Bayesian Learning for Domain-Invariant Speaker Verification and Anti-Spoofing

Hynek Hermansky; Jin Li; Johan Rohdin; Kong Aik Lee; Man-Wai Mak

REVIEW 3 major objections 5 minor 47 references

Bayesian Learning for Domain-Invariant Speaker Verification and Anti-Spoofing

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Modeling uncertainty in frequency-wise normalization weights improves speaker verification and anti-spoofing under domain mismatch.

desk verdict A clean Bayes-by-Backprop reweighting of RFN with promising dev-set numbers, but the uncertainty story is not actually tested and the significance claims outrun the evidence. read the letter →

arxiv 2506.07536 v1 pith:LCUR76RT submitted 2025-06-09 eess.AS

Jin Li , Man-Wai Mak , Johan Rohdin , Kong Aik Lee , Hynek Hermansky This is my paper

classification eess.AS

keywords speakerverificationanti-spoofingdomaingeneralizationBayesianlearningvariationalinferencefrequency-wisenormalizationASVspoof5weightedRFN

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 claims that the normalization weights in relaxed instance frequency-wise normalization (RFN) are a source of overfitting when a speaker-verification or anti-spoofing network is tested on a new domain, because different frequency bins carry different amounts of domain information and the weights themselves are uncertain. To remove this bottleneck, it assigns each frequency bin its own sigmoid-weighted normalization term and treats the weight vector as random with a diagonal-Gaussian variational posterior, calling the resulting layer Bayesian Weighted RFN (BWRFN). On cross-dataset speaker verification, BWRFN-R-vector lowers equal error rate on the unseen HI-MIA domain from 12.65% (plain R-vector) to 8.15%, and on the ASVspoof 5 development sets it reports lower error than the compared baselines for both anti-spoofing (Track 1) and spoofing-robust verification (Track 2). If these results hold, the paper establishes a plug-in normalization layer that makes embeddings more robust to domain mismatch without changing the rest of the network.

What carries the argument

The central object is the Bayesian Weighted RFN (BWRFN) layer. It replaces the scalar relaxation $\lambda$ of RFN with frequency-wise sigmoid-weighted vectors $\mathbf{w}_1$ and $\mathbf{w}_2$ acting on the layer-normalization and instance-frequency-normalization branches, then treats $\mathbf{w}=[\mathbf{w}_1;\mathbf{w}_2]$ as random with a diagonal-Gaussian variational posterior $q(\mathbf{w})=\mathcal{N}(\mathbf{w};\boldsymbol{\mu}_w,\mathrm{diag}(\boldsymbol{\sigma}_w^2))$. Training maximizes the variational lower bound of Eq. 7 with one Monte Carlo sample per segment, drawn via the reparameterization trick $\mathbf{w}^{(k)}=\boldsymbol{\mu}_w+\boldsymbol{\sigma}_w\odot\boldsymbol{\epsilon}^{(k)}$, plus the closed-form KL divergence of Eq. 9. This prevents the network from overcommitting to any single frequency weighting; at inference the expected embedding output is used, which corresponds to setting the weights to their posterior mean $\boldsymbol{\mu}_w$.

What would settle it

Train BWRFN with the KL term removed or with $\boldsymbol{\sigma}_w$ frozen at zero, and compare EER on HI-MIA and the ASVspoof 5 dev sets; if performance does not degrade, the variational posterior is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The core discovery the paper argues for is that accounting for the uncertainty of frequency-wise normalization weights yields domain-invariant speaker embeddings. By modeling $\mathbf{w}=[\mathbf{w}_1^\top,\mathbf{w}_2^\top]^\top$ with a variational posterior $q(\mathbf{w})=\mathcal{N}(\mathbf{w};\boldsymbol{\mu}_w,\mathrm{diag}(\boldsymbol{\sigma}_w^2))$ and optimizing the evidence lower bound of Eq. 7, the BWRFN layer learns to weight the layer-normalization and instance-frequency-normalization branches per frequency bin without committing to a single fixed weight vector. The paper reports that this reduces EER on the unseen HI-MIA domain from 12.65% (R-vector) and 12.96% (RFN-R-vector) to 8.15%, and that BWRFN-ResNet attains minDCF 0.13 and EER 11.36% on the ASVspoof 5 Track 1 development set and min a-DCF 0.125, min t-DCF 0.205, and t-EER 5.83% on Track 2, which it states is better than the current state of the art for the closed condition. The authors interpret this as evidence that Bayesian learning generalizes better than fixed-weight or learned-weight normalization when the test domain is unseen.

