Pith. sign in

REVIEW 3 major objections 6 minor 26 references

In the TidyVoice2026 cross-lingual speaker-verification challenge, a post-hoc projection that removes the thirty principal language directions from speaker embeddings, followed by adaptive symmetric score normalization, outperformed the bas

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 04:08 UTC pith:YL7XOZL3

load-bearing objection Useful challenge paper with a solid held-out signal for NAP, but the leaderboard 'wins' claim is contradicted by the paper's own discussion and needs fixing before publication. the 3 major comments →

arxiv 2607.22923 v1 pith:YL7XOZL3 submitted 2026-07-24 cs.CL

Simple Language Normalization Wins: Cross-Lingual Speaker Verification for the TidyVoice 2026 Challenge

Nina Hosseini-Kivanani This is my paper
classification cs.CL
keywords speaker verificationcross-lingual mismatchnuisance attribute projectionlanguage subspacescore normalizationmultilingual speechTidyVoice2026embedding compensation
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.

Cross-lingual speaker verification fails when the same speaker's enrollment and test speech are in different languages. This paper argues that the damage is mostly confined to a low-dimensional subspace of the embedding space, and that removing it with a classical back-end transform—Nuisance Attribute Projection (NAP)—is enough to beat far more complex systems. Using the official pretrained encoder as-is, the author estimates language directions from same-speaker cross-language differences in the development data, projects them out, and re-scores with adaptive symmetric normalization. Development equal-error rate drops from 2.97% with cosine scoring and 2.70% with AS-Norm to 2.18%, and the same system posts the best evaluation score of 8.40 on the challenge's unseen languages. The paper's broader point: before building new encoders or adding adversarial objectives, try removing the nuisance directions you already know are there.

Core claim

The author's central claim is that language mismatch in a strong 256-dimensional speaker embedding is a low-rank additive nuisance. Difference vectors between the same speaker's mean embeddings in two different languages concentrate along roughly thirty dominant directions; removing those directions with NAP and applying AS-Norm turns the official baseline encoder into the best system in the TidyVoice2026 challenge: 2.18% development EER and an evaluation score of 8.40. Because the projection is learned entirely from development speakers, needs no language labels at test time, and adds one matrix-vector multiplication per embedding, the result implies the gain comes from the back-end, not fr

What carries the argument

The load-bearing object is the NAP projection matrix P = I − V V^T, where V is the 256×k matrix whose columns are the top k eigenvectors of the covariance of L2-normalized cross-language, same-speaker mean-embedding differences. The dimension k is swept over development trials and fixed at k=30. Applying P to each embedding, L2-normalizing, and then scoring with cosine similarity plus adaptive symmetric score normalization (AS-Norm, a cohort-based score normalization using top cohort scores on both enrollment and test sides) strips the estimated language subspace from the representation. A single 256×256 eigendecomposition estimates P, and applying it costs one matrix–vector multiplication p

Load-bearing premise

The load-bearing premise is that the 30-dimensional language subspace estimated from development speakers who appear in at least two languages—and chosen by sweeping development EER—captures language-general directions that also span the 38 unseen evaluation languages; the paper itself notes that the global estimate may under-compensate very low-resource languages, so if those directions are dataset-specific artifacts rather than language-universal phonotactic/prosodic struct

What would settle it

Randomly permute the language labels used to form same-speaker difference vectors and rerun NAP: if the shuffled projection still reduces development EER, the improvement is not due to language structure. Stronger still, estimate P from only a subset of the development languages and score the remaining development languages (or the official evaluation set): if EER does not drop relative to AS-Norm alone, the claim that the subspace transfers to unseen languages is falsified.

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

If this is right

  • The method is encoder-agnostic: any fixed-dimensional embedding extractor can be language-compensated without retraining or language labels at inference time.
  • The largest absolute gains are in the hardest trials—target speakers switching language while impostors stay in one language—so the transform is specifically counteracting language mismatch rather than broadly denoising scores.
  • Adversarial language-invariance training and models trained from scratch on the challenge data perform worse than the pretrained encoder plus NAP, making a post-hoc back-end correction the more cost-effective route in this setting.
  • The held-out evaluation ordering matches development, and the best system's score of 8.40 tops the leaderboard, indicating the development-estimated subspace transfers to unseen languages.

Where Pith is reading between the lines

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

  • If language directions are universal across languages, then a NAP matrix estimated from any sufficiently diverse subset of languages should help on entirely held-out languages; this can be tested by splitting the development languages into two parts and checking whether the projection trained on one half improves the other half.
  • The k=30 elbow in the sweep suggests the language information occupies a fairly well-defined low-rank subspace; a principled rule for choosing k from the eigenvalue spectrum (for example, a gap or a variance threshold) might replace the trial sweep.
  • A single global linear projection may under-compensate low-resource languages whose language directions differ from the majority—the paper flags this itself—so family-specific or hierarchical nuisance subspaces are a natural testable extension.
  • For future cross-lingual evaluations, this result suggests adding NAP-plus-score-normalization as a default baseline before investing in adversarial or self-supervised adaptation; it is cheap, transparent, and hard to beat here.

