Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Pushing the Performance of Synthetic Speech Detection with Kolmogorov-Arnold Networks and Self-Supervised Learning Models

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

Pith's one-line read Swapping the MLP for a Kolmogorov-Arnold layer pushes ASVspoof 2021 LA error down to 0.70 percent.

desk verdict A plausible drop-in improvement with an inflated headline number due to a questionable TCM baseline reproduction. read the letter →

arxiv 2506.14153 v1 pith:WYJZPTU5 submitted 2025-06-17 cs.SD cs.CLeess.AS

classification cs.SDcs.CLeess.AS
keywords syntheticspeechdetectiondeepfakeKolmogorov-ArnoldNetworksGroup-RationalKANXLSR-ConformerASVspoof2021self-supervisedlearninganti-spoofing
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 tries to establish that a single architectural swap improves synthetic-speech detection: replace the Multi-Layer Perceptron projection that maps self-supervised speech features into a Conformer encoder with a Group-Rational Kolmogorov-Arnold Network (GR-KAN). On ASVspoof 2021, the resulting XLSR-GRKAN-Conformer model reaches 0.80% and 0.70% equal error rate on the logical-access track under fixed- and variable-length evaluation, and it improves deepfake-track results as well. The authors report average relative EER reductions of 29.1% across four SSL backbones (WavLM, XLS-R, UniSpeech-SAT, and mHuBERT-147), supporting the claim that the KAN projector generalizes beyond one architecture. If correct, the result matters because it upgrades a leading SSL-based countermeasure without new data, new losses, or a redesigned backbone.

What carries the argument

The load-bearing object is the Group-Rational Kolmogorov-Arnold Network (GR-KAN) used as the projection layer between the frozen self-supervised encoder (XLS-R) and the Conformer encoder. KAN is a neural architecture based on the Kolmogorov-Arnold representation theorem: it represents a multivariate function as a sum of learnable univariate functions, here implemented as B-splines. GR-KAN replaces the B-spline with a rational function, shares that function across groups of input channels, and uses a variance-preserving initialization so it can be initialized directly from an MLP's linear-layer weights. This layer does the dimensionality reduction that the MLP previously did, while the paper argues its learnable univariate activations approximate high-dimensional SSL feature mappings more flexibly and lose less information.

What would settle it

Re-run baseline 2 with the original authors' released code and checkpoints; if the TCM baseline reproduces near 1.07% EER on 21LA (Fix) rather than the in-house 1.74%, the proposed model's headline 54.0% relative gain against TCM collapses to a few percent, and only the smaller gains against the non-TCM baseline and the cross-SSL ablation remain as evidence.

Watch

Extended reading notes

Core claim

The central claim is that KAN-based projectors are a better interface between self-supervised speech representations and sequence models than the fixed-activation MLP. Concretely, replacing the SeLU-activated linear projection in XLSR-Conformer with a GR-KAN layer yields state-of-the-art equal error rates of 0.80% (fixed length) and 0.70% (variable length) on the ASVspoof 2021 LA evaluation set, and 1.95% and 2.31% on the 21DF set for the non-TCM variant. Adding the temporal-channel modeling (TCM) module further lowers LA EER to 0.80% fixed and 0.70% variable. The paper also demonstrates consistent gains when the same projector swap is applied to WavLM Large, UniSpeech-SAT, and mHuBERT-147 features, and argues that GR-KAN's group-wise rational activations and variance-preserving initialization avoid the feature attenuation that an MLP projection can cause in high-dimensional SSL embeddings.

Load-bearing premise

The headline relative gains assume the reproduced XLSR-Conformer+TCM baseline (1.74% EER on 21LA fixed) is a valid comparison point, but the paper does not verify it against the original authors' implementation, and that baseline scores worse than the simpler non-TCM baseline (1.07%) in the same table.

Editorial extensions

If this is right

  • The same GR-KAN projector can be dropped into other SSL-plus-classifier pipelines, not only XLSR-Conformer; the ablation shows gains for WavLM, UniSpeech-SAT, and mHuBERT-147.
  • XLSR-GRKAN-Conformer sets a new reported EER of 0.70% on ASVspoof 2021 LA variable-length evaluation, meaning fewer spoofed utterances pass a speaker-verification countermeasure.
  • Because GR-KAN can load MLP weights, an already trained SSL-MLP system can be converted to the KAN projector and fine-tuned rather than trained from scratch.
  • The largest relative gains appear when the projector is combined with temporal-channel modeling (TCM); on the DF track the gains are smaller but still consistent.

