Pith. sign in

REVIEW 5 minor 41 references

The paper claims that language-match and duration metadata, used only to calibrate fixed speaker-verification scores, improve decisions across language conditions and embedding systems.

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 · deepseek-v4-flash

2026-08-01 20:40 UTC pith:PEOSH7RC

load-bearing objection Well-controlled empirical backend paper: metadata-conditioned MLP fusion calibration gives modest but real gains, with the paper's own LID experiment marking the boundary.

arxiv 2607.16532 v1 pith:PEOSH7RC submitted 2026-07-17 eess.AS

AMECxSV: Adaptive Metadata-Driven Embedding-Fusion Calibration for X-Lingual Speaker Verification

Xin Wei , Shi He , Yihe Yuan , Huang-Cheng Chou , Sudarsana Reddy Kadiri , Shrikanth Narayanan This is my paper
classification eess.AS
keywords automatic speaker verificationembedding-fusion calibrationmetadata-driven calibrationcross-lingual speaker verificationscore fusionselective classificationabstentiontrial-level side information
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that fixed front-end scores in X-lingual speaker verification become more reliable when each trial's language-match status and duration statistics are used as calibration context in a small learned calibration network. On a development-derived speaker-disjoint split with six fixed embedding systems, this score+metadata fusion lowers full-coverage equal error rate from 2.15% to 1.85% while also improving calibration loss and detection costs; two external single-score sources improve from 3.15% to 2.42% and from 0.64% to 0.43%. A confidence-gated variant trades coverage for risk, reaching 0.10% accepted-trial error at 80% coverage internally and 0.03% at 79% coverage externally. Controls — a matched score-only head, shuffled metadata, and a metadata-only head — support the interpretation that the gain is calibration context, not added capacity or a metadata shortcut. The authors carefully bound the claim to metadata-available settings: with imperfect predicted language labels the benefit shrinks and at the strictest operating point reverses.

Core claim

AMECxSV is a backend that fixes the embedding extractors and learns a calibrated target posterior from the concatenation of trial scores and trial metadata. The central discovery is that adding aligned metadata — a language-match indicator and duration-reliability features (minimum duration, duration ratio, inverse-square-root short-duration risk) — to a flat-vector MLP calibration head improves both ranking and calibration over score-only fusion on the same architecture. On the paper's held-out split, the full-coverage head reduces EER from 2.15% to 1.85%, C_llr from 0.09 to 0.07, and actual detection cost (actDCF) at target prior 0.001 from 0.64 to 0.47, with paired speaker-clustered boots

What carries the argument

The carrying mechanism is a flat-vector MLP calibration head that consumes the concatenation of a trial's score vector (K fixed front-end scores) and a metadata vector made of the language-match indicator L_i and duration-reliability features r_i (minimum duration, duration ratio, inverse-square-root short-duration risk). A deterministic feature expansion builds score summaries, order statistics, pairwise score terms, score-metadata interactions, and squared metadata terms; the MLP output is affine-calibrated on a validation split and mapped to a target posterior. The work it does is to let the decision boundary shift per trial condition — language-matched vs cross-language, balanced vs imba

Load-bearing premise

The whole benefit rests on having accurate language-match metadata at inference time; with a language-identification model that agrees with true labels only about 78% of the time, the gain degrades and at the strictest operating point becomes worse than ignoring metadata.

What would settle it

Compute the EER of a head trained only on language-match and duration features (no scores) on the paper's held-out split; if it falls well below 50%, metadata are carrying speaker evidence and the calibration-context interpretation collapses.

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

If this is right

  • Full-coverage decisions improve across score sources: internal six-front-end EER drops from 2.15% to 1.85%, two external single-score sources drop from 3.15% to 2.42% and from 0.64% to 0.43%, and the dual external source reaches 0.43%, with parallel gains in calibration loss and detection cost.
  • The gain is causal to aligned metadata, not to model size or a metadata shortcut: a matched score-only head with identical capacity, a shuffled-metadata head, and a metadata-only head all fail to reproduce the improvement.
  • The confidence-gate variant gives a tunable coverage-risk trade-off: at 0.80 coverage the accepted-trial EER is 0.10% internally, and at 0.79 coverage the best external head reaches 0.03% — selective metrics that must be read together with coverage.
  • Metadata benefits are condition-dependent: short-duration trials improve in calibration but not in ranking, and the language-blind duration-only variant still beats the score-only control, giving a fallback when language labels are unavailable.
  • The method is a fixed-front-end backend: it leaves embeddings untouched and requires language and duration metadata at inference, so the paper's claims do not extend to language-blind evaluation, official challenge scoring, or embedding-level language mismatch.