Load-bearing premise

The load-bearing premise is that the variational posterior, and specifically the uncertainty it encodes, is what improves domain generalization; because inference uses only the posterior mean, the gains could instead come from the KL regularizer or from random weight perturbations during training, a possibility the paper does not test.

Editorial extensions

If this is right

Inserting BWRFN into the R-vector network lowers EER on the unseen HI-MIA domain to 8.15%, from 12.65% for R-vector and 12.96% for RFN-R-vector.
BWRFN-R-vector also improves overall cross-dataset EER to 12.38% from 13.36% for R-vector, with the best placement being the second residual block (L2).
On ASVspoof 5 Track 1 (closed, development), BWRFN-ResNet reports a minDCF of 0.13 and an EER of 11.36%, the lowest among the compared anti-spoofing systems.
On ASVspoof 5 Track 2 (closed, development), BWRFN-ResNet reports min a-DCF 0.125, min t-DCF 0.205, and t-EER 5.83%, which the paper states is better than the compared state-of-the-art SASV systems.
Because the Bayesian layer is inserted in place of existing normalization layers, it can be added to other embedding front-ends without changing the training loss or architecture, which the authors list as future work.

Reading between the lines

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

A natural disambiguation the paper does not run is to ablate the KL term or fix $\boldsymbol{\sigma}_w=0$ during training; if EER on HI-MIA and the ASVspoof 5 dev sets is unchanged, the gains come from regularization rather than from posterior uncertainty, and if it degrades, the Bayesian mechanism is doing real work.
The same BWRFN layer could be evaluated with test-time Monte Carlo sampling from $q(\mathbf{w})$ to produce a distribution over embeddings, which would give calibrated verification scores and confidence estimates that the paper leaves implicit.
Since the normalization layer is front-end-agnostic, the approach should transfer to other audio tasks where frequency-wise domain shift is the bottleneck, such as acoustic scene classification or language recognition.
The paper reports development-set results for ASVspoof 5; the evaluation-set extension is an obvious next check and would indicate whether the development-set advantages persist on the official protocol.

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.

Desk Editor's Note

read the letter

Short version: the paper adds variational Bayesian weights to the per-frequency normalization parameters in RFN, calls it BWRFN, and reports better EERs on cross-dataset ASV and on ASVspoof 5 development data. That is a legitimate and straightforward engineering idea—Bayes by Backprop on two small weight vectors—and the benchmark coverage is decent: cross-dataset ASV, anti-spoofing, and SASV. If the gains hold up, it is a useful module for robust speaker embeddings.

But the central narrative is oversold. The posterior predictive in Eq. 10 is never used; inference substitutes the mean weight, so no uncertainty is propagated at test time. The improvement may simply be KL regularization during training. That can still be a useful regularization trick, but the paper does not present it that way.

Concrete soft spots, in proportion:

- No significance tests and no multiple seeds. The abstract says "significantly better," but the table shows single-run EERs. HI-MIA improvement is large (8.15 vs 12.65), but CN-Celeb is worse than plain R-vector (16.96 vs 16.01). - The BWRFN insertion layer L2 is selected from test Overall EERs in Table 2. That is selection on the test set. - Section 4.1.2 promises evaluation-set results for ASVspoof 5, but Tables 3 and 4 report only development-set numbers. Calling it state-of-the-art is premature. - WRFN is introduced as a stepping stone and not analyzed separately beyond one row, which is fine but should be stated more clearly.

Credit where due: the variational math is standard and correctly laid out (ELBO, reparameterization, closed-form KL). The evaluation uses external baselines and datasets, so there is no circularity. Related work is appropriately cited, including RFN and Bayes-by-Backprop. The writing is clear enough that the method is reproducible in principle.

This is a speech-community paper, not a methods paper. A serious referee should see it, but the authors should be pushed for multi-seed results, evaluation-set numbers, validation-based layer selection, and an ablation that actually uses the posterior predictive via MC sampling instead of the mean. If the gain persists, it is a solid incremental contribution; if it is only the KL regularizer, that is still publishable but needs an honest reframing. My call: engage, do not desk-reject.

Referee Report

3 major / 5 minor

Summary. The paper proposes a Bayesian extension of relaxed instance frequency-wise normalization (RFN) for domain-invariant speaker verification and anti-spoofing. It introduces trainable frequency-wise weights into RFN (WRFN) and then treats those weights probabilistically via variational inference with a diagonal-Gaussian posterior, optimizing the ELBO (Eq. 7). Experiments cover cross-dataset ASV on FFSVC/CN-Celeb/HI-MIA, and anti-spoofing and spoofing-robust ASV on ASVspoof 5 Track 1 and Track 2. The headline results report An EER improvement on HI-MIA (8.15% vs 12.65% for R-vector) and improved development-set metrics on ASVspoof 5, leading the authors to claim that 'BWRFN is significantly better than WRFN and RFN.' The main conceptual contribution is that modeling uncertainty in the normalization weights improves domain generalization.

