arxiv: 2604.19209 · v1 · submitted 2026-04-21 · 💻 cs.SD

Recognition: unknown

Audio Spoof Detection with GaborNet

Waldek Maciejko

Authors on Pith no claims yet

Pith reviewed 2026-05-10 02:01 UTC · model grok-4.3

classification 💻 cs.SD

keywords audio spoof detectionGabor filtersraw audio processingneural network front-endfeature extractiondata augmentationRawNet2RawGAT-ST

0 comments

The pith

Gabor filter banks serve as an ingestion layer for raw audio in neural networks built for spoof detection.

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

The paper examines a new front-end layer called GaborNet that replaces sinc-based convolution with a bank of Gabor filters applied directly to raw audio samples. Because Gabor outputs are complex, the work applies post-processing steps such as squared modulus or Gaussian lowpass pooling to make the features usable by later layers. These modified layers are inserted into the established RawNet2 and RawGAT-ST architectures, which were originally designed for distinguishing genuine from spoofed speech, while training is further supported by augmenting the data with codec conversions, room responses, and additive noise. A reader would care because end-to-end models that start from raw waveforms avoid manual feature engineering, and a filter bank that reduces frequency-domain distortions could produce more reliable detection of voice spoofs.

Core claim

An ingestion layer built from a bank of Gabor filters, named GaborNet, together with the required modifications for complex-valued outputs, can be integrated into the RawNet2 and RawGAT-ST architectures for audio spoof detection, and standard audio augmentation techniques using codecs, room impulse responses, and additive noises further support training of these models.

What carries the argument

GaborNet: a neural ingestion layer that convolves the raw input waveform with a bank of Gabor filters and applies post-processing (squared modulus or Gaussian lowpass pooling) to convert the complex results into real-valued features suitable for subsequent network stages.

If this is right

Frequency-domain distortions caused by truncating sinc functions are reduced when Gabor filters are used instead.
Raw audio can be processed directly inside established spoof-detection networks without intermediate hand-crafted features.
Augmenting training data with codec conversions, reverberation, and noise improves the robustness of the resulting detectors.
The same GaborNet front-end can be dropped into both RawNet2 and RawGAT-ST, showing architectural flexibility.

Where Pith is reading between the lines

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

The approach may transfer to other raw-audio tasks such as speaker verification or environmental sound classification.
Because Gabor filters are complex, they implicitly retain phase information that purely real-valued filters discard, which could help against phase-manipulated spoofs.
Direct head-to-head accuracy and latency comparisons against SincNet baselines on multiple datasets would clarify whether the gains are consistent.

Load-bearing premise

That Gabor filters, after the modifications needed to handle their complex outputs, extract features from raw audio that are more useful and less distorted than those produced by sinc filters for the task of identifying spoofs.

What would settle it

A controlled experiment in which the GaborNet versions of RawNet2 or RawGAT-ST achieve equal or higher equal-error rates than the original sinc-based versions on a standard benchmark such as ASVspoof would show that the filter replacement does not deliver the intended improvement.

Figures

Figures reproduced from arXiv: 2604.19209 by Waldek Maciejko.

**Figure 1.** Figure 1: Three examples of sinc functions at the training stage across three epochs: 0 𝑡ℎ , 50𝑡ℎ and 100𝑡ℎ. The top row presents the characteristics in the time domain, and the bottom row presents them in the frequency domain [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗

**Figure 2.** Figure 2: Three examples of Gabor filters at the training stage across three epochs: 0 𝑡ℎ , 50𝑡ℎ and 100𝑡ℎ. The top row shows characteristics in the time domain, and the bottom row presents characteristics in the frequency domain [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: The schema of Filter Map Scaling applied in Gabor RawNet2. Waldemar Maciejko Page 11 of 10 [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

**Figure 4.** Figure 4: The schema of the Top K-Pooling layer applied in RawGAT-ST separately to time, frequency and fusion domains [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

**Figure 5.** Figure 5: Comparison of training curves of investigated architectures. Waldemar Maciejko Page 12 of 10 [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

read the original abstract

An direction of development in the extraction of features from audio signals is based on processing raw samples in the time domain. Such an approach appears to be effective, especially in the era of neural networks. An example is SincNet. In this solution, the core of the neural network layer is a set of sinc functions that are convolved with the input signal. Due to the finite length of sinc functions, distortions appear in the frequency domain of the convolved signal, the same as in the case of windowing the signal. Recently, a new approach has been developed that uses Gabor filters to replace sinc functions. Due to the complex results, further modifications had to be applied, such as squared modulus or Gaussian Lowpass Pooling. In this work, an ingestion layer based on a bank of Gabor filters, named GaborNet, and its modifications are intensively examined within the popular RawNet2 and RawGAT- ST architectures. These have been developed for the purpose of audio spoof detection. Another issue that has been investigated was audio augmentation using codec conversions, room responses, and additive noises.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper swaps the sinc layer for a Gabor filter bank inside RawNet2 and RawGAT-ST, handles the complex outputs with squared modulus or Gaussian pooling, and tests standard augmentations, but the abstract contains no performance numbers at all.