Where Pith is reading between the lines

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

  • If the calibration-context interpretation is right, the same structure should transfer to other trial side information — estimated channel noise, enrollment quality, or sensor type — in verification tasks beyond speech; a natural next experiment is to feed such variables into the same MLP and look for the same EER and detection-cost pattern.
  • The predicted-language result hints at a practical reliability threshold: at roughly 78% language-identification agreement, the language-match feature no longer pays off at the strictest operating point, so deployment could gate language metadata on LID confidence and fall back to duration-only calibration below that threshold.
  • Because a metadata-only head stays near chance, the paper's design offers a diagnostic tool: per-condition score-to-posterior maps learned by the MLP could be inspected to see which language and duration conditions require the most recalibration, without any speaker-identity training.
  • A boundary test of the paper's own scope: run the trained score+metadata head on the protocol's official language-blind phase with the language-match feature stripped, keeping only duration metadata; the paper's duration-only variant (about 2.06% EER) predicts this head should still beat score-only, which would clarify whether reliability context generalizes beyond the development split.

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

0 major / 5 minor

Summary. The paper proposes AMECxSV, a metadata-driven calibration backend for X-lingual speaker verification that feeds fixed front-end trial scores together with trial metadata (binary language-match indicator and duration-derived features) into a small MLP to produce calibrated target posteriors, with an optional posterior-confidence abstention gate. The main empirical claim is that adding aligned metadata to a multi-score MLP backend improves full-coverage speaker-verification discrimination and calibration compared with a score-only MLP control. On a large speaker-disjoint held-out split derived from the TidyVoice X-ASV development protocol, AMEC-FC reduces full-coverage EER from 2.15% to 1.85% (internal), and external single-score experiments show reductions from 3.15% to 2.42% (Official) and 0.64% to 0.43% (LI-MSV). The paper includes a set of careful controls (matched score-only, shuffled metadata, metadata-only), bootstrap confidence intervals, and a speaker-clustered paired bootstrap for the primary comparison. Claims are explicitly scoped to metadata-available settings; an experiment with a frozen language-ID model shows that the benefit degrades and can reverse at P_tar=0.001 if the predicted language-match signal is noisy.

Significance. If the empirical results hold, the paper makes a useful practical contribution: a lightweight, architecture-agnostic calibration layer that can improve ASV decisions under metadata-available conditions without retraining front ends. The manuscript is unusually careful in its experimental design: speaker-disjoint splits, a matched architecture score-only control, a metadata permutation control, a metadata-only negative control, and speaker-clustered bootstrap intervals for the primary comparison. The authors explicitly distinguish full-coverage from selective metrics, separate development evidence from official challenge scores, and disclose the dependence on reliable language metadata. They also provide a codebase link. The main limitation—degradation with imperfect predicted language metadata—is disclosed and tested, though a quantitative LID-accuracy threshold would improve the deployment guidance. Overall, the central claim is well supported.

minor comments (5)
  1. [Appendix B, Table IX] The LID experiment is informative but the paper stops short of giving a practical threshold. Please add one or two sentences in Section V.G or Appendix B stating the LID agreement level at which AMEC-FC's full-coverage advantages vanish, if such a level can be read from the data, or explicitly state that no threshold is established and that reliable LID is required for stringent operating points.
  2. [Section III-D, Eq. (14)] The target coverage parameter Γ is not defined. Specify that Γ is the desired target coverage on the validation set and state the value used for the reported 0.80 coverage results.
  3. [Figure 3] The inline text around the scatter plot is difficult to parse (e.g., 'a 3.15 2.42 0.21 0.64 0.43 0.18 0.43 0.038'). Please provide a conventional figure with clear axis labels, group legends, and metric annotations, or move these numbers into a table.
  4. [Table VII, first row] The row label 'Linear score-only raw six raw scores' contains a redundant word; it should read 'six raw scores' or 'raw score vector'.
  5. [Section IV-C] The authors state that rich-feature MLP controls use a fixed seed and are not a multi-seed sweep. Since the main results also use a single seed, add a sentence acknowledging that trial-level bootstrap CIs do not capture seed variability, and optionally report a small multi-seed run for the primary comparison.