Significance. The idea of placing Bayesian uncertainty on per-frequency normalization weights is plausible and the insertion is architecture-agnostic, which could make the method broadly useful if the gains are robust. The paper builds on a clearly defined prior normalization method (RFN) and evaluates across multiple tasks and external baselines, and there is no circularity in the evaluation setup. However, the evidence for the central claim is currently incomplete: the inference procedure described after Eq. 10 is ambiguous about whether posterior uncertainty is actually used at test time, the ASVspoof results are reported only on the development set despite a stated intent to evaluate on the evaluation set, and the word 'significantly' is used without any statistical testing or multiple-seed analysis. With these issues resolved, the contribution would be a modest but useful empirical study; as it stands, the central claim is not fully supported.

major comments (3)

[§3, Eq. (10); §4.2] The paper defines the posterior predictive distribution as an expectation over the posterior (Eq. 10) but then states that 'For simplicity, we use the expected value of embedding layer output during inference.' It is never specified whether this expectation is computed by Monte Carlo sampling from q(w) at test time or by substituting w with its mean μ_w. If the latter, the training-time covariance σ_w is discarded and the final system reduces to a deterministic WRFN with weights μ_w; in that case the reported gains cannot be attributed to uncertainty propagation and would arise only from the KL term in the ELBO. Please state the exact inference recipe and, if sampling is used, the number of Monte Carlo samples at inference.
[§4.1.2; Tables 3–4] Section 4.1.2 says that 'we built our systems on the training and development set and evaluated them on the evaluation set (see Table 2 of [26])', but Tables 3 and 4 report results only 'on the development set'. The claim in Section 5 that the method 'outperforms the current state-of-the-art approaches for the closed condition of Track 2 of ASVspoof 5' is not supported by the reported evidence unless the evaluation-set numbers are supplied, since the official challenge ranking is based on the evaluation set. Either provide the evaluation-set results or restrict the claims to development-set performance.
[§5, Tables 1–2; Abstract] The abstract states that 'BWRFN is significantly better than WRFN and RFN,' but no significance tests, confidence intervals, or multiple-seed results are reported anywhere in the paper. In addition, Table 2 selects the BWRFN insertion location L2 based on the same Overall EER column that is later used as the headline result (12.38%), so the reported number may be optimistic due to selection on the test set. Please report repeated runs or a held-out validation split for model selection, and replace 'significantly' with a statistically grounded statement or remove it if no such analysis is available.

minor comments (5)

[§2.1] The acronym for instance frequency-wise normalization is given as 'INF' in the text but the equations and the rest of the paper use 'IFN'; please pick one notation and use it consistently.
[§3, Eq. (9)] The sentence 'where σw,f and σw,f are the parameters of the variational posterior' contains a duplicated symbol; it should likely read 'μw,f and σw,f' or refer to σ^2_{w,f}.
[§4.2] The text says that utterances are cut into 2-second segments, but the BWRFN paragraph states that a single Monte Carlo sample is used for 'each 4-second speech segment'; please make the segment length consistent.
[§4.2] In the BWRFN-R-vector paragraph, the phrase 'same training settings as the RFN-vector network' appears to refer to the R-vector network; please correct the model name.
[§5, Table 1] The observation that deterministic WRFN increases HI-MIA EER relative to the plain R-vector (14.16% vs 12.65%), while BWRFN sharply reduces it (8.15%), is interesting and deserves an explicit discussion, since it suggests that the per-frequency weighting itself can hurt in the absence of the Bayesian training scheme.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: BWRFN is benchmarked against external datasets and baselines; the Eq. 10 vs. mean-inference gap is an implementation/interpretation issue, not a circular derivation.

full rationale

The derivation chain is self-contained and externally validated. BWRFN is defined in Eqs. 6-9 as a variational extension of WRFN, with variational parameters mu_w and sigma_w trained on speaker or anti-spoofing data; the reported EER, minDCF, and t-EER are computed on held-out domains (unseen HI-MIA, ASVspoof 5 development sets) and compared against independent baselines such as R-vector, ECAPA-TDNN, RFN-R-vector, and published ASVspoof 5 systems. No fitted parameter is renamed as a prediction, and no load-bearing claim is justified by a self-citation: RFN is adopted from reference [14], a non-overlapping prior work, and the ASVspoof 5 fusion in [33] is used only as a downstream component. The one internal inconsistency is that Eq. 10 defines a posterior predictive expectation over w, but Section 3 states "For simplicity, we use the expected value of embedding layer output during inference," which means sigma_w does not propagate to the test-time output. This weakens the paper's mechanistic claim that uncertainty accounts for the gains, because the observed improvements could come from the KL regularizer alone; however, this is an interpretation gap rather than a circular reduction, since the reported numbers compare the resulting deterministic network against external benchmarks. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on learned per-frequency weights, standard variational inference assumptions, and an inference choice that replaces the posterior predictive distribution with the mean weights. No invented entities are introduced.

