Pith. sign in

REVIEW 2 major objections 5 minor 52 references

Better uncertainty estimates for speaker embeddings come from joint inter- and intra-speaker hardness, plus source-prior calibration under domain shift.

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 ↗

T0 review · grok-4.5

2026-07-11 11:12 UTC pith:PSVFZHVJ

load-bearing objection Solid incremental extension of the authors' own U3-xi line: joint inter imes intra hardness scaling works in-domain; UCDA helps cross-domain but does not fully fix minDCF under uncertainty-aware scoring. the 2 major comments →

arxiv 2607.04937 v1 pith:PSVFZHVJ submitted 2026-07-06 cs.SD

Towards Robust Uncertainty-Aware Speaker Modeling

Junjie Li , Yang Xiao , Kong Aik Lee This is my paper
classification cs.SD
keywords speaker verificationuncertainty estimationuncertainty-aware softmaxdomain adaptationcross-domainspeaker embeddingsGaussian posterior pooling
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.

Speaker recognition systems turn variable-length speech into fixed embeddings, but those embeddings are less trustworthy when noise, channel, or duration make some frames unreliable. Prior uncertainty-aware models already treat each embedding as a Gaussian whose variance is meant to measure reliability, yet that variance is often poorly estimated and becomes miscalibrated when the test domain differs from training. This paper claims that uncertainty learning improves when the loss multiplies two complementary hardness signals—how separable the target speaker is from the nearest rival, and how tightly the sample aligns with its best class prototype—and feeds that product into an uncertainty-aware scale factor. Separately, it claims that a lightweight, label-free adaptation step that freezes the embedding network and only adjusts the uncertainty branch so that target-domain uncertainty vectors match a fixed source-domain Gaussian prior restores calibration across domains. On VoxCeleb1 and the mismatched CNCeleb set the combined approach yields lower equal-error rates and more reliable uncertainty-aware scores, giving practitioners a concrete way to keep reliability estimates honest under real acoustic variation.

Core claim

The authors establish that uncertainty estimates for speaker embeddings become more faithful to actual reliability when the uncertainty-aware scale factor is driven by the product of inter-speaker separability and intra-speaker prototype alignment, and that residual domain-induced miscalibration can be reduced by aligning target uncertainty vectors to a source-domain Gaussian prior while updating only the uncertainty pathway.

What carries the argument

Inter- and Intra-Speaker-Aware Uncertainty Softmax: the scale factor s_u is formed by multiplying the inter-speaker hardness gap by an exponential of the maximum prototype cosine, then using that product both to shape the bias matrix Lambda and to modulate the Mahalanobis-style confidence term that multiplies the softmax logits.

Load-bearing premise

That a single Gaussian fitted on source-domain uncertainty vectors remains a valid calibration target for an unseen target domain, and that correcting only the variance branch is enough to restore reliability without distorting the speaker embeddings themselves.

What would settle it

Train the full model on VoxCeleb2, then apply UCDA on CNCeleb (or another mismatched corpus) and check whether the Pearson correlation between predicted utterance-level uncertainty and measured discrimination quality fails to improve, or whether uncertainty-aware scoring raises rather than lowers EER and minDCF relative to the unadapted baseline.

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

If this is right

  • Speaker-verification pipelines can replace ordinary AAM-Softmax with the joint inter-/intra- hardness scale and obtain more reliable uncertainty without changing the encoder architecture.
  • Cross-domain deployments can run a short, unsupervised UCDA pass that updates only the pooling-layer variance branch, leaving embeddings and classifiers frozen.
  • Uncertainty-aware cosine scoring becomes usable under domain shift once UCDA has aligned the target uncertainty distribution to the source prior.
  • The same hardness product can be dropped into AM-Softmax or SphereFace2 losses, extending the method beyond the original angular-margin setting.

Where Pith is reading between the lines

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

  • Because UCDA never needs target labels or target statistics, it could be applied on-the-fly to streaming enrollment or test utterances as they arrive.
  • The same source-prior alignment idea may transfer to other biometric embeddings (face, voiceprint) whose uncertainty also drifts under channel change.
  • If the residual distribution mismatch after UCDA remains large, a mixture-of-Gaussians or longer-tailed source prior might be a natural next experiment.

