds; this is standard hyperparameter selection, not a fitted parameter renamed as a prediction. The disclosed 2/10 layer-collapse exclusion (Section 5, Table 1 note: "Pruning ratios were optimized on a single seed... 2 failed under this aggressive compression due to layer collapse, yielding n=8 for confidence intervals in Table 1") is a robustness/selection limitation — the reported 0.850 AUC is conditional on successful training — but the paper states it explicitly, and it does not make the AUC value equal to the training objective or to a fitted value. The QAT loss (Eq. 1–2) is a new training objective compared against the standard STE QAT baseline; it is not defined in terms of the reported AUC, and the comparison is an empirical result rather than a tautology. There are no self-citations and no imported uniqueness theorems. Hidden hyperparameter details or the n=8 confidence interval are correctness/robustness concerns, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claims are empirical; the unstated tuning parameters above are the main hidden degrees of freedom. No new physical or architectural entities are introduced—the angle-QAT is a loss function, not a new component.

free parameters (4)
  • Per-layer structured pruning ratios = not reported (produce 2.1k, 1.1k, 622-param models in Fig. 3)
    Chosen by Optuna multi-objective search over validation FPR and parameter count (Sec. 3.1); the headline 2.1k configuration is defined by these undisclosed ratios.
  • Repel weight λ in Eq. (1) = not reported
    Weights the non-target class repulsion in the angle QAT loss; the claimed 1–4% INT4 gain depends on this value and it is not stated.
  • Soft-to-hard annealing schedule = not reported
    The quantization path uses annealing 'inspired by [23]' (Sec. 3.2) but the schedule (temperature, epochs) is not specified.
  • Self-distillation loss weight (CE vs KL) = not reported
    Post-pruning fine-tuning uses a 'weighted combination' of cross-entropy and KL (Sec. 3.1); the weight is not given.
axioms (4)
  • ad hoc to paper Frozen full-precision classifier weight vectors are correct angular targets for INT4-quantized penultimate features
    Equations (1)–(2) assume W_FP rows are fixed prototypes and that pulling quantized features to them corrects quantization-induced angular error; no independent evidence is provided that this geometry is the right objective at 4-bit.
  • domain assumption Per-window zero-mean/unit-variance normalization over 200 ms is causal and preserves discriminative spectral patterns
    Sec. 2.1 'Amplitude-agnostic preprocessing'; assumes the normalization is computed only over the available causal window and that removing absolute energy does not hurt noisy-speech detection.
  • domain assumption Frame-level independent classification (no temporal smoothing or future context) is the right causal evaluation protocol
    Sec. 4 'Evaluation'; deployed streaming VADs typically add smoothing/hangover, so this protocol is conservative but may not reflect final system behavior; comparisons to non-causal prior work are indirect.
  • domain assumption Quoted prior-art AUC values (esp. AtomicVAD 0.869 causal from [10]) are accurate
    Table 1's comparison and the SOTA argument rely on numbers taken from the cited papers without re-running those systems under the authors' protocol.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment." pith.science (2026). https://pith.science/paper/6MJHJOAQ

@misc{pith2026260725870,
  author       = {Pith},
  title        = {Pith review of: VAD to the Bone: Ultra-Tiny Speech Activity Detection for Edge Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6MJHJOAQ}},
  note         = {Machine review of arXiv:2607.25870}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Voice activity detection (VAD) triggers downstream speech processing in always-on systems under strict memory, latency, and compute constraints. Recent compact models report strong accuracy but rely on components that are not widely supported: learnable filterbanks, recurrent layers, or non-causal post-processing. We propose kiloVAD, designed for embedded inference using standard Mel features, CNN-only layers, and tunable context/spectral parameters. We introduce per-layer structured pruning with self-distillation and angle-based quantization-aware training (QAT) that outperforms standard QAT by 1-4%. Evaluated per-frame under causal conditions, kiloVAD achieves 0.850 AUC on AVA-Speech with 2.1 k parameters and 200 ms context, establishing a new state of the art for causal, deployment-ready VAD.

Figures

Figures reproduced from arXiv: 2607.25870 by Gorkem Ulkar, Scott Veidenheimer, Shanza Iftikhar, Sheila Seidel, Stephen Bauer.