free parameters (3)

Frequency-wise normalization weights w1, w2 (variational means mu_w) = Learned during training (values not reported)
These per-frequency weights are the core of WRFN/BWRFN; they are fit to the training data and determine the normalization output, so the method's performance depends on their fitted values.
Variational covariance parameters sigma_w = Learned during training (values not reported)
Diagonal posterior variances learned by optimizing the ELBO; they control sampling noise during training and thus the regularization effect.
Relaxation coefficient lambda = 0.5 (fixed)
Degree of relaxation between LN and IFN; set to 0.5 following RFN [14], not tuned for BWRFN.

assumptions (5)

domain assumption RFN as defined in Eq. 5 removes domain-dependent variability along the frequency axis
Adopted from [14] without re-derivation; the whole method builds on this premise.
ad hoc to paper Per-frequency weighting of LN and IFN outputs improves domain invariance
This is the central hypothesis motivating WRFN (Eq. 6); it is asserted, not derived.
standard math Variational posterior is diagonal Gaussian and prior is standard Gaussian (Eq. 9)
Standard variational inference assumptions; the paper invokes [19,20].
ad hoc to paper K=1 Monte Carlo sample during training is sufficient
Set for efficiency following [32]; the effect of estimator noise on training is not analyzed.
domain assumption Using the expected value of the embedding output at inference approximates the posterior predictive distribution (Eq. 10)
The paper replaces the integral over w with the mean weights, which discards posterior variance; this assumption is load-bearing for the 'uncertainty accounted for' narrative.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bayesian Learning for Domain-Invariant Speaker Verification and Anti-Spoofing." pith.science (2026). https://pith.science/paper/LCUR76RT

@misc{pith2026250607536,
  author       = {Pith},
  title        = {Pith review of: Bayesian Learning for Domain-Invariant Speaker Verification and Anti-Spoofing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LCUR76RT}},
  note         = {Machine review of arXiv:2506.07536}
}

read the original abstract

The performance of automatic speaker verification (ASV) and anti-spoofing drops seriously under real-world domain mismatch conditions. The relaxed instance frequency-wise normalization (RFN), which normalizes the frequency components based on the feature statistics along the time and channel axes, is a promising approach to reducing the domain dependence in the feature maps of a speaker embedding network. We advocate that the different frequencies should receive different weights and that the weights' uncertainty due to domain shift should be accounted for. To these ends, we propose leveraging variational inference to model the posterior distribution of the weights, which results in Bayesian weighted RFN (BWRFN). This approach overcomes the limitations of fixed-weight RFN, making it more effective under domain mismatch conditions. Extensive experiments on cross-dataset ASV, cross-TTS anti-spoofing, and spoofing-robust ASV show that BWRFN is significantly better than WRFN and RFN.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 45 canonical work pages

[26]

C. M. Bishop and N. M. Nasrabadi, Pattern Recognition and Ma- chine Learning. Springer, 2006, vol. 4

work page 2006
[1]

It is wid ely applied across various applications, including biometric au- thentication on personal smart devices [3]

Introduction Automatic speaker veriﬁcation (ASV) seeks to authenticate a speaker’s identity by analyzing their voice [1, 2]. It is wid ely applied across various applications, including biometric au- thentication on personal smart devices [3]. Recently, insp ired by the robust feature extraction abilities of deep neural ne t- works (DNNs), numerous deep l...

work page
[2]

Bayesian Learning for Domain-Invariant Speaker Verification and Anti-Spoofing

Preliminaries In this section, we deﬁne the notations and present the preli mi- nary on relaxed instance frequency-wise normalization. arXiv:2506.07536v1 [eess.AS] 9 Jun 2025 2.1. Relaxed Instance Frequency-wise Normalization An audio signal processing system typically applies spectr al- domain analysis on each audio channel on a frame-by-frame ba- sis. P...

work page Pith review arXiv 2025
[3]