Circularity Check

0 steps flagged

No circularity: the AMEC-FC result is an empirical, control-checked comparison with fitted parameters evaluated on untouched held-out data; LID dependence is disclosed and tested, not circular.

full rationale

The paper contains no derivation that is equivalent to its inputs. The main comparison (AMEC-FC vs MultiScore-FC) is empirical: the MLP backend is trained on the calibration split, affine calibration (Eq. 10) is fit on validation, and all headline numbers are computed on the development-derived held-out test partition that is never used for training, model selection, threshold tuning, or feature selection. The explicit statement that the affine step 'is not the source of ranking gains' is consistent with EER being threshold-independent, and the matched score-only, shuffled-metadata, and metadata-only controls in Table IV rule out the main shortcut interpretations. Abstention results are repeatedly labeled as accepted-trial selective metrics with coverage, not full-coverage EER, so no metric is being renamed or re-labeled as a prediction. The only significant boundary condition—reliance on reliable language-match metadata at inference—is disclosed in Section V-G and directly stress-tested in Appendix B, Table IX with a frozen LID model; degraded performance there is a limitation, not a circularity. No load-bearing self-citation or imported uniqueness theorem appears; references to TidyVoice and other systems are external sources. Consequently the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on a trained MLP, affine rescaling, and a validation-tuned abstention threshold, all standard supervised calibrations on held-out splits. The key domain assumption is that language/duration metadata are available and reliable enough to condition calibration. No new physical or theoretical entities are introduced; AMECxSV is an algorithmic backend.

free parameters (4)
  • MLP backend parameters h_theta = trained on calibration split (not enumerated)
    The learned fusion/calibration mapping that produces the reported EER/Cllr gains.
  • Affine calibration alpha, beta = fitted on validation split (Eq. 10)
    Adjusts posterior scale; the paper states it is not the source of ranking gains.
  • Abstention coverage threshold eta_kappa = quantile chosen on validation for target coverage Gamma (Eq. 14)
    Defines the accepted set for all selective metrics; tuning on validation may affect accepted-trial EER.
  • Duration metadata map coefficients = [1, 1, 1] for dmin, dmin/dmax, 1/sqrt(dmin)
    Hand-chosen representation in Eq. 6 emphasizing short-duration risk; not fitted but ad hoc.
axioms (5)
  • standard math Logistic sigmoid maps affine logits to calibrated posteriors; scores are treated as log-likelihood ratios.
    Eq. 10 and Appendix C; standard ASV calibration assumption.
  • domain assumption Oracle language-match labels in the development protocol are accurate in the oracle condition.
    Section IV-A and Table IX; if language labels are noisy, the gain shrinks or reverses.
  • domain assumption Speaker-disjoint split with cross-partition non-target trials removed is a valid held-out evaluation.
    Section IV-A; disclosed but not equivalent to official evaluation results.
  • domain assumption Metadata (language match, duration) do not encode speaker identity and act only as calibration context.
    Supported by the metadata-only near-chance control, but assumed in the interpretation.
  • domain assumption Fixed front-end scores are reliable enough to fuse; calibration cannot fix embedding-level language mismatch.
    Section V-G limitation; the method inherits upstream score quality.

pith-pipeline@v1.3.0-alltime-deepseek · 12784 in / 12369 out tokens · 128093 ms · 2026-08-01T20:40:08.391878+00:00 · methodology

0 comments
read the original abstract

In X-lingual automatic speaker verification (ASV), fixed front-end scores vary in reliability with language match, duration, and score source. We propose AMECxSV, an adaptive metadata-driven embedding-fusion calibration backend for metadata-available settings. AMECxSV fuses trial scores with metadata to produce calibrated target posteriors, with optional posterior-confidence abstention; metadata serve as calibration context, not speaker evidence. On a development-derived speaker-disjoint held-out split, score+metadata heads reduce equal error rate (EER) from 3.15% to 2.42% for the official TidyVoice score source and from 0.64% to 0.43% for LI-MSV; the dual-score head reaches 0.43% full-coverage EER. At 0.79 coverage, abstention yields 0.03% accepted-trial EER, not a full-coverage metric. Matched score-only, metadata-permutation, and metadata-only controls support a calibration-context interpretation and limit claims to metadata-available scoring.