Reading between the lines

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

  • Editorial inference: if the projector swap is as backbone-agnostic as the ablation suggests, the same replacement could transfer to other SSL downstream tasks that feed high-dimensional representations into sequence models, such as speaker verification or speech emotion recognition.
  • Editorial inference: the pattern of larger gains on higher-error SSL backbones (UniSpeech-SAT and mHuBERT) hints that GR-KAN helps most when the SSL representation is less perfectly aligned with the task; a testable prediction is that gains shrink as the SSL backbone is already near ceiling.
  • Editorial inference: the reported 60.55% relative average improvement blends large gains against the reproduced TCM baseline and smaller gains against the non-TCM baseline, so a fair benchmark comparison would separate those two reference points; the cross-SSL ablation is the cleaner evidence for the projector's value.
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 replacing the MLP projection layer in the XLSR-Conformer architecture with a Group-Rational Kolmogorov-Arnold Network (GR-KAN) for synthetic speech detection. The authors evaluate the approach on the ASVspoof 2021 LA and DF benchmarks, reporting a 60.55% relative EER improvement over a reproduced TCM baseline and new state-of-the-art results (0.80% and 0.70% EER on 21LA fixed and variable conditions). They also present an ablation study showing consistent EER reductions when the MLP is replaced with GR-KAN across four different SSL feature extractors (WavLM, XLS-R, UniSpeech-SAT, mHuBERT-147). The manuscript includes a theoretical overview of KAN/GR-KAN, a description of the modified architecture, and experiments with fixed- and variable-length utterance settings.

Significance. If validated, the result is practically significant: a simple drop-in replacement of the MLP projector with GR-KAN yields consistent, often large, EER reductions across multiple SSL backbones, and the code is publicly released. The ablation in Table 3 is the strongest evidence, showing relative EER reductions of 29.1% on average across four SSL models and two tasks. However, the headline 60.55% relative improvement and the claimed SOTA status rest on the reproduction of the XLSR-Conformer+TCM baseline, which is not independently verified and whose reported numbers are inconsistent with the source paper. The qualitative conclusion that GR-KAN helps is supported, but the magnitude and the SOTA claim need careful re-baselining and correction of internal inconsistencies.

major comments (4)
  1. [Abstract, §4.3, Table 1] The headline 60.55% relative improvement is the average of the 54.0% (21LA Fix) and 67.1% (21LA Var) gains reported against Baseline 2 (XLSR-Conformer+TCM). However, the reproduced Baseline 2 is substantially worse than the reproduced Baseline 1 (1.74% vs 1.07% EER on 21LA Fix; 2.13% vs 1.07% on 21LA Var), which contradicts the original TCM paper's claim that TCM improves the Conformer baseline by about 26% relatively. Because the 54.0% and 67.1% figures are computed against this weak reproduction, the abstract's headline gain and the claim of new state-of-the-art results are not independently auditable. Please provide external verification of the TCM reproduction (e.g., released checkpoints or a direct comparison with the official numbers from [6]) or recompute the claimed relative improvements against a baseline whose reproduction is verified. The same issue applies to the DF rows, where the relative gains over Baseline 2 are only 7.3% and 5.4%.
  2. [§4.3, Table 1] The claimed 17.8% relative improvement for 21LA (Var) over Baseline 1 compares the proposed variable-length result (0.88% EER) with Baseline 1's fixed-length result (1.07% EER), because Table 1 lists no variable-length entry for Baseline 1. This is not an apples-to-apples comparison. Please obtain a variable-length Baseline 1 evaluation or remove the claim.
  3. [§5, Table 1] The conclusion states that the model achieves 'a relative reduction of EER of up to 54.0% on LA set, and 67.1% on DF set,' but Table 1 assigns 67.1% to 21LA (Var), while the 21DF reductions are at most 23.5% (and only 5-7% against Baseline 2). This internal inconsistency should be corrected, and the abstract's 60.55% figure should be recomputed with consistent track labels.
  4. [§4.3, Table 1 and Table 3] Several reported gains are small in absolute terms (e.g., 1.07→1.05 EER on 21LA Fix, and 2.55→2.31 on 21DF Var), and the paper does not report repeated-run variance or any significance test. Since the central claim is that replacing the MLP projector improves detection performance, please provide evidence that these differences are not within run-to-run variability, or temper the corresponding claims.
minor comments (5)
  1. [§3.1] There are typos in the text: 'Out projected features' should be 'Our projected features', and 'learns to capture the most relevant captures' should probably be 'learns to capture the most relevant features'.
  2. [§4.2] Please report the GR-KAN configuration (number of layers, group size k, output dimension, and whether the weights were initialized from the baseline MLP) and the Top-5 ensemble selection criterion; these details are needed for reproducibility.
  3. [§4.2] The RawBoost settings differ between LA and DF tracks; state explicitly whether these settings were chosen before seeing evaluation results or selected on the development set, to rule out evaluation-set tuning.
  4. [Table 3] The SSL model references are given as URLs in footnotes; consider listing them in the reference list for completeness.
  5. [§2.3, Eq. (6)] Equation (6) is hard to read; define I, k, O, and w clearly and check the floor expression notation in the published version.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical comparisons against the public ASVspoof 2021 benchmark, not derivations that reduce to their inputs.