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

2 major / 5 minor

Summary. The paper proposes a robust uncertainty-aware speaker modeling framework with two components. First, an Inter- and Intra-Speaker-Aware Uncertainty Softmax extends UAAM-Softmax by modulating the uncertainty-aware scale s_u with both inter-speaker cosine-gap hardness Λ_i and an intra-speaker prototype-alignment term Λ_j (Eqs. 11–13), so that predicted covariance better reflects embedding reliability. Second, Uncertainty-Calibrated Domain Adaptation (UCDA) performs label-free adaptation by freezing the encoder, mean branch and classifier and updating only the uncertainty (Gaussian-posterior) module so that target-domain utterance-level uncertainty vectors are pulled toward a fixed source-domain Gaussian prior via an NLL objective (Eqs. 14–15). Experiments on ECAPA-TDNN trained on VoxCeleb2 show consistent in-domain gains on VoxCeleb1 (especially Exp. 5 with uncertainty-aware scoring) and partial recovery of cross-domain performance on CNCeleb after UCDA (Tables I–II, Figs. 3–4).

Significance. If the claims hold, the work supplies a practical, lightweight route to more reliable uncertainty estimates for speaker verification under both clean and mismatched conditions. The joint hardness formulation is a natural, non-circular extension of the authors’ prior UAAM construction and is shown to transfer to AM-Softmax and SphereFace2; the correlation analysis in Fig. 3 supplies direct evidence that the learned uncertainty tracks discrimination quality. UCDA is attractive for deployment because it is unsupervised, freezes the speaker-discriminative path, and requires only a source-domain uncertainty prior. The honest reporting of Exp. 3 failure and residual minDCF degradation under uncertainty-aware scoring strengthens credibility. The contribution is incremental rather than foundational, but it is useful for the uncertainty-aware speaker-recognition community.

major comments (2)
  1. §IV, Eqs. (14)–(15) and the freeze statement: the central claim that UCDA “mitigates uncertainty miscalibration” rests on the assumption that a fixed source-domain Gaussian N(µ_src, σ²_src) estimated on VoxCeleb2 is a valid calibration target for CNCeleb. Table II shows that after UCDA the uncertainty-aware cosine score still yields minDCF values (0.524–0.528) that remain high relative to the cosine-only baseline, and Fig. 4 explicitly states that the domain discrepancy “is not completely eliminated.” Freezing the mean branch and classifier prevents any compensatory adjustment of embedding geometry. The paper should either (a) provide a quantitative calibration metric (e.g., ECE or reliability diagrams before/after UCDA) that demonstrates genuine miscalibration reduction, or (b) qualify the claim to “partial distribution alignment that improves EER under uncertainty-aware scoring.”
  2. Table I, Exp. 3 vs. Exp. 5: the bias-only joint formulation (Eq. 11) degrades performance while the multiplicative scale formulation (Eq. 13) succeeds. The authors attribute this to the larger stabilizing constant λ required for positive-definiteness. This is a load-bearing design choice; without an ablation that isolates λ (or an alternative positive-definite construction that keeps λ comparable across formulations), it remains unclear whether the intra-speaker term itself is beneficial or whether the gain is an artifact of how hardness is injected into s_u. A short controlled ablation would strengthen the claim that both inter- and intra-speaker factors are necessary.
minor comments (5)
  1. Throughout: numerous typos and inconsistent spellings (“coressponding,” “uncertianty,” “V oxCeleb,” “modulatings u,” “U 3-xi”). A careful proof-reading pass is needed.
  2. Fig. 2 caption and surrounding text: the figure is useful but the five difficulty settings are only loosely described; labeling each panel with the exact (Λ_i, Λ_j) regime would make the motivation clearer.
  3. Eq. (6)–(7): the claim that (6) is an efficient approximation of the Mahalanobis form (7) is plausible but not formally justified; a one-sentence derivation or reference would help.
  4. Table I gray rows: the uncertainty-aware cosine score sometimes drives minDCF to 1.000 on CNCeleb before UCDA; a brief discussion of score calibration or threshold sensitivity would aid interpretation.
  5. §V: the exact value of λ used for each experiment (0.5 vs 1.2) should be stated in the experimental protocol rather than only in the discussion of Exp. 3.