5 interpolates between IFN and LN style normalization

Bayesian Relaxed Instance Frequency-wise Normalization Eq. 5 interpolates between IFN and LN style normalization. However, the degree of relaxation, λ, of individual frequency bins are ﬁxed and equal. We hypothesize that different fre- quency bins have different levels of domain dependence and should be weighted differently. We therefore extend RFN to wei...

work page
[4]

Evaluation Tasks The experiments were conducted on cross-dataset ASV tasks, anti-spooﬁng tasks, and spooﬁng-robust automatic speaker ver- iﬁcation (SASV) tasks

Experimental Setup 4.1. Evaluation Tasks The experiments were conducted on cross-dataset ASV tasks, anti-spooﬁng tasks, and spooﬁng-robust automatic speaker ver- iﬁcation (SASV) tasks. 4.1.1. Cross-Dataset Speaker V eriﬁcation We followed [12] for cross-dataset speaker veriﬁcation. To as- sess robustness, we designed a setup using CN-Celeb [22] (te st- in...

work page
[5]

For cross-dataset ASV evaluation, the EER is reported after the ﬁnal training epoc h

Experiments and Analysis This section presents the experimental results and analysi s of domain generalization performance. For cross-dataset ASV evaluation, the EER is reported after the ﬁnal training epoc h. For the closed condition of Track 1, both minDCF and EER were used as evaluation metrics, while the closed condition of Track 2 was assessed using ...

work page
[6]

Results demonstrate that accounting f or the uncertainty in the frequency-dependent weights can mit i- gate domain generalization issues

Conclusions and Future Work In this paper, we have introduced the BWRFN-R-vector net- work, which integrates Bayesian learning into the weightin g of the frequency components in the relaxed instance frequency - wise normalization. Results demonstrate that accounting f or the uncertainty in the frequency-dependent weights can mit i- gate domain generalizat...

work page
[7]

Linguistics, Artiﬁcial Intelligence and Lang uage and Speech Technologies: from Research to Applications

Acknowledgements This work was supported by the RGC of Hong Kong SAR, Grant No. PolyU 15228223 and the PolyU’s Research Student Attachment Programme. The work was supported by Czech Ministry of Education, Y outh and Sports (MoE) through the OP JAK project “Linguistics, Artiﬁcial Intelligence and Lang uage and Speech Technologies: from Research to Applicat...

work page

Show all 47 references

[8]

Mak and J.-T

M.-W. Mak and J.-T. Chien, Machine learning for speaker recog- nition. Cambridge University Press, 2020

2020
[9]

Modeling supraseg- mental information using ﬁnite difference network for end- to-end speaker veriﬁcation,

J. Li, M.-W. Mak, N. Y an, and L. Wang, “Modeling supraseg- mental information using ﬁnite difference network for end- to-end speaker veriﬁcation,” in Proc. APSIPA ASC, 2023, pp. 119–124

2023
[10]

Speaker veriﬁcation makes its debut in smartphone,

K. A. Lee, B. Ma, and H. Li, “Speaker veriﬁcation makes its debut in smartphone,” SLTC Newsletter, February 2013

2013
[11]

Speaker recognition based on dee p learning: An overview,

Z. Bai and X.-L. Zhang, “Speaker recognition based on dee p learning: An overview,” Neural Networks , vol. 140, pp. 65–99, 2021

2021
[12]

Spooﬁng-aware speaker veriﬁcation robust against domain and channel mismatches,

C. Zeng, X. Miao, X. Wang, E. Cooper, and J. Y amagishi, “Spooﬁng-aware speaker veriﬁcation robust against domain and channel mismatches,” in Proc. STL, 2024, pp. 1150–1157

2024
[13]

A comprehensive survey on transfer learning,

F. Zhuang, Z. Qi, K. Duan, D. Xi, Y . Zhu, H. Zhu, H. Xiong, an d Q. He, “A comprehensive survey on transfer learning,” Proceed- ings of the IEEE , vol. 109, no. 1, pp. 43–76, 2021

2021
[14]

Domain gener- alization: A survey,

K. Zhou, Z. Liu, Y . Qiao, T. Xiang, and C. C. Loy, “Domain gener- alization: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, pp. 4396–4415, 2022

2022
[15]

The CORAL+ algorith m for unsupervised domain adaptation of PLDA,

K. A. Lee, Q. Wang, and T. Koshinaka, “The CORAL+ algorith m for unsupervised domain adaptation of PLDA,” in Proc. ICASSP, 2019, pp. 5821–5825

2019
[16]

Un- supervised domain adaptation via domain adversarial train ing for speaker recognition,

Q. Wang, W. Rao, S. Sun, L. Xie, E. S. Chng, and H. Li, “Un- supervised domain adaptation via domain adversarial train ing for speaker recognition,” in Proc. ICASSP, 2018, pp. 4889–4893