full rationale

The paper's central claim is empirical: replacing the MLP projection with GR-KAN in XLSR-Conformer, with or without TCM, lowers equal error rates on the ASVspoof 2021 LA and DF evaluation sets. The results are measured against a public, externally defined benchmark and compared with previously published systems. No parameter is fitted to the evaluation labels, and no predicted quantity is defined in terms of the measured outcome. The theoretical sections on the Kolmogorov-Arnold representation theorem and GR-KAN equations motivate the architecture but do not assert that the KAN formulation itself generates the EER numbers; those numbers come from held-out evaluation. The only self-referential element is reliance on the baseline papers' configurations ([5], [6]) when reproducing XLSR-Conformer and XLSR-Conformer+TCM, which is standard experimental practice and is not load-bearing circularity: the comparison remains against external data. The in-house reproduction of XLSR-Conformer+TCM shows a reversed ordering relative to the original TCM paper (Table 1 gives Baseline 2 as 1.74% EER on 21LA Fix while Baseline 1 is 1.07%), and the conclusion labels the 67.1% gain as 'DF' whereas Table 1 assigns it to 21LA (Var). These are correctness, comparability, and reporting risks, not circular reasoning, because even a mistuned baseline would not make the proposed model's evaluation reduce to its architectural definition. Accordingly, the paper is self-contained against external benchmarks and no circular step is present.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central empirical claim rests on the ASVspoof benchmark being meaningful, on the fidelity of in-house reproductions of the two baseline models, and on the applicability of GR-KAN as a feature projector. No new mathematical axioms or entities are introduced, but the paper's headline relative improvements depend on the reproduced TCM baseline whose configuration is not independently validated.

free parameters (2)
  • Top-5 validation ensemble = 5
    Section 4.2: final results are averaged over the top 5 best models on the validation set; this hand-chosen protocol affects the reported EER and is applied to both baselines and proposed models.
  • RawBoost augmentation configuration per track = LA: convolutive + impulsive additive noise; DF: stationary colored noise
    Section 4.2: different augmentation settings are used for LA and DF training; these choices are adopted from the baseline paper and affect the relative gains.
assumptions (3)
  • domain assumption ASVspoof 2021 evaluation protocol is a valid measure of synthetic speech detection performance.
    Section 4.1 uses the challenge's LA and DF tracks with EER and min t-DCF as metrics; the claim of SOTA assumes this benchmark is meaningful.
  • domain assumption The reproduced XLSR-Conformer and XLSR-Conformer+TCM baselines faithfully match the original published implementations.
    Section 4.2 says all configurations from baseline papers were retained; however Baseline 2's reproduced 21LA Fix EER (1.74%) is worse than Baseline 1 (1.07%), raising doubt about this assumption.
  • domain assumption The Kolmogorov-Arnold representation theorem provides a valid justification for using GR-KAN as a feature projector in this setting.
    Section 2.2 presents the theorem as motivation, but no proof is given that SSL feature projection benefits from it in this context.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pushing the Performance of Synthetic Speech Detection with Kolmogorov-Arnold Networks and Self-Supervised Learning Models." pith.science (2026). https://pith.science/paper/WYJZPTU5