Circularity Check

1 steps flagged

Minor non-load-bearing self-citation of authors' prior UAAM/U3-xi base; new inter×intra factors and UCDA NLL are independent proposals validated by held-out empirical tables, not definitional reductions.

specific steps
  1. self citation load bearing [§III, Eqs. 5–7 and surrounding text; Table I Exp. 2]
    "To address this limitation, U3-xi [22] firstly incorporates uncertainty information explicitly into the Softmax formulation, which is called uncertianty-aware AAM (UAAM): LUAAM = ... su = ||ϕs|| / √((ϕs)⊤(Λ+Σs)ϕs) ... Exp. 2 reports the performance of previous uncertainty-aware framework, U3-xi trained with UAAM-Softmax [22]"

    The base uncertainty-aware scale and loss that all subsequent variants (1⃝–4⃝) modify is taken directly from the authors' concurrent/prior arXiv [22]; the paper's starting point and one of the main comparison systems are therefore self-cited rather than independently re-derived. This is not load-bearing for the new claims (the inter×intra multipliers and UCDA are original and evaluated separately), but it is the sole instance of the self-citation pattern.

full rationale

The paper is a standard empirical ML methods contribution in speaker verification. The Inter- and Intra-Speaker-Aware Uncertainty Softmax is explicitly constructed as multiplicative extensions (Eqs. 10, 13) of the authors' own prior UAAM scale (Eqs. 5–7 from [22]), and UCDA is a conventional source-prior NLL (Eqs. 14–15) with a freeze of non-variance parameters. Neither step claims a first-principles derivation, uniqueness theorem, or out-of-sample prediction that is forced by construction or by a fitted parameter renamed as a result. All performance claims rest on direct comparisons (Tables I–II) against the prior method and baselines on held-out VoxCeleb1/CNCeleb trials, with uncertainty-quality correlations (Fig. 3) and distribution plots (Fig. 4) reported as post-hoc observations. Self-citation of the base architecture is present and heavy but supplies only the starting point; the incremental factors and adaptation objective are new and independently tested. No self-definitional loop, no fitted-input-as-prediction, and no uniqueness imported from prior author work appear. Score 2 reflects the single minor self-citation pattern without elevating it to load-bearing circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central empirical claim rests on the linear-Gaussian frame model and diagonal covariance inherited from prior xi-vector work, on hand-chosen stabilizing constants λ that keep Λ positive-definite, on the modeling choice that source-domain uncertainty statistics form a suitable prior for any target domain, and on the freeze-only-uncertainty adaptation policy. No new physical entities are postulated; the invented objects are algorithmic (the joint hardness scale and the UCDA objective).

free parameters (4)
  • stabilizing constant λ in Λ = 0.5 (inter-only); 1.2 (joint)
    Hand-set to the smallest feasible positive-definite value (λ=0.5 for inter-only, λ=1.2 for joint inter×intra); the paper notes that larger λ weakens uncertainty modulation and explains Exp. 3’s failure.
  • AAM scale s and margin m schedule = s=32; m=0→0.2
    Default WeSpeaker values (s=32; m ramped 0→0.2 over epochs 20–40) control the base softmax geometry that the uncertainty scale multiplies.
  • UCDA learning rate and epoch count = 5 epochs; best LR=1e-7
    Adaptation is run for 5 epochs at LR ∈ {1e-5, 1e-6, 1e-7}; best reported RI uses 1e-7. These are free choices that materially affect Table II.
  • source-domain uncertainty prior (μ_src, σ²_src) = estimated once on VoxCeleb2 training set
    Empirical mean and variance of source utterance-level uncertainty vectors (Eq. 15); treated as fixed targets for the NLL objective on every target sample.
axioms (5)
  • domain assumption Frame-level features follow a linear-Gaussian model z_t = h + ε_t with diagonal precision L_t, yielding closed-form posterior mean and covariance (Eqs. 1–2).
    Inherited from xi-vector literature [21]; never re-justified here but required for the entire pooling and Σ_s pipeline.
  • domain assumption Covariance of the speaker embedding is diagonal, allowing element-wise BN/FC transforms (Eqs. 3–4).
    Stated explicitly in §II; full covariance would change both computation and the su approximation.
  • ad hoc to paper Sample difficulty for uncertainty supervision is adequately captured by the product of inter-speaker cosine gap Λ_i and exp(max_j cos θ_j) intra term.
    Motivated by Fig. 2 intuition but not derived; alternative hardness measures are not ablated beyond the four numbered variants.
  • ad hoc to paper Source-domain uncertainty distribution is a valid calibration target for target-domain utterances under domain shift.
    Core premise of UCDA (§IV); if source and target reliability patterns differ structurally, NLL alignment can mis-calibrate rather than help.
  • ad hoc to paper Freezing encoder, mean branch, BN/FC shared path, and classifier while updating only the uncertainty prediction module preserves speaker-discriminative geometry.
    Stated as design choice in §IV; no ablation that unfreezes other modules or compares to full fine-tuning.
invented entities (2)
  • Inter- and Intra-Speaker-Aware Uncertainty Softmax (joint hardness scale su) no independent evidence
    purpose: Supervise the variance branch with a scale that multiplies inter-speaker separability and intra-speaker prototype alignment so that predicted Σ_s better tracks embedding reliability.
    New loss construction (Eqs. 11–13); independent evidence is only the empirical tables in this paper, not an external falsifiable prediction.
  • Uncertainty-Calibrated Domain Adaptation (UCDA) no independent evidence
    purpose: Label-free alignment of target uncertainty vectors to a fixed source Gaussian prior by updating only the uncertainty head.
    New adaptation procedure (Eq. 14); evidence is limited to CNCeleb after VoxCeleb2 training in this manuscript.