read the letter

The one thing to know is that this work swaps the sinc convolution layer for a Gabor filter bank inside RawNet2 and RawGAT-ST, names the result GaborNet, and checks how it behaves with standard audio augmentations for spoof detection. The authors note that Gabor outputs are complex, so they apply squared modulus or Gaussian low-pass pooling to make them usable. They then run the modified networks on the usual spoof detection task. This is a straightforward empirical extension of the SincNet line of work. It does a decent job of documenting the necessary changes to the filter bank and integrating it with existing codebases. The main limitation is the lack of any quantitative results in the abstract. There are no error rates, no baseline comparisons, and no mention of how much the new layer changes detection accuracy. That makes it hard to judge whether the Gabor approach is better, worse, or about the same as the original sinc version. The citation pattern looks normal for this subfield, pulling in the RawNet and SincNet papers. This paper is aimed at researchers who build raw-waveform models for audio authentication or deepfake detection. A reader who wants to try Gabor filters in their own pipeline might get some implementation ideas from it. It deserves a serious referee only if the full manuscript includes the missing performance numbers and proper statistical comparisons. Without those, it is too thin for review. Recommendation: If the results are there and look reasonable, accept for review; otherwise desk reject.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces GaborNet, an ingestion layer consisting of a bank of Gabor filters for raw time-domain audio processing in spoof detection. It replaces sinc functions from prior work like SincNet, applies post-processing modifications (squared modulus or Gaussian Lowpass Pooling) to address complex-valued outputs, integrates the layer into RawNet2 and RawGAT-ST backbones, and evaluates the resulting systems together with standard audio augmentations (codec conversion, room impulse responses, additive noise). The central contribution is framed as an intensive empirical examination rather than a theoretical derivation.

Significance. If the reported experiments demonstrate competitive or improved equal error rates on standard spoof detection benchmarks relative to sinc-based baselines, the work would provide a practical alternative for learnable filter banks in raw-audio anti-spoofing pipelines. The explicit treatment of complex-output handling and the combination with established augmentation pipelines add incremental engineering value, though the absence of any numerical results in the abstract limits immediate assessment of impact.

major comments (1)

[Abstract] Abstract: the claim of an 'intensive examination' is not supported by any reported metrics, baselines, or error bars, so the soundness of the empirical conclusions cannot be evaluated from the provided summary.

minor comments (2)

[Abstract] The abstract and introduction should explicitly state the datasets (e.g., ASVspoof 2019/2021), evaluation metrics, and number of runs to allow readers to gauge reproducibility.
[Methods] Notation for the Gabor filter parameters (center frequency, bandwidth, etc.) and the exact form of the squared-modulus and Gaussian Lowpass Pooling operations should be given in a dedicated methods subsection with equations.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback. We address the major comment below.

read point-by-point responses

Referee: Abstract: the claim of an 'intensive examination' is not supported by any reported metrics, baselines, or error bars, so the soundness of the empirical conclusions cannot be evaluated from the provided summary.

Authors: We agree that the abstract does not currently include numerical results, which limits immediate assessment of the empirical claims. In the revised manuscript we will update the abstract to report key equal error rates (EER) achieved by the GaborNet variants within RawNet2 and RawGAT-ST, together with direct comparisons to the corresponding SincNet baselines and a brief indication of the augmentation pipeline. This change will allow readers to evaluate the strength of the empirical examination from the abstract itself. revision: yes

Circularity Check

0 steps flagged

No significant circularity: purely empirical examination

full rationale

The paper describes an empirical study replacing the sinc layer in RawNet2 and RawGAT-ST with a Gabor-filter bank (GaborNet) plus documented post-processing (squared modulus or Gaussian low-pass pooling) and standard audio augmentations. No equations, derivations, predictions, or first-principles claims are present; performance numbers are obtained from experiments on spoof-detection benchmarks. No self-citation is used to justify a uniqueness theorem or to force a result by construction. The central claim is limited to 'intensive examination' of the modified architectures, which is self-contained against external benchmarks and does not reduce to any fitted input or renamed ansatz.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The work relies on standard assumptions from signal processing about Gabor filter properties and neural network training on augmented data, with no free parameters or new axioms stated.

invented entities (1)

GaborNet no independent evidence
purpose: Ingestion layer using a bank of Gabor filters for raw audio feature extraction in spoof detection networks
Newly named component introduced in the paper, built on standard Gabor filters plus modifications for complex outputs.

pith-pipeline@v0.9.0 · 5477 in / 1154 out tokens · 45133 ms · 2026-05-10T02:01:08.111602+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 12 canonical work pages

[1]

Deepscatteringspectrum

Andén,J.,Mallat,S.,2014. Deepscatteringspectrum. IEEETransactionsonSignalProcessing62,4114–4128. doi:10.1109/TSP.2014.2326991. Brümmer, N., de Villiers, E.,

work page doi:10.1109/tsp.2014.2326991 2014
[2]