@misc{pith2026250614153,
  author       = {Pith},
  title        = {Pith review of: Pushing the Performance of Synthetic Speech Detection with Kolmogorov-Arnold Networks and Self-Supervised Learning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WYJZPTU5}},
  note         = {Machine review of arXiv:2506.14153}
}
read the original abstract

Recent advancements in speech synthesis technologies have led to increasingly advanced spoofing attacks, posing significant challenges for automatic speaker verification systems. While systems based on self-supervised learning (SSL) models, particularly the XLSR-Conformer model, have demonstrated remarkable performance in synthetic speech detection, there remains room for architectural improvements. In this paper, we propose a novel approach that replaces the traditional Multi-Layer Perceptron in the XLSR-Conformer model with a Kolmogorov-Arnold Network (KAN), a novel architecture based on the Kolmogorov-Arnold representation theorem. Our results on ASVspoof2021 demonstrate that integrating KAN into the SSL-based models can improve the performance by 60.55% relatively on LA and DF sets, further achieving 0.70% EER on the 21LA set. These findings suggest that incorporating KAN into SSL-based models is a promising direction for advances in synthetic speech detection.

Figures

Figures reproduced from arXiv: 2506.14153 by the authors.

Figure 1
Figure 1. Architecture of the baseline XLSR-Conformer model. The XLSR-Conformer+TCM baseline only modifies the MHSA module. 3.2. XLSR-GRKAN-Conformer Model Our proposed architecture1 with XLSR-Conformer enhanced by GR-KAN is illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture of feature projector with GR-KAN. This architectural modification is primarily motivated by the inherent requirement to reduce the dimensionality of speech representations derived from SSL models before they can serve as input to the sequence-to-sequence models like Conformer - a scenario where KAN’s strengths become particularly ad￾vantageous. Furthermore, the GR-KAN’s inherent capability 1 Implementat… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 21 canonical work pages

  1. [6]

    Temporal-channel modeling in multi-head self- attention for synthetic speech detection,

    D.-T. Truong, R. Tao, T. Nguyen, H.-T. Luong, K. A. Lee, and E. S. Chng, “Temporal-channel modeling in multi-head self- attention for synthetic speech detection,” inInterspeech 2024, 2024, pp. 537–541

  2. [1]

    Introduction In recent years, speech synthesis technologies have achieved remarkable progress, enabling the generation of increasingly more natural and convincing synthetic voices. While these ad- vancements in text-to-speech (TTS) and voice conversion (VC) systems demonstrate the potential of conversational AI applica- tions in human-computer interaction...

  3. [2]

    Pushing the Performance of Synthetic Speech Detection with Kolmogorov-Arnold Networks and Self-Supervised Learning Models

    Theoretical Formulation 2.1. Multi-Layer Perceptrons (MLPs) A Multi-Layer Perceptron (MLP) is a fully connected feedfor- ward neural network consisting of multiple layers of neurons. Each neuron in a layer is connected to every node in the follow- ing layer, and the node applies a nonlinear activation function to the weighted sum of its inputs. The founda...

  4. [3]

    Proposed Methodology 3.1. Baseline Model Architectures We use two state-of-the-art architectures as our baseline, XLSR-Conformer [6] and its variant XLSR-Conformer+TCM with an additional temporal-channel dependency modelling (TCM) module. As can be seen in Figure 1, the XLSR- Conformer baseline comprises two main parts: (i) the pre- trained XLS-R [25], wh...

  5. [4]

    Experiments 4.1. Datasets and Evaluation metrics The training and development datasets are sourced from the ASVspoof 2019 [28] logical access (LA) track, including bonafide speech and synthetic speech generated from two speech synthesis techniques: voice conversion and text-to- speech. We evaluate our model on the ASVspoof 2021 [29] logical access (LA) an...

  6. [5]

    Our model replaces fully-connected layer by GR-KAN, which is powerful in func- tion approximation and dimensionality reduction

    Conclusion In this paper, we proposed a novel architecture utilised GR- KAN for anti-spoofing speech systems. Our model replaces fully-connected layer by GR-KAN, which is powerful in func- tion approximation and dimensionality reduction. This place- ment can enhance the system’s continual learning capability when GR-KAN is positioned between the SSL model...