pith-pipeline@v1.1.0-grok45 · 16824 in / 4025 out tokens · 35592 ms · 2026-07-11T11:12:06.856794+00:00 · methodology

0 comments
read the original abstract

Speaker embeddings aggregate frame-level acoustic features into compact representations for speaker recognition. Recent uncertainty-aware speaker modeling approaches further characterize the reliability of speaker embeddings by estimating their associated uncertainty. However, existing methods often suffer from inaccurate uncertainty estimation and uncertainty miscalibration under domain shifts. To address these challenges, we propose a robust uncertainty modeling framework from both estimation and adaptation perspectives. Specifically, we introduce an Inter- and Intra-Speaker-Aware Uncertainty Softmax that incorporates both inter-speaker separability and intra-speaker variability into uncertainty learning, enabling uncertainty estimates to better capture the reliability of speaker embeddings. Furthermore, we propose an Uncertainty-Calibrated Domain Adaptation (UCDA) framework to mitigate uncertainty miscalibration caused by domain mismatch. Extensive experiments on both in-domain and cross-domain benchmarks demonstrate that the proposed approach consistently improves uncertainty reliability and speaker recognition robustness.

Figures

Figures reproduced from arXiv: 2607.04937 by Junjie Li, Kong Aik Lee, Yang Xiao.

Figure 1
Figure 1. Figure 1: Architecture of the uncertainty-aware speaker model. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cosine similarity distributions between sample embeddings and class prototypes across different classification difficulty [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Relationship between utterance-level uncertainty and speaker discrimination quality on VoxCeleb1. Each point [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Uncertainty distribution on in-domain and cross [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

52 extracted references · 4 linked inside Pith

  1. [1]

    V oice biometrics,

    J. A. Markowitz, “V oice biometrics,”Communications of the ACM, vol. 43, no. 9, pp. 66–73, 2000

    2000

  2. [2]

    A personalized acoustic interface for wearable human–machine interaction,

    Z. Lin, G. Zhang, X. Xiao, C. Au, Y . Zhou, C. Sun, Z. Zhou, R. Yan, E. Fan, S. Siet al., “A personalized acoustic interface for wearable human–machine interaction,”Advanced Functional Materials, vol. 32, no. 9, p. 2109430, 2022

    2022

  3. [3]

    Speaker recognition for surveillance applica- tion,

    E. Kiktova and J. Juhar, “Speaker recognition for surveillance applica- tion,”Journal of Electrical and Electronics Engineering, vol. 8, no. 2, p. 19, 2015

    2015

  4. [4]

    Overview of speaker modeling and its applications: From the lens of deep speaker representation learning,

    S. Wang, Z. Chen, K. A. Lee, Y . Qian, and H. Li, “Overview of speaker modeling and its applications: From the lens of deep speaker representation learning,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  5. [5]