Figures

Figures reproduced from arXiv: 2607.16532 by Huang-Cheng Chou, Shi He, Shrikanth Narayanan, Sudarsana Reddy Kadiri, Xin Wei, Yihe Yuan.

Figure 1
Figure 1. Figure 1: Overview of AMECxSV pipeline: adaptive metadata-driven embedding-fusion calibration backend for fixed embedding front ends. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Coverage-risk curve for AMEC-ABS. Lower coverage retains higher [PITH_FULL_IMAGE:figures/full_fig_p006_2.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

41 extracted references · 6 linked inside Pith

  1. [1]

    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,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 5329–5333

    2018

  2. [2]

    ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in TDNN based speaker verification,

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

    2020

  3. [3]

    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,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023, pp. 1–5

    2023

  4. [4]

    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, pp. 5301–5305

    2023

  5. [5]

    ERes2NetV2: Boosting short-duration speaker verification performance with computational efficiency,

    Y . Chen, S. Zheng, H. Wang, L. Cheng, Q. Chen, S. Zhang, and J. Li, “ERes2NetV2: Boosting short-duration speaker verification performance with computational efficiency,” inProc. Interspeech, 2024, pp. 3245– 3249

    2024

  6. [6]

    WavLM: Large-scale self-supervised pre- training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Yoshioka, X. Xiao, J. Wu, L. Zhou, S. Ren, Y . Qian, Y . Qian, J. Wu, M. Zeng, X. Yu, and F. Wei, “WavLM: Large-scale self-supervised pre- training for full stack speech processing,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  7. [7]

    Tackling the score shift in cross-lingual speaker verification by exploiting language infor- mation,

    J. Thienpondt, B. Desplanques, and K. Demuynck, “Tackling the score shift in cross-lingual speaker verification by exploiting language infor- mation,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 7187–7191

    2022

  8. [8]

    Qual- ity measure functions for calibration of speaker recognition systems in various duration conditions,

    M. I. Mandasari, R. Saeidi, M. McLaren, and D. A. van Leeuwen, “Qual- ity measure functions for calibration of speaker recognition systems in various duration conditions,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 21, no. 11, pp. 2425–2438, 2013

    2013

  9. [9]

    Quality measures based calibration with duration and noise dependency for speaker recognition,

    M. I. Mandasari, R. Saeidi, and D. A. van Leeuwen, “Quality measures based calibration with duration and noise dependency for speaker recognition,”Speech Communication, vol. 72, pp. 126–137, 2015

    2015

  10. [10]

    TidyV oice: A curated multilingual dataset for speaker verification derived from common voice,

    A. Farhadipour, J. Marquenie, S. Madikeri, and E. Chodroff, “TidyV oice: A curated multilingual dataset for speaker verification derived from common voice,” arXiv preprint arXiv:2601.16358, 2026. [Online]. Available: https://arxiv.org/abs/2601.16358

    arXiv 2026

  11. [11]

    TidyV oice 2026 challenge evaluation plan,

    A. Farhadipour, J. Marquenie, S. Madikeri, T. Vukovic, V . Dellwo, K. Reid, F. M. Tyers, I. Siegert, and E. Chodroff, “TidyV oice 2026 challenge evaluation plan,” arXiv preprint arXiv:2601.21960, 2026. [Online]. Available: https://arxiv.org/abs/2601.21960

    arXiv 2026

  12. [12]

    The DET curve in assessment of detection task performance,

    A. Martin, G. Doddington, T. Kamm, M. Ordowski, and M. Przybocki, “The DET curve in assessment of detection task performance,” in Proc. European Conference on Speech Communication and Technology (Eurospeech), 1997, pp. 1895–1898

    1997

  13. [13]

    Application-independent evaluation of speaker detection,

    N. Br ¨ummer and J. du Preez, “Application-independent evaluation of speaker detection,”Computer Speech & Language, vol. 20, no. 2–3, pp. 230–275, 2006

    2006

  14. [14]

    The BOSARIS toolkit: Theory, algorithms and code for surviving the new DCF,

    N. Br ¨ummer and E. de Villiers, “The BOSARIS toolkit: Theory, algorithms and code for surviving the new DCF,” arXiv preprint arXiv:1304.2865, 2013. [Online]. Available: https://arxiv.org/abs/1304. 2865