  7. [7]

    An overview of voice conversion and its challenges: From statistical modeling to deep learning,

    B. Sisman, J. Yamagishi, S. King, and H. Li, “An overview of voice conversion and its challenges: From statistical modeling to deep learning,”IEEE/ACM Trans. Audio, Speech and Lang. Proc., vol. 29, p. 132–157, Nov. 2020. [Online]. Available: https://doi.org/10.1109/TASLP.2020.3038524

  8. [8]

    Human perception of audio deepfakes,

    N. M. M ¨uller, K. Pizzi, and J. Williams, “Human perception of audio deepfakes,” inProceedings of the 1st International Workshop on Deepfake Detection for Audio Multimedia, ser. DDAM ’22. New York, NY , USA: Association for Computing Machinery, 2022, p. 85–91. [Online]. Available: https: //doi.org/10.1145/3552466.3556531

Show all 37 references
  1. [9]

    Vsasv: a vietnamese dataset for spoofing-aware speaker verification,

    V . Hoang, V . T. Pham, H. N. Xuan, P. Nhi, P. Dat, and T. T. T. Nguyen, “Vsasv: a vietnamese dataset for spoofing-aware speaker verification,” inInterspeech 2024, 2024, pp. 4288–4292

  2. [10]

    V oice spoofing countermeasures: Taxonomy, state-of-the-art, experimental analysis of generalizability, open challenges, and the way forward,

    A. Khan, K. M. Malik, J. Ryan, and M. Saravanan, “V oice spoofing countermeasures: Taxonomy, state-of-the-art, experimental analysis of generalizability, open challenges, and the way forward,” 2022. [Online]. Available: https: //arxiv.org/abs/2210.00417

  3. [11]

    A conformer-based classifier for variable-length utterance process- ing in anti-spoofing,

    E. Rosello, A. Gomez-Alanis, A. M. Gomez, and A. Peinado, “A conformer-based classifier for variable-length utterance process- ing in anti-spoofing,” inINTERSPEECH 2023, 2023, pp. 5281– 5285

  4. [12]

    Robust audio deep- fake detection using ensemble confidence calibration,

    C. Y . Kwok, D.-T. Truong, and J. Q. Yip, “Robust audio deep- fake detection using ensemble confidence calibration,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

  5. [13]

    Fine-tuning wav2vec2 for speaker recognition,

    N. Vaessen and D. A. Van Leeuwen, “Fine-tuning wav2vec2 for speaker recognition,” inICASSP 2022-2022 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7967–7971

  6. [14]

    Exploring wav2vec 2.0 fine tuning for improved speech emotion recognition,

    L.-W. Chen and A. Rudnicky, “Exploring wav2vec 2.0 fine tuning for improved speech emotion recognition,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

  7. [15]

    Exploring speaker age estimation on different self-supervised learning models,

    D.-T. Truong, T. T. Anh, and C. E. Siong, “Exploring speaker age estimation on different self-supervised learning models,” in 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2022, pp. 1950– 1955

  8. [16]

    Estimation of speaker age and height from speech signal using bi-encoder transformer mixture model,

    T. Gupta, T. D. Truong, T. T. Anh, and E. S. Chng, “Estimation of speaker age and height from speech signal using bi-encoder transformer mixture model,” inInterspeech 2022, 2022, pp. 1978– 1982

  9. [17]

    Nes2net: A lightweight nested architecture for foundation model driven speech anti-spoofing,

    T. Liu, D.-T. Truong, R. K. Das, K. A. Lee, and H. Li, “Nes2net: A lightweight nested architecture for foundation model driven speech anti-spoofing,”arXiv preprint arXiv:2504.05657, 2025

  10. [18]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998

  11. [19]

    Conformer: Convolution-augmented transformer for speech recognition,

    A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y . Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y . Wu, and R. Pang, “Conformer: Convolution-augmented transformer for speech recognition,”

  12. [20]

    The kolmogorov–arnold representation theorem revisited,

    J. Schmidt-Hieber, “The kolmogorov–arnold representation theorem revisited,”Neural Networks, vol. 137, pp. 119– 126, 2021. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S0893608021000289

  13. [21]

    Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,

    J. weon Jung, H.-S. Heo, H. Tak, H. jin Shim, J. S. Chung, B.-J. Lee, H.-J. Yu, and N. Evans, “Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” 2021. [Online]. Available: https://arxiv.org/abs/2110.01200