Figure 1
Figure 1. Figure 1: AVA-Speech AUC vs. input context length. Error bars show 95% confidence intervals (n=10 seeds). Performance plateaus near 200 ms. model-size reduction to suit application requirements. • A novel angle-aware self-distilling QAT approach that out￾performs standard QAT by 1–4% under INT4 quantization. • Analysis of design tradeoffs including Mel bins, input con￾text, and compression ratio versus performance. … view at source ↗
Figure 3
Figure 3. Figure 3: AVA-Speech AUC vs. parameter count for pruned models. Per-layer pruning outperforms uniform global prun￾ing, especially at low parameter counts. Knowledge distillation (KD) improves performance at all compression levels. Labeled points: 622 parameters retains 0.831 AUC; 2.1 k parameters achieves 0.851 AUC, within 1.3% of baseline. 4. Experimental Setup Training data. We train on LibriSpeech train-clean-100… view at source ↗
Figure 4
Figure 4. Figure 4: AVA-Speech AUC for FP32, INT8, and INT4 models. Colors denote quantization method; open vs. filled markers in￾dicate 2.1 k vs. 10 k parameters. 6. Discussion We presented kiloVAD, a VAD co-designed from the ground up for aggressive compression and microcontroller deployment. Our architecture, featuring a 1×1 adapter layer to decouple in￾put resolution from internal channels, global average pooling for cont… view at source ↗

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

37 extracted references · 5 linked inside Pith

  1. [1]

    reported values are not directly comparable because of two competing inference pro- tocols

    Introduction V oice activity detection (V AD) is a critical front-end compo- nent for speech processing systems deployed on edge devices. Before any downstream processing such as speech recognition, speaker verification, or keyword spotting can engage, a V AD must first determine whether speech is present, enabling the de- vice to keep power-hungry models...

    Pith/arXiv arXiv 2026

  2. [2]

    Model Architecture The kiloV AD architecture is optimized for cross-platform em- bedded support under standard operator constraints

    Deployment Oriented System Design 2.1. Model Architecture The kiloV AD architecture is optimized for cross-platform em- bedded support under standard operator constraints. Our model processes Mel spectrogram features through depthwise separa- ble convolutional blocks, global average pooling, and a linear classifier. We make four deployment-driven design c...

  3. [3]

    Compression Methods 3.1. Per-Layer Structured Pruning Strategy We use structured pruning viatorch-pruning[15], which constructs a dependency graph to identify parameter groups that must be pruned together to preserve functional structure. We apply magnitude-based pruning withℓ 2-norm importance to systematically remove channels contributing least to the o...

  4. [4]

    This mixture exposes the model to both environmental noise (wind) and diverse acoustic interference (DNS)

    Experimental Setup Training data.We train on LibriSpeech train-clean-100 [25] with three noise conditions: 25% clean speech, 25% mixed with synthetic wind noise [26] at−5dB SNR, and 50% mixed with DNS Challenge noise [27] at SNRs of{−10,−5,0,5,10}dB, with half of these samples including simulated room reverber- ation. This mixture exposes the model to bot...

  5. [5]

    Results Structured pruning.Figure 3 compares pruning strategies across compression levels. Per-layer pruning ratios, optimized via multi-objective search, consistently outperform uniform global pruning, particularly below 5 k parameters where global pruning degrades sharply. Notably, global pruning fails entirely below 2 k parameters (the curve terminates...

  6. [6]

    Discussion We presented kiloV AD, a V AD co-designed from the ground up for aggressive compression and microcontroller deployment. Our architecture, featuring a 1×1 adapter layer to decouple in- put resolution from internal channels, global average pooling for context-flexible inference, and a fully convolutional back- bone, was specifically engineered to...