    Pith/arXiv arXiv 2013

  15. [15]

    The PA V algorithm optimizes binary proper scoring rules,

    N. Br ¨ummer and J. du Preez, “The PA V algorithm optimizes binary proper scoring rules,” arXiv preprint arXiv:1304.2331, 2013. [Online]. Available: https://arxiv.org/abs/1304.2331

    Pith/arXiv arXiv 2013

  16. [16]

    A comparison of linear and non-linear calibrations for speaker recognition,

    N. Br ¨ummer, A. Swart, and D. A. van Leeuwen, “A comparison of linear and non-linear calibrations for speaker recognition,” inProc. The Speaker and Language Recognition Workshop (Odyssey), 2014, pp. 14– 18

    2014

  17. [17]

    Robustness of quality-based score calibration of speaker recognition systems with respect to low-SNR and short-duration conditions,

    A. Nautsch, R. Saeidi, C. Rathgeb, and C. Busch, “Robustness of quality-based score calibration of speaker recognition systems with respect to low-SNR and short-duration conditions,” inProc. The Speaker and Language Recognition Workshop (Odyssey), 2016, pp. 358–365

    2016

  18. [18]

    Trial- based calibration for speaker recognition in unseen conditions,

    Y . Lei, L. Ferrer, A. Lawson, M. McLaren, and N. Scheffer, “Trial- based calibration for speaker recognition in unseen conditions,” inProc. The Speaker and Language Recognition Workshop (Odyssey), 2014, pp. 19–25

    2014

  19. [19]

    A speaker verification backend with robust performance across conditions,

    L. Ferrer, M. McLaren, and N. Br ¨ummer, “A speaker verification backend with robust performance across conditions,”Computer Speech & Language, vol. 71, p. 101258, 2022

    2022

  20. [20]

    A generative approach to condition-aware score calibration for speaker verification,

    B. J. Borgstr ¨om, “A generative approach to condition-aware score calibration for speaker verification,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 891–901, 2023

    2023

  21. [21]

    The IDLAB V oxSRC-20 submission: Large margin fine-tuning and quality-aware score calibration in DNN based speaker verification,

    J. Thienpondt, B. Desplanques, and K. Demuynck, “The IDLAB V oxSRC-20 submission: Large margin fine-tuning and quality-aware score calibration in DNN based speaker verification,” inProc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 5814–5818

    2021

  22. [22]

    Selective classification for deep neural networks,

    Y . Geifman and R. El-Yaniv, “Selective classification for deep neural networks,” inAdvances in Neural Informa- tion Processing Systems (NeurIPS), 2017, pp. 4878–4887. [Online]. Available: https://proceedings.neurips.cc/paper/2017/hash/ 4a8423d5e91fda00bb7e46540e2b0cf1-Abstract.html

    2017

  23. [23]

    Toward fail-safe speaker recognition: Trial-based calibration with a reject option,

    L. Ferrer, M. K. Nandwana, M. McLaren, D. Castan, and A. Lawson, “Toward fail-safe speaker recognition: Trial-based calibration with a reject option,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 27, no. 1, pp. 140–153, 2019

    2019

  24. [24]

    V oxceleb: A large-scale speaker identification dataset,

    A. Nagrani, J. S. Chung, and A. Zisserman, “V oxceleb: A large-scale speaker identification dataset,” inProceedings of Interspeech 2017, 2017, pp. 2616–2620

    2017

  25. [25]

    Dis- entangled representation learning for multilingual speaker recognition,

    K. Nam, Y . Kim, J. Huh, H. S. Heo, J.-w. Jung, and J. S. Chung, “Dis- entangled representation learning for multilingual speaker recognition,” inProc. Interspeech, 2023, pp. 5316–5320

    2023

  26. [27]

    3D-Speaker-Toolkit: An open- source toolkit for multimodal speaker verification and diarization,

    Y . Chen, S. Zheng, H. Wang, L. Cheng, T. Zhu, R. Huang, C. Deng, Q. Chen, S. Zhang, W. Wang, and X. Li, “3D-Speaker-Toolkit: An open- source toolkit for multimodal speaker verification and diarization,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  27. [28]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv preprint arXiv:1607.06450, 2016. [Online]. Available: https: //arxiv.org/abs/1607.06450