    Deep speaker: an end-to-end neural speaker embedding system,

    C. Li, X. Ma, B. Jiang, X. Li, X. Zhang, X. Liu, Y . Cao, A. Kannan, and Z. Zhu, “Deep speaker: an end-to-end neural speaker embedding system,”arXiv preprint arXiv:1705.02304, 2017

    Pith/arXiv arXiv 2017

  6. [6]

    X- vectors: Robust dnn embeddings for speaker recognition,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, “X- vectors: Robust dnn embeddings for speaker recognition,” in2018 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 5329–5333

    2018

  7. [7]

    Revisiting the statistics pooling layer in deep speaker embedding learning,

    S. Wang, Y . Yang, Y . Qian, and K. Yu, “Revisiting the statistics pooling layer in deep speaker embedding learning,” in2021 12th International Symposium on Chinese Spoken Language Processing (ISCSLP). IEEE, 2021, pp. 1–5

    2021

  8. [8]

    Attentive statistics pooling for deep speaker embedding,

    K. Okabe, T. Koshinaka, and K. Shinoda, “Attentive statistics pooling for deep speaker embedding,” inInterspeech 2018, 2018, pp. 2252–2256

    2018

  9. [9]

    Self-attentive speaker embeddings for text-independent speaker verification

    Y . Zhu, T. Ko, D. Snyder, B. Mak, and D. Povey, “Self-attentive speaker embeddings for text-independent speaker verification.” inInterspeech, vol. 2018, 2018, pp. 2–6

    2018

  10. [10]

    Self multi-head attention for speaker recognition,

    M. India, P. Safari, and J. Hernando, “Self multi-head attention for speaker recognition,” inInterspeech 2019, 2019, pp. 4305–4309

    2019

  11. [11]

    Multi- query multi-head attention pooling and inter-topk penalty for speaker verification,

    M. Zhao, Y . Ma, Y . Ding, Y . Zheng, M. Liu, and M. Xu, “Multi- query multi-head attention pooling and inter-topk penalty for speaker verification,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6737–6741

    2022

  12. [12]

    Ecapa-tdnn: Em- phasized channel attention, propagation and aggregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Em- phasized channel attention, propagation and aggregation in tdnn based speaker verification,” inInterspeech 2020, 2020, pp. 3830–3834

    2020

  13. [13]

    Vector-based attentive pooling for text-independent speaker verification

    Y . Wu, C. Guo, H. Gao, X. Hou, and J. Xu, “Vector-based attentive pooling for text-independent speaker verification.” inInterspeech, 2020, pp. 936–940

    2020

  14. [14]

    Data uncertainty learning in face recognition,

    J. Chang, Z. Lan, C. Cheng, and Y . Wei, “Data uncertainty learning in face recognition,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 5710–5719

    2020

  15. [15]

    Sdd-fiqa: Unsupervised face image quality assessment with similarity distribution distance,

    F.-Z. Ou, X. Chen, R. Zhang, Y . Huang, S. Li, J. Li, Y . Li, L. Cao, and Y .-G. Wang, “Sdd-fiqa: Unsupervised face image quality assessment with similarity distribution distance,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 7670– 7679

    2021

  16. [16]

    Probabilistic face embeddings,

    Y . Shi and A. K. Jain, “Probabilistic face embeddings,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6902–6911

    2019

  17. [17]

    Map: Multimodal uncertainty-aware vision-language pre-training model,

    Y . Ji, J. Wang, Y . Gong, L. Zhang, Y . Zhu, H. Wang, J. Zhang, T. Sakai, and Y . Yang, “Map: Multimodal uncertainty-aware vision-language pre-training model,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 23 262–23 271

    2023

  18. [18]

    Magface: A universal repre- sentation for face recognition and quality assessment,

    Q. Meng, S. Zhao, Z. Huang, and F. Zhou, “Magface: A universal repre- sentation for face recognition and quality assessment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 14 225–14 234

    2021

  19. [19]

    Fast and reliable probabilistic face embeddings based on constrained data uncertainty estimation,

    K. Chen, T. Yi, and Q. Lv, “Fast and reliable probabilistic face embeddings based on constrained data uncertainty estimation,”Image and Vision Computing, vol. 121, p. 104429, 2022

    2022

  20. [20]