2018
[17]

Unsupervised domain adaptation for i-vector speak er recognition,

D. Garcia-Romero, A. McCree, S. Shum, N. Brummer, and C. V a- quero, “Unsupervised domain adaptation for i-vector speak er recognition,” in Proc. Odyssey, 2014, pp. 260–264

2014
[18]

Learning to gen- eralize: Meta-learning for domain generalization,

D. Li, Y . Y ang, Y .-Z. Song, and T. Hospedales, “Learning to gen- eralize: Meta-learning for domain generalization,” in Proce. the AAAI, 2018, pp. 3490–3497

2018
[19]

Meta-generalization for domain-invariant speaker veriﬁ cation,

H. Zhang, L. Wang, K. A. Lee, M. Liu, J. Dang, and H. Meng, “Meta-generalization for domain-invariant speaker veriﬁ cation,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 31, pp. 1024–1036, 2023

2023
[20]

Mu- tual information-based embedding decoupling for generali zable speaker veriﬁcation,

J. Li, J. Han, S. Deng, T. Zheng, Y . He, and G. Zheng, “Mu- tual information-based embedding decoupling for generali zable speaker veriﬁcation,” in Proc. Interspeech, 2023, pp. 3147–3151

2023
[21]

Dom ain generalization with relaxed instance frequency-wise norm aliza- tion for multi-device acoustic scene classiﬁcation,

B. Kim, S. Y ang, J. Kim, H. Park, J. Lee, and S. Chang, “Dom ain generalization with relaxed instance frequency-wise norm aliza- tion for multi-device acoustic scene classiﬁcation,” in Proc. In- terspeech, 2022, pp. 2393–2397

2022
[22]

A review of uncertainty quantiﬁcation in deep learning: Techniques, applications and challenges,

M. Abdar, F. Pourpanah, S. Hussain, D. Rezazadegan, L. L iu, M. Ghavamzadeh, P . Fieguth, X. Cao, A. Khosravi, U. R. Acharya et al. , “A review of uncertainty quantiﬁcation in deep learning: Techniques, applications and challenges,” Information Fusion , vol. 76, pp. 243–297, 2021

2021
[23]

Gaussian process neural networks for speech recog ni- tion

M. W. Lam, S. Hu, X. Xie, S. Liu, J. Y u, R. Su, X. Liu, and H. Meng, “Gaussian process neural networks for speech recog ni- tion.” in Proc. Interspeech, 2018, pp. 1778–1782

2018
[24]

Bayesian neural network language modeling for speech reco gni- tion,

B. Xue, S. Hu, J. Xu, M. Geng, X. Liu, and H. Meng, “Bayesian neural network language modeling for speech reco gni- tion,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 2900–2917, 2022

2022
[25]

Bayesian x-vector: Bayesian neural network based x-vecto r sys- tem for speaker veriﬁcation,

X. Li, J. Zhong, J. Y u, S. Hu, X. Wu, X. Liu, and H. Meng, “Bayesian x-vector: Bayesian neural network based x-vecto r sys- tem for speaker veriﬁcation,” in Proc. Odyssey, 2020, pp. 365– 371

2020
[27]

V ariational dropout and the local reparameterization trick,

D. P . Kingma, T. Salimans, and M. Welling, “V ariational dropout and the local reparameterization trick,” Proc. NIPS , pp. 2575– 2583, 2015

2015
[28]

Weight uncertainty in neural network,

C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wiers tra, “Weight uncertainty in neural network,” in Proc. ICML, 2015, pp. 1613–1622

2015
[29]

CN-Celeb: multi-genre speaker recogni - tion,

L. Li, R. Liu, J. Kang, Y . Fan, H. Cui, Y . Cai, R. Vipperla, T. F. Zheng, and D. Wang, “CN-Celeb: multi-genre speaker recogni - tion,” Speech Communication, vol. 137, pp. 77–91, 2022

2022
[30]

Hi-mia: A far-ﬁeld text-depend ent speaker veriﬁcation database and the baselines,

X. Qin, H. Bu, and M. Li, “Hi-mia: A far-ﬁeld text-depend ent speaker veriﬁcation database and the baselines,” in Proc. ICASSP, 2020, pp. 7609–7613

2020
[31]

The interspeech 2020 far-ﬁeld speaker veriﬁcation challe nge,

X. Qin, M. Li, H. Bu, W. Rao, R. K. Das, S. Narayanan, and H. Li, “The interspeech 2020 far-ﬁeld speaker veriﬁcation challe nge,” arXiv preprint arXiv:2005.08046, 2020