  14. [22]

    Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation,

    H. Tak, M. Todisco, X. Wang, J. weon Jung, J. Yamagishi, and N. Evans, “Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation,” 2022. [Online]. Available: https://arxiv.org/abs/2202.12233

  15. [23]

    Leveraging positional-related local-global dependency for syn- thetic speech detection,

    X. Liu, M. Liu, L. Wang, K. A. Lee, H. Zhang, and J. Dang, “Leveraging positional-related local-global dependency for syn- thetic speech detection,” inICASSP 2023 - 2023 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

  16. [24]

    Light convolutional neu- ral network with feature genuinization for detection of synthetic speech attacks,

    Z. Wu, R. Das, J. Yang, and H. Li, “Light convolutional neu- ral network with feature genuinization for detection of synthetic speech attacks,” 10 2020, pp. 1101–1105

  17. [25]

    Xls-r: Self-supervised cross- lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Conneau, and M. Auli, “Xls-r: Self-supervised cross- lingual speech representation learning at scale,” 2021. [Online]. Available: https://arxiv.org/abs/2111.09296

  18. [26]

    Given an input speech signalO, theT-length output SSL features are denoted as X=SSL(O) = (x t ∈R D|t= 1, ..., T), withDbeing the output dimension of the SSL model

    model, utilised as the feature extractor to capture contex- tualised representations from the high-dimensional speech sig- nal; and (ii) the Conformer Encoder. Given an input speech signalO, theT-length output SSL features are denoted as X=SSL(O) = (x t ∈R D|t= 1, ..., T), wit...

  19. [27]

    Kan: Kolmogorov-arnold networks,

    Z. Liu, Y . Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljaˇci´c, T. Y . Hou, and M. Tegmark, “Kan: Kolmogorov-arnold networks,” 2024. [Online]. Available: https://arxiv.org/abs/2404. 19756

  20. [28]

    Suitability of kans for computer vision: A preliminary investigation,

    B. Azam and N. Akhtar, “Suitability of kans for computer vision: A preliminary investigation,”arXiv preprint arXiv:2406.09087, 2024

  21. [29]

    An investigation on the potential of kan in speech enhancement,

    H. Li, Y . Hu, C. Chen, and E. S. Chng, “An investigation on the potential of kan in speech enhancement,”arXiv preprint arXiv:2412.17778, 2024

  22. [30]

    Multilayer feedfor- ward networks are universal approximators,

    K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedfor- ward networks are universal approximators,”Neural networks, vol. 2, no. 5, pp. 359–366, 1989

  23. [31]

    Kolmogorov-arnold transformer,

    X. Yang and X. Wang, “Kolmogorov-arnold transformer,” 2024. [Online]. Available: https://arxiv.org/abs/2409.10594

  24. [33]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” 2020. [Online]. Available: https://arxiv.org/abs/ 2006.11477

  25. [34]

    End-to-end anti-spoofing with rawnet2,

    H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and A. Larcher, “End-to-end anti-spoofing with rawnet2,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 6369–6373

  26. [35]

    Asvspoof 2019: Future horizons in spoofed and fake audio detection,

    M. Todisco, X. Wang, V . Vestman, M. Sahidullah, H. Delgado, A. Nautsch, J. Yamagishi, N. W. D. Evans, T. H. Kinnunen, and K. A. LEE, “Asvspoof 2019: Future horizons in spoofed and fake audio detection,” inInterspeech, 2019. [Online]. Available: https://api.semanticscholar.org...

  27. [36]

    Asvspoof 2021: Towards spoofed and deepfake speech de- tection in the wild,

    X. Liu, X. Wang, M. Sahidullah, J. Patino, H. Delgado, T. Kin- nunen, M. Todisco, J. Yamagishi, N. Evans, A. Nautsch, and K. A. Lee, “Asvspoof 2021: Towards spoofed and deepfake speech de- tection in the wild,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vo...

  28. [37]

    Raw- boost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,

    H. Tak, M. Kamble, J. Patino, M. Todisco, and N. Evans, “Raw- boost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,” inICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 20...

  29. [2020]

    Available: https://arxiv.org/abs/2005.08100

    [Online]. Available: https://arxiv.org/abs/2005.08100

Pith tools

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