    Xi+: Un- certainty supervision for robust speaker embedding,

    J. Li, K. A. Lee, D.-T. Truong, T. Liu, and M.-W. Mak, “Xi+: Un- certainty supervision for robust speaker embedding,” inICASSP 2026- 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 18 847–18 851

    2026

  21. [21]

    Xi-vector embedding for speaker recognition,

    K. A. Lee, Q. Wang, and T. Koshinaka, “Xi-vector embedding for speaker recognition,”IEEE Signal Processing Letters, vol. 28, pp. 1385– 1389, 2021

    2021

  22. [22]

    U3-xi: Pushing the boundaries of speaker recog- nition via incorporating uncertainty,

    J. Li and K. A. Lee, “U3-xi: Pushing the boundaries of speaker recog- nition via incorporating uncertainty,”arXiv preprint arXiv:2601.15719, 2026

    arXiv 2026

  23. [23]

    Disentangling voice and content with self-supervision for speaker recognition,

    T. Liu, K. A. Lee, Q. Wang, and H. Li, “Disentangling voice and content with self-supervision for speaker recognition,”Advances in Neural Information Processing Systems, vol. 36, pp. 50 221–50 236, 2023

    2023

  24. [24]

    Incorporating uncertainty from speaker embedding estimation to speaker verification,

    Q. Wang, K. A. Lee, and T. Liu, “Incorporating uncertainty from speaker embedding estimation to speaker verification,” inICASSP 2023- 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  25. [25]

    Cosine scoring with uncertainty for neural speaker embedding,

    Q. Wang and K. A. Lee, “Cosine scoring with uncertainty for neural speaker embedding,”IEEE Signal Processing Letters, vol. 31, pp. 845– 849, 2024

    2024

  26. [26]

    Scoring of Large-Margin Embeddings for Speaker Verification: Cosine or PLDA?

    Q. Wang, K. A. Lee, and T. Liu, “Scoring of Large-Margin Embeddings for Speaker Verification: Cosine or PLDA?” inInterspeech 2022, 2022, pp. 600–604

    2022

  27. [27]

    Analysis of the abc classification backends for nist sre24,

    S. Cumani, A. Silnova, S. Barahona, L. Mo ˇsner, O. Plchot, and J. Ro- hdin, “Analysis of the abc classification backends for nist sre24,” in Proc. Interspeech 2025, 2025, pp. 3978–3982

    2025

  28. [28]

    Can you trust your model’s uncer- tainty? evaluating predictive uncertainty under dataset shift,

    Y . Ovadia, E. Fertig, J. Ren, Z. Nado, D. Sculley, S. Nowozin, J. Dillon, B. Lakshminarayanan, and J. Snoek, “Can you trust your model’s uncer- tainty? evaluating predictive uncertainty under dataset shift,”Advances in neural information processing systems, vol. 32, 2019

    2019

  29. [29]

    Pseudo-speaker distribution learning in voice anonymization,

    L. Chen, W. Gu, K. A. Lee, W. Guo, and Z.-H. Ling, “Pseudo-speaker distribution learning in voice anonymization,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  30. [30]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 4690– 4699

    2019

  31. [31]

    Scaleface: Uncertainty-aware deep metric learning,

    R. Kail, K. Fedyanin, N. Muravev, A. Zaytsev, and M. Panov, “Scaleface: Uncertainty-aware deep metric learning,” in2023 IEEE 10th Interna- tional Conference on Data Science and Advanced Analytics (DSAA), 2023, pp. 1–10

    2023

  32. [32]

    Improving training and inference of face recognition models via random temperature scaling,

    L. Shang, M. Huang, W. Shi, Y . Liu, Y . Liu, W. Steven, B. Sun, X. Xie, and Y . Qiao, “Improving training and inference of face recognition models via random temperature scaling,” inProceedings of the Thirty- Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth...