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

3 major / 6 minor

Summary. The paper reports the authors' TidyVoice 2026 challenge submission. Starting from the official SimAM-ResNet34 encoder, it estimates a low-dimensional language subspace from within-speaker cross-language embedding differences and applies Nuisance Attribute Projection (NAP) to remove those directions, followed by Adaptive Symmetric score normalization (AS-Norm). The paper reports that this reduces development EER from 2.97% (cosine) and 2.70% (AS-Norm) to 2.18%, and obtains a Codabench evaluation score of 8.40, which the authors claim 'ranks at the top of the leaderboard'. It also reports comparisons with ECAPA encoders, a DANN language-adversarial variant, WavLM front-ends, and LDA/PLDA back-ends.

Significance. The core idea is simple and practically relevant: a post-hoc, training-free NAP back-end can reduce cross-lingual mismatch in a strong neural embedding system. The held-out Codabench score (8.40 vs 8.92 for the same encoder with AS-Norm only) independently supports the direction of the claim despite the development-set hyperparameter selection. The paper also provides useful per-condition analysis of language-match/mismatch trials and claims release of a complete pipeline. However, the leaderboard-rank claim is not supported by any external evidence and is internally contradicted in Section 4, and the main evaluation metric is ambiguously defined. These issues affect the central message and need to be resolved before the paper can be recommended for publication.

major comments (3)
  1. [§1, §4, Table 1] The title and abstract assert that the system 'ranks at the top of the leaderboard' and that 'simple language normalization wins'. Section 4, however, states: 'Even when other systems obtain slightly better absolute scores on the leaderboard...' and Table 1 lists only the authors' own submitted systems, with no leaderboard excerpt or comparison to other teams. The manuscript therefore provides no evidence about the rank among all challenge participants. The supported claim is the internal comparison: NAP+AS-Norm outperforms the authors' own Champion+AS-Norm (8.92) and other own back-ends. This distinction is load-bearing for the title, abstract, and introduction. Please provide the actual leaderboard standings or revise all claims to the supported within-system comparison.
  2. [§2.3, §3.1] The main method's hyperparameters are selected on the development set: the language-subspace dimension k is swept over development trials and fixed at k=30, and the AS-Norm cohort size and top-N (N=300 vs N=400) are also development choices. The reported headline improvement to 2.18% EER is therefore in-sample with respect to these choices, and no error bars or repeated runs are given. The held-out Codabench score (8.40 vs 8.92 for the same encoder with AS-Norm only) mitigates the overfitting concern for the qualitative conclusion, but the manuscript should explicitly state that the development numbers are selected-on-development, report the full k-sweep, and ideally provide confidence intervals or a validation split for the development EER.
  3. [§3.3, Table 1] The evaluation metric is ambiguous. The text refers to 'a Codabench evaluation score of 8.40', then to 'the overall evaluation EER of 8.40%', while Table 1 labels the column 'Eval score' (lower is better) and separately gives EER/minDCF rows for Task 1 and Task 2, with 'Overall 8.40 / 0.6485 12.15 / 0.6299'. The reader cannot determine whether 8.40 is the Task 1 overall EER, a combined official score, or something else. Because the central claim rests on this number, the official scoring definition and the relationship between 'Eval score' and the per-condition EER rows must be specified precisely.
minor comments (6)
  1. [Abstract, §1, §2.1] The abstract and introduction say evaluation contains 2,200 speakers in 40 unseen languages, while §2.1 says 38 unseen languages. Reconcile this discrepancy.
  2. [§2.3] The difference vectors are L2-normalized and stacked into Δ, but it is not stated whether the columns are mean-centered before computing the empirical covariance C. If the mean of the difference vectors is not exactly zero, the top eigenvectors may be biased by a constant offset rather than language directions. Please clarify.
  3. [§2.4 / Availability] The paper states that 'We release a complete and vectorized preprocessing, scoring, and submission pipeline', but no URL, repository name, or availability statement is provided. Add a link or state where the code can be obtained.
  4. [Table 1] Define 'Task 1' and 'Task 2' in the caption or text, and explicitly state the units of the 'Eval score' row. Also explain why the 'Eval score' column shows a single number while the bottom half reports EER/minDCF for two tasks.
  5. [§3.2] The sentence 'Table 1 and report EER for each combination' appears ungrammatical; presumably it should read 'Table 1 reports EER for each combination.'
  6. [§2.2] The term 'SimAM ResNet34 champion' is used for the baseline encoder. Clarify whether this is the official challenge champion model or the authors' own implementation, and cite the exact source.