  7. [7]

    Acknowledgments The authors thank Bibek Gupta and Rupesh Kumar for their help configuring on-premises machines and Azure training workflows, Adam Rowell for his help creating the online demo, and Kao Kitichotkul, Alessandro Ragano, and Henning Hase- mann for their feedback on the manuscript

  8. [8]

    The authors reviewed and validated all outputs and retain full responsibility for the work presented

    Generative AI Use Disclosure Claude (Anthropic) was used as a writing assistant for manuscript editing and as a programming aid during code de- velopment. The authors reviewed and validated all outputs and retain full responsibility for the work presented

  9. [9]

    Silero vad: Pre-trained enterprise-grade voice activity detector (vad) models,

    S. Team, “Silero vad: Pre-trained enterprise-grade voice activity detector (vad) models,” https://github.com/snakers4/silero-vad, 2025, version 6.0.0, released August 26, 2025

    2025

  10. [10]

    An efficient transformer-based model for voice activity detection,

    Y . Zhao and B. Champagne, “An efficient transformer-based model for voice activity detection,” in2022 IEEE 32nd Inter- national Workshop on Machine Learning for Signal Processing (MLSP), 2022, pp. 1–6

    2022

  11. [11]

    A transformer-based voice activity detector,

    B. Karan et al., “A transformer-based voice activity detector,” in Interspeech, vol. 2024, 2024, pp. 3819–3823

    2024

  12. [12]

    Pyannote.audio: Neural building blocks for speaker diarization,

    H. Bredin et al., “Pyannote.audio: Neural building blocks for speaker diarization,” inICASSP 2020 - 2020 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 7124–7128

    2020

  13. [13]

    Marblenet: Deep 1d time-channel separable convolutional neural network for voice activity detection,

    F. Jia, S. Majumdar, and B. Ginsburg, “Marblenet: Deep 1d time-channel separable convolutional neural network for voice activity detection,” 2021. [Online]. Available: https: //arxiv.org/abs/2010.13886

    Pith/arXiv arXiv 2021

  14. [14]

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

    H. Chae and S. Lee, “Small-footprint convolutional neural net- work with reduced feature map for voice activity detection,” in ICASSP 2024 - 2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), 2024, pp. 12 266– 12 270

    2024

  15. [15]

    Sincqdr-vad: A noise-robust voice activity de- tection framework leveraging learnable filters and ranking-aware optimization,

    C.-C. Wang et al., “Sincqdr-vad: A noise-robust voice activity de- tection framework leveraging learnable filters and ranking-aware optimization,”arXiv preprint arXiv:2508.20885, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    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,” inProceedings of Interspeech 2022, 09 2022, pp. 5363–5367

    2022

  17. [17]

    Sg-vad: Stochastic gates based speech activity detection,

    J. Svirsky and O. Lindenbaum, “Sg-vad: Stochastic gates based speech activity detection,” inICASSP 2023-2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  18. [18]

    Atomicvad: A tiny voice activity detection model for efficient inference in intelligent iot systems,

    A. J. Soto-Vergel, P. Sankaran, J. C. Velez, R. Amaya-Mier, and D. Ramirez-Rios, “Atomicvad: A tiny voice activity detection model for efficient inference in intelligent iot systems,”Internet of Things, p. 101822, 2025

    2025

  19. [19]

    A V A-Speech: A Densely Labeled Dataset of Speech Activity in Movies,

    S. Chaudhuri et al., “A V A-Speech: A Densely Labeled Dataset of Speech Activity in Movies,” inProceedings of Interspeech, 2018, pp. 1239–1243. [Online]. Available: https://www.isca-speech. org/archive/Interspeech 2018/abstracts/2028.html

    2018

  20. [20]

    Accurate and structured pruning for effi- cient automatic speech recognition,

    H. Jiang, L. L. Zhang, Y . Li, Y . Wu, S. Cao, T. Cao, Y . Yang, J. Li, M. Yang, and L. Qiu, “Accurate and structured pruning for effi- cient automatic speech recognition,” inProc. Interspeech, 2023, pp. 4104–4108

    2023

  21. [21]

    Quantization aware training with absolute-cosine regularization for automatic speech recognition,

    H. D. Nguyen, A. Alexandridis, and A. Mouchtaris, “Quantization aware training with absolute-cosine regularization for automatic speech recognition,” inProc. Interspeech, 2020, pp. 3366–3370

    2020

  22. [22]

    Tempo- ral properties of spontaneous speech—a syllable-centric perspec- tive,

    S. Greenberg, H. Carvey, L. Hitchcock, and S. Chang, “Tempo- ral properties of spontaneous speech—a syllable-centric perspec- tive,”Journal of Phonetics, vol. 31, no. 3-4, pp. 465–485, 2003