2020 arXiv
[32]

V oxce leb: Large-scale speaker veriﬁcation in the wild,

A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “V oxce leb: Large-scale speaker veriﬁcation in the wild,” Computer Speech & Language, vol. 60, p. 101027, 2020

2020
[33]

ASVspoof 5: crowd- sourced speech data, deepfakes, and adversarial attacks at scale,

X. Wang, H. Delgado, H. Tak, J. weon Jung, H. jin Shim, M. Todisco, I. Kukanov, X. Liu, M. Sahidullah, T. H. Kinnunen , N. Evans, K. A. Lee, and J. Y amagishi, “ASVspoof 5: crowd- sourced speech data, deepfakes, and adversarial attacks at scale,” in Proc. ASVspoof 2024, 2024, pp. 1–8

2024
[34]

MUSAN: A music, speech , and noise corpus,

D. Snyder, G. Chen, and D. Povey, “MUSAN: A music, speech , and noise corpus,” arXiv preprint arXiv:1510.08484, 2015

2015 arXiv
[35]

Image method for efﬁcient ly sim- ulating small-room acoustics,

J. B. Allen and D. A. Berkley, “Image method for efﬁcient ly sim- ulating small-room acoustics,” The Journal of the Acoustical So- ciety of America, vol. 65, 1979

1979
[36]

Far-ﬁeld end-to-end text-dep endent speaker veriﬁcation based on mixed training data with trans fer learning and enrollment data augmentation

X. Qin, D. Cai, and M. Li, “Far-ﬁeld end-to-end text-dep endent speaker veriﬁcation based on mixed training data with trans fer learning and enrollment data augmentation.” in Proc. Interspeech, 2019, pp. 4045–4049

2019
[37]

Data augmentation using deep generative models for embedding based speaker recog- nition,

S. Wang, Y . Y ang, Z. Wu, Y . Qian, and K. Y u, “Data augmentation using deep generative models for embedding based speaker recog- nition,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 28, pp. 2598–2609, 2020

2020
[38]

ECAPA - TDNN: emphasized channel attention, propagation and aggre ga- tion in TDNN based speaker veriﬁcation,

B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA - TDNN: emphasized channel attention, propagation and aggre ga- tion in TDNN based speaker veriﬁcation,” in Interspeech, 2020, pp. 3830–3834

2020
[39]

Auto-encoding variational bayes,

D. P . Kingma, “Auto-encoding variational bayes,” arXiv preprint arXiv:1312.6114, 2013

2013 arXiv
[40]

BUT systems and analyses for the ASVspoof 5 Challenge,

J. Rohdin, L. Zhang, P . Oldˇ rich, V . Stanˇ ek, D. Mihola, J. Peng, T. Stafylakis, D. Beveraki, A. Silnova, J. Brukner, and L. Bu rget, “BUT systems and analyses for the ASVspoof 5 Challenge,” in Proc. ASVspoof, 2024, pp. 24–31

2024
[41]

R aw- boost: A raw data boosting and augmentation method applied t o automatic speaker veriﬁcation anti-spooﬁng,

H. Tak, M. Kamble, J. Patino, M. Todisco, and N. Evans, “R aw- boost: A raw data boosting and augmentation method applied t o automatic speaker veriﬁcation anti-spooﬁng,” in Proc. ICASSP , 2022, pp. 6382–6386

2022
[42]

IDV oice team system description for ASVSpoof5 Challenge,

A. Okhotnikov, I. Y akovlev, N. Torgashov, R. Makarov, E. G´ omez, P . Malov, A. Alenin, and A. Balykin, “IDV oice team system description for ASVSpoof5 Challenge,” in Proc. ASVspoof, 2024, pp. 43–47

2024
[43]

Whispeak speech deepfake dete ction systems for the ASVspoof5 Challenge,

P . Falez and T. Marteau, “Whispeak speech deepfake dete ction systems for the ASVspoof5 Challenge,” in Proc. ASVspoof, 2024, pp. 32–35

2024
[44]

ASVspoof 5 Chall enge: advanced ResNet architectures for robust voice spooﬁng det ec- tion,

A.-T. Dao, M. Rouvier, and D. Matrouf, “ASVspoof 5 Chall enge: advanced ResNet architectures for robust voice spooﬁng det ec- tion,” in Proc. ASVspoof, 2024

2024
[45]

Spooﬁng-robust speaker veriﬁcation using parallel embed ding fusion: BTU speech group’s approach for ASVspoof5 Challenge,