    2023

  33. [33]

    Adaspeaker: Learning discriminative speaker representations with gradient-aware adaptive scaling,

    J. Liu, X. Wang, and J. Meng, “Adaspeaker: Learning discriminative speaker representations with gradient-aware adaptive scaling,” inPro- ceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 660–668

    2025

  34. [34]

    Wespeaker: A research and production oriented speaker embedding learning toolkit,

    H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xiang, Y . Deng, and Y . Qian, “Wespeaker: A research and production oriented speaker embedding learning toolkit,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  35. [35]

    Additive margin softmax for face verification,

    F. Wang, J. Cheng, W. Liu, and H. Liu, “Additive margin softmax for face verification,”IEEE Signal Processing Letters, vol. 25, no. 7, pp. 926–930, 2018

    2018

  36. [36]

    Cosface: Large margin cosine loss for deep face recognition,

    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 recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 5265–5274

    2018

  37. [37]

    Exploring binary classification loss for speaker verification,

    B. Han, Z. Chen, and Y . Qian, “Exploring binary classification loss for speaker verification,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  38. [38]

    Sphereface2: Binary classification is all you need for deep face recognition,

    Y . Wen, W. Liu, A. Weller, B. Raj, and R. Singh, “Sphereface2: Binary classification is all you need for deep face recognition,” inICLR, 2022

    2022

  39. [39]

    Cam++: A fast and efficient network for speaker verification using context-aware masking,

    H. Wang, S. Zheng, Y . Chen, L. Cheng, and Q. Chen, “Cam++: A fast and efficient network for speaker verification using context-aware masking,” inProc. Interspeech 2023, 2023, pp. 5301–5305

    2023

  40. [40]

    Golden gemini is all you need: Finding the sweet spots for speaker verification,

    T. Liu, K. A. Lee, Q. Wang, and H. Li, “Golden gemini is all you need: Finding the sweet spots for speaker verification,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 2324–2337, 2024

    2024

  41. [41]

    BUT system description to voxceleb speaker recognition challenge 2019,

    H. Zeinali, S. Wang, A. Silnova, P. Mat ˇejka, and O. Plchot, “BUT system description to voxceleb speaker recognition challenge 2019,” arXiv preprint arXiv:1910.12592, 2019

    Pith/arXiv arXiv 2019

  42. [42]

    Deep coral: Correlation alignment for deep do- main adaptation,

    B. Sun and K. Saenko, “Deep coral: Correlation alignment for deep do- main adaptation,” inEuropean conference on computer vision. Springer, 2016, pp. 443–450

    2016

  43. [43]

    Learning transferable features with deep adaptation networks,

    M. Long, Y . Cao, J. Wang, and M. Jordan, “Learning transferable features with deep adaptation networks,” inInternational conference on machine learning. PMLR, 2015, pp. 97–105

    2015

  44. [44]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. March, and V . Lempitsky, “Domain-adversarial training of neural networks,”Journal of machine learning research, vol. 17, no. 59, pp. 1–35, 2016

    2016

  45. [45]

    Robust cross- domain speaker verification with multi-level domain adapters,

    W. Huang, B. Han, S. Wang, Z. Chen, and Y . Qian, “Robust cross- domain speaker verification with multi-level domain adapters,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 781–11 785

    2024

  46. [46]

    Self-supervised learning based domain adaptation for robust speaker verification,

    Z. Chen, S. Wang, and Y . Qian, “Self-supervised learning based domain adaptation for robust speaker verification,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 5834–5838

    2021

  47. [47]

    What uncertainties do we need in bayesian deep learning for computer vision?

    A. Kendall and Y . Gal, “What uncertainties do we need in bayesian deep learning for computer vision?”Advances in neural information processing systems, vol. 30, 2017

    2017

  48. [48]

    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

    Pith/arXiv arXiv 2015

  49. [49]

    A study on data augmentation of reverberant speech for robust speech recogni- tion,

    T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recogni- tion,” in2017 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2017, pp. 5220–5224

    2017

  50. [50]

    V oxCeleb2: Deep Speaker Recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxCeleb2: Deep Speaker Recognition,” inInterspeech 2018, 2018, pp. 1086–1090

    2018

  51. [51]

    V oxceleb: a large-scale speaker identification dataset,

    A. Nagrani, J. S. Chung, and A. Zisserman, “V oxceleb: a large-scale speaker identification dataset,”arXiv preprint arXiv:1706.08612, 2017

    Pith/arXiv arXiv 2017

  52. [52]

    Cn-celeb: a challenging chinese speaker recogni- tion dataset,

    Y . Fan, J. Kang, L. Li, K. Li, H. Chen, S. Cheng, P. Zhang, Z. Zhou, Y . Cai, and D. Wang, “Cn-celeb: a challenging chinese speaker recogni- tion dataset,” inICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp. 7604–7608

    2020