    Pith/arXiv arXiv 2016

  28. [29]

    Dropout: A simple way to prevent neural networks from over- fitting,

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhut- dinov, “Dropout: A simple way to prevent neural networks from over- fitting,”Journal of Machine Learning Research, vol. 15, no. 56, pp. 1929–1958, 2014

    1929

  29. [30]

    Searching for activation functions,

    P. Ramachandran, B. Zoph, and Q. V . Le, “Searching for activation functions,” arXiv preprint arXiv:1710.05941, 2017. [Online]. Available: https://arxiv.org/abs/1710.05941

    Pith/arXiv arXiv 2017

  30. [31]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inInternational Conference on Learning Representations (ICLR), 2019. [Online]. Available: https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  31. [32]

    Simple attention mod- ule based speaker verification with iterative noisy label detection,

    X. Qin, N. Li, C. Weng, D. Su, and M. Li, “Simple attention mod- ule based speaker verification with iterative noisy label detection,” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6722–6726

    2022

  32. [33]

    Language-invariant multilingual speaker verification for the TidyV oice 2026 challenge,

    Z. Li, X. Miao, J. Liu, and M. Li, “Language-invariant multilingual speaker verification for the TidyV oice 2026 challenge,” arXiv preprint arXiv:2603.08092, 2026. [Online]. Available: https://arxiv.org/abs/2603. 08092

    arXiv 2026

  33. [34]

    Seamless: Multilingual expressive and streaming speech translation,

    L. Barrault, Y .-A. Chung, M. C. Meglioli, D. Dale, N. Dong, P.-A. Duquenne, H. Elsahar, H. Gong, K. Heffernan, J. Hoffman et al., “Seamless: Multilingual expressive and streaming speech translation,” arXiv preprint arXiv:2312.05187, 2023. [Online]. Available: https://arxiv.org/abs/2312.05187

    Pith/arXiv arXiv 2023

  34. [35]

    w2v-BERT: Combining contrastive learning and masked lan- guage modeling for self-supervised speech pre-training,

    Y .-A. Chung, Y . Zhang, W. Han, C.-C. Chiu, J. Qin, R. Pang, and Y . Wu, “w2v-BERT: Combining contrastive learning and masked lan- guage modeling for self-supervised speech pre-training,” inProc. IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2021, pp. 244–250

    2021

  35. [36]

    Efron and R

    B. Efron and R. J. Tibshirani,An Introduction to the Bootstrap. Chap- man and Hall/CRC, 1994

    1994

  36. [37]

    Confidence intervals for evaluation in machine learning,

    L. Ferrer and P. Riera, “Confidence intervals for evaluation in machine learning,” Computer software, accessed: 2026-07-01. [Online]. Available: https://github.com/luferrer/ConfidenceIntervals

    2026

  37. [38]

    Analysis of score normalization in multilingual speaker recognition,

    P. Matejka, O. Novotny, O. Plchot, L. Burget, M. D. Sanchez, and J. Cernocky, “Analysis of score normalization in multilingual speaker recognition,” inProc. Interspeech, 2017, pp. 1567–1571

    2017

  38. [39]

    Within-class covariance normalization for SVM-based speaker recognition,

    A. O. Hatch, S. Kajarekar, and A. Stolcke, “Within-class covariance normalization for SVM-based speaker recognition,” inProc. Interspeech, 2006, pp. 1471–1474

    2006

  39. [40]

    Front- end factor analysis for speaker verification,

    N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouellet, “Front- end factor analysis for speaker verification,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 19, no. 4, pp. 788–798, 2011

    2011

  40. [41]

    On the generalized distance in statistics,

    P. C. Mahalanobis, “On the generalized distance in statistics,”Proceed- ings of the National Institute of Sciences of India, vol. 2, no. 1, pp. 49–55, 1936. APPENDIXA STATISTICAL ANDMETADATA-CONTROLEVIDENCE These retained controls support the main metadata-calibration interpretation without adding long implementation tables. TABLE VI SPEAKER-CLUSTERED PAI...

    1936

  41. [2021]

    Available: https://arxiv.org/abs/2106.04624

    [Online]. Available: https://arxiv.org/abs/2106.04624

    Pith/arXiv arXiv