The bosaris toolkit: Theory, algorithms and code for surviving the new dcf,

The bosaris toolkit: Theory, algorithms and code for surviving the new dcf. ArXiv abs/1304.2865. URL: https://api.semanticscholar.org/CorpusID:14392885. Cohen, A., Rimon, I., Aflalo, E., Permuter, H.H.,

work page arXiv
[3]

Alain de Cheveigné and Hideki Kawahara

Comparison of parametric representations for monosyllabic word recognition in continuously spoken sentences. IEEE Transactions on Acoustics, Speech, and Signal Processing 28, 357–366. doi:10.1109/TASSP.1980.1163420. Gao,H.,Ji,S.,2022. Graphu-nets. IEEETransactionsonPatternAnalysisandMachineIntelligence44,4948–4960. doi:10.1109/TPAMI.2021. 3081010. Gupta, ...

work page doi:10.1109/tassp.1980.1163420 1980
[4]

EURASIP Journal on Audio, Speech,and MusicProcessing2024

Vulnerability issues in Automatic Speaker Verification (ASV) systems. EURASIP Journal on Audio, Speech,and MusicProcessing2024. URL:http://dx.doi.org/10.1186/s13636-024-00328-8, doi:10.1186/s13636-024-00328-8. He, K., Zhang, X., Ren, S., Sun, J.,

work page doi:10.1186/s13636-024-00328-8
[5]

ArXiv abs/2004.00526

Improved rawnet with filter-wise rescaling for text-independent speaker verification using raw waveforms. ArXiv abs/2004.00526. URL:https://api.semanticscholar.org/CorpusID:226202021. Knyazev, B., Taylor, G.W., Amer, M.R.,

work page arXiv 2004
[6]

URL:https://api.semanticscholar.org/CorpusID:195069083

Understanding attention and generalization in graph neural networks, in: Neural Information Processing Systems. URL:https://api.semanticscholar.org/CorpusID:195069083. Ko,T.,Peddinti,V.,Povey,D.,Seltzer,M.L.,Khudanpur,S.,2017.Astudyondataaugmentationofreverberantspeechforrobustspeechrecognition, in: 2017 IEEE International Conference on Acoustics, Speech ...

work page doi:10.1109/icassp.2017 2017
[7]

Vggsound: A Large-Scale Audio-Visual Dataset

Cgcnn: Complex gabor convolutional neural network on raw speech, in: ICASSP 2020 - 2020 IEEE InternationalConferenceonAcoustics,SpeechandSignalProcessing(ICASSP),pp.7724–7728. doi:10.1109/ICASSP40776.2020.9054220. Oppenheim, A.V., Schafer, R.W.,

work page doi:10.1109/icassp40776.2020.9054220 2020
[8]

1021–1028

Speaker recognition from raw waveform with sincnet, in: 2018 IEEE Spoken Language Technology Workshop (SLT), pp. 1021–1028. doi:10.1109/SLT.2018.8639585. Snyder, D., Chen, G., Povey, D.,

work page doi:10.1109/slt.2018.8639585 2018
[9]

Audioclip: Extending clip to image, text and audio

Musan: A music, speech, and noise corpus. ArXiv abs/1510.08484. URL:https://api. semanticscholar.org/CorpusID:15676318. Tak, H., Jee-weon Jung, Patino, J., Kamble, M.R., Todisco, M., Evans, N.W.D., 2021a. End-to-end spectro-temporal graph attention networks for speaker verification anti-spoofing and speech deepfake detection. ArXiv abs/2107.12710. URL:htt...

work page doi:10.1109/icassp43922.2022.9746213 2022
[10]

Computer Speech & Language 64, 101114

Asvspoof 2019: A large-scale public database of synthesized, converted and replayed speech. Computer Speech & Language 64, 101114. URL:https: //www.sciencedirect.com/science/article/pii/S0885230820300474, doi:https://doi.org/10.1016/j.csl.2020.101114. Wang, Y., Getreuer, P., Hughes, T., Lyon, R.F., Saurous, R.A.,

work page doi:10.1016/j.csl.2020.101114 2019
[11]

2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 5670–5674URL:https://api.semanticscholar.org/ CorpusID:3100199

Trainable frontend for robust and far-field keyword spotting. 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 5670–5674URL:https://api.semanticscholar.org/ CorpusID:3100199. Zeghidour, N., Teboul, O., de Chaumont Quitry, F., Tagliasacchi, M.,

2017
[12]

ArXiv abs/2101.08596

Leaf: A learnable frontend for audio classification. ArXiv abs/2101.08596. URL:https://api.semanticscholar.org/CorpusID:231662084. Zeghidour, N., Usunier, N., Kokkinos, I., Schaiz, T., Synnaeve, G., Dupoux, E.,

work page arXiv
[13]

5509–5513

Learning filterbanks from raw speech for phone recognition, in: 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 5509–5513. doi:10.1109/ICASSP.2018. 8462015. Waldemar Maciejko Page 10 of 10 Audio Spoof Detection with GaborNet Figure 1:Three examples of sinc functions at the training stage across three epochs:0𝑡ℎ,...

work page doi:10.1109/icassp.2018 2018