    2003

  23. [23]

    Depgraph: Towards any structural pruning,

    G. Fang, X. Ma, M. Song, M. B. Mi, and X. Wang, “Depgraph: Towards any structural pruning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 16 091–16 101

    2023

  24. [24]

    Optuna: A next-generation hyperparameter optimization framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next-generation hyperparameter optimization framework,” in KDD, 2019

    2019

  25. [25]

    The high-dimensional geome- try of binary neural networks,

    A. G. Anderson and C. P. Berg, “The high-dimensional geome- try of binary neural networks,”arXiv preprint arXiv:1705.07199, 2017

    Pith/arXiv arXiv 2017

  26. [26]

    Neural networks weights quantization: Target none-retraining ternary (tnt),

    T. Zhang, L. Zhu, Q. Zhao, and K. Shin, “Neural networks weights quantization: Target none-retraining ternary (tnt),” in2019 Fifth Workshop on Energy Efficient Machine Learning and Cognitive Computing – NeurIPS Edition (EMC2-NIPS). IEEE, Dec. 2019, pp. 62–65

    2019

  27. [27]

    Feature affinity assisted knowledge distillation and quantization of deep neural networks on label-free data,

    Z. Li, B. Yang, P. Yin, Y . Qi, and J. Xin, “Feature affinity assisted knowledge distillation and quantization of deep neural networks on label-free data,”IEEE Access, vol. 11, pp. 78 042–78 051, 2023

    2023

  28. [28]

    Sphereface: Deep hypersphere embedding for face recognition,

    W. Liu, Y . Wen, Z. Yu, M. Li, B. Raj, and L. Song, “Sphereface: Deep hypersphere embedding for face recognition,” inCVPR, 2017, pp. 6738–6746

    2017

  29. [29]

    Cosface: Large margin cosine loss for deep face recog- nition,

    H. Wang, Y . Wang, Z. Zhou, X. Ji, D. Gong, J. Zhou, Z. Li, and W. Liu, “Cosface: Large margin cosine loss for deep face recog- nition,” inCVPR, 2018, pp. 5265–5274

    2018

  30. [30]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, I. Kotsia, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” inCVPR, 2019

    2019

  31. [31]

    Soft-to-hard vector quanti- zation for end-to-end learning compressible representations,

    E. Agustsson, F. Mentzer, M. Tschannen, L. Cavigelli, R. Tim- ofte, L. Benini, and L. Van Gool, “Soft-to-hard vector quanti- zation for end-to-end learning compressible representations,” in NeurIPS, 2017, pp. 1141–1151

    2017

  32. [32]

    Quantized neural networks: Training neural networks with low precision weights and activations,

    I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y . Ben- gio, “Quantized neural networks: Training neural networks with low precision weights and activations,”Journal of Machine Learning Research, vol. 18, no. 187, pp. 1–30, 2018

    2018

  33. [33]

    Lib- rispeech: an asr corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: an asr corpus based on public domain audio books,” in2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  34. [34]

    Simulating wind noise with airflow speed-dependent characteristics,

    D. Mirabilii, A. Lodermeyer, F. Czwielong, S. Becker, and E. A. Habets, “Simulating wind noise with airflow speed-dependent characteristics,” in2022 International Workshop on Acoustic Sig- nal Enhancement (IWAENC), 2022, pp. 1–5

    2022

  35. [35]

    The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,

    C. K. Reddyet al., “The interspeech 2020 deep noise suppression challenge: Datasets, subjective testing framework, and challenge results,” inInterspeech, 2020

    2020

  36. [36]

    Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi,

    M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Son- deregger, “Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi,” inProc. Interspeech 2017, 2017, pp. 498–502

    2017

  37. [37]

    Speech model pre-training for end-to-end spoken language understanding,

    L. Lugosch, M. Ravanelli, P. Ignoto, V . S. Tomar, and Y . Bengio, “Speech model pre-training for end-to-end spoken language understanding,” 2019. [Online]. Available: https: //arxiv.org/abs/1904.03670

    Pith/arXiv arXiv 2019

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.