O. Kurnaz, S. C. Demirtas ¸, A. B. J. Mishra, and C. Hanilc ¸i, “Spooﬁng-robust speaker veriﬁcation using parallel embed ding fusion: BTU speech group’s approach for ASVspoof5 Challenge,” in Proc. ASVspoof, 2024, pp. 138–143

2024
[46]

AASIST3: KAN-enhanced AASIST speech deepfake detection using SSL features and additional regularization for the ASVspoof 20 24 Challenge,

K. Borodin, V . Kudryavtsev, D. Korzh, A. Eﬁmenko, G. Mkrtchian, M. Gorodnichev, and O. Y . Rogov, “AASIST3: KAN-enhanced AASIST speech deepfake detection using SSL features and additional regularization for the ASVspoof 20 24 Challenge,” in Proc. ASVspoof, 2024, pp. 48–55

2024
[47]

The SHADOW team submission to the ASVSpoof 2024 Challenge,

J. A. Villalba, T. Feng, T. Thebaud, J. Lee, S. Narayanan , and N. Dehak, “The SHADOW team submission to the ASVSpoof 2024 Challenge,” in Proc. ASVspoof, 2024, pp. 36–42

2024

Pith tools

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

[1] [26]

C. M. Bishop and N. M. Nasrabadi, Pattern Recognition and Ma- chine Learning. Springer, 2006, vol. 4

work page 2006

[2] [1]

It is wid ely applied across various applications, including biometric au- thentication on personal smart devices [3]

Introduction Automatic speaker veriﬁcation (ASV) seeks to authenticate a speaker’s identity by analyzing their voice [1, 2]. It is wid ely applied across various applications, including biometric au- thentication on personal smart devices [3]. Recently, insp ired by the robust feature extraction abilities of deep neural ne t- works (DNNs), numerous deep l...

work page

[3] [2]

Bayesian Learning for Domain-Invariant Speaker Verification and Anti-Spoofing

Preliminaries In this section, we deﬁne the notations and present the preli mi- nary on relaxed instance frequency-wise normalization. arXiv:2506.07536v1 [eess.AS] 9 Jun 2025 2.1. Relaxed Instance Frequency-wise Normalization An audio signal processing system typically applies spectr al- domain analysis on each audio channel on a frame-by-frame ba- sis. P...

work page Pith review arXiv 2025

[4] [3]

5 interpolates between IFN and LN style normalization

Bayesian Relaxed Instance Frequency-wise Normalization Eq. 5 interpolates between IFN and LN style normalization. However, the degree of relaxation, λ, of individual frequency bins are ﬁxed and equal. We hypothesize that different fre- quency bins have different levels of domain dependence and should be weighted differently. We therefore extend RFN to wei...

work page

[5] [4]

Evaluation Tasks The experiments were conducted on cross-dataset ASV tasks, anti-spooﬁng tasks, and spooﬁng-robust automatic speaker ver- iﬁcation (SASV) tasks

Experimental Setup 4.1. Evaluation Tasks The experiments were conducted on cross-dataset ASV tasks, anti-spooﬁng tasks, and spooﬁng-robust automatic speaker ver- iﬁcation (SASV) tasks. 4.1.1. Cross-Dataset Speaker V eriﬁcation We followed [12] for cross-dataset speaker veriﬁcation. To as- sess robustness, we designed a setup using CN-Celeb [22] (te st- in...

work page

[6] [5]

For cross-dataset ASV evaluation, the EER is reported after the ﬁnal training epoc h

Experiments and Analysis This section presents the experimental results and analysi s of domain generalization performance. For cross-dataset ASV evaluation, the EER is reported after the ﬁnal training epoc h. For the closed condition of Track 1, both minDCF and EER were used as evaluation metrics, while the closed condition of Track 2 was assessed using ...

work page

[7] [6]

Results demonstrate that accounting f or the uncertainty in the frequency-dependent weights can mit i- gate domain generalization issues

Conclusions and Future Work In this paper, we have introduced the BWRFN-R-vector net- work, which integrates Bayesian learning into the weightin g of the frequency components in the relaxed instance frequency - wise normalization. Results demonstrate that accounting f or the uncertainty in the frequency-dependent weights can mit i- gate domain generalizat...

work page

[8] [7]

Linguistics, Artiﬁcial Intelligence and Lang uage and Speech Technologies: from Research to Applications

Acknowledgements This work was supported by the RGC of Hong Kong SAR, Grant No. PolyU 15228223 and the PolyU’s Research Student Attachment Programme. The work was supported by Czech Ministry of Education, Y outh and Sports (MoE) through the OP JAK project “Linguistics, Artiﬁcial Intelligence and Lang uage and Speech Technologies: from Research to Applicat...

work page