Circularity Check

1 steps flagged

Development EER after NAP is in-sample because k=30 is tuned on the same dev trials, but the held-out Codabench score gives independent support; no equation-level circularity.

specific steps
  1. other [Section 2.3 (NAP sweep) and Section 3.1 (Development set performance)]
    "A sweep over k∈{1,2,3,5,8,10,15,20,30,40,50} on development trials with AS-Norm shows monotonic improvements up to k=30 (2.18% EER), then slight degradation at k=50. We fix k=30 for all NAP experiments. ... Applying NAP with k=30 to champion embeddings and re-running AS-Norm reduces EER to 2.18% with a minDCF of 0.73, our best development performance."

    The 2.18% development EER used as evidence that NAP 'reduces development EER' is the minimum of the k-sweep performed on those same development trials: k=30 is selected because it yields that EER. Reporting the sweep minimum as the demonstration of improvement is in-sample model selection rather than an independent validation. This is bounded, because the subsequent Codabench evaluation score of 8.40 was obtained with a fixed configuration on a held-out set, so the central claim has independent external grounding.

full rationale

The paper's derivation chain is not circular at the equation level: NAP is a standard nuisance-subspace projection, the language subspace is estimated from cross-language same-speaker differences, and the reported evaluation score comes from an external Codabench leaderboard with a configuration fixed before evaluation. The only self-fitting element is the development-side presentation: the hyperparameter k and AS-Norm settings are tuned on the development trials, and the same trials' EER is then reported as the improvement, which is statistical selection rather than a prediction. This does not vitiate the held-out result. The leaderboard rank claim in Section 1 is internally contradicted by Section 4's admission that other systems may score better, and Table 1 lists only the authors' own runs; however, that is an evidence/correctness gap, not circularity. No self-citation chain, renamed known result, or uniqueness-imported-from-authors pattern is present. Overall, the central claim has independent content through the held-out Codabench score, so the circularity score is low.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The central claim rests on a small set of statistical modeling choices: one selected hyperparameter k=30, AS-Norm cohort settings, and two domain assumptions about the transferability of a language subspace estimated from development data to unseen languages. The paper also trusts the official baseline and challenge infrastructure. No new physical or theoretical entities are introduced; the 'language subspace' is a PCA subspace estimated from data, not a postulated entity.

free parameters (2)
  • k (number of projected language directions) = 30
    Chosen by sweeping k∈{1,2,3,5,8,10,15,20,30,40,50} on development trials (§2.3); selected for lowest development EER, then fixed for all NAP experiments and evaluation. This is the load-bearing number for the NAP improvement.
  • AS-Norm cohort size and top-N = 500 speakers, ~12 utterances each; N=300 for baseline, N=400 for NAP
    Back-end hyperparameters chosen per configuration with no sensitivity analysis; they affect the score normalization but not the NAP subspace itself (§2.3).
axioms (3)
  • domain assumption Difference vectors between mean embeddings of the same speaker in two languages isolate language-induced nuisance variation (Eq. 1); channel/session variability is not a confound.
    If the same-speaker differences are dominated by recording conditions, PCA will not capture language and NAP will remove speaker/recording information instead. §2.3.
  • domain assumption A low-rank subspace computed from development languages transfers to evaluation languages not seen in training (the paper says both 38 and 40 unseen languages).
    This is the load-bearing generalization premise; the paper's evidence is one official held-out leaderboard score. §2.3, §3.3.
  • domain assumption The reported baseline and challenge setup (SimAM-ResNet34 pretrained on VoxBlink2/VoxCeleb2, official EERs, hidden evaluation labels) are accurate as provided.
    All improvements are relative gains over the official baseline; the paper does not train the baseline from scratch or audit the challenge infrastructure. §2.2.

pith-pipeline@v1.3.0-alltime-deepseek · 7746 in / 14720 out tokens · 142140 ms · 2026-08-01T04:08:10.481658+00:00 · methodology

0 comments
read the original abstract

Cross-lingual mismatch remains a key source of overall degradation in modern speaker verification. The TidyVoice2026 Challenge targets this setting with text-independent verification, comprising 3,666 training and 808 development speakers in 40 languages and 2,200 evaluation speakers in 38 unseen languages, without language labels at test time. Starting from the official SimAM-ResNet34 baseline pretrained on VoxBlink2 and VoxCeleb2 and fine-tuned on TidyVoice, we revisit Nuisance Attribute Projection (NAP) as a simple language-normalization step in the embedding space. We estimate a compact language subspace from cross-language same-speaker differences and project embeddings onto its orthogonal complement before cosine scoring with Adaptive Symmetric score normalization. This reduces development EER from 2.97\% with cosine and 2.70\% with AS-Norm to 2.18\% and yields a Codabench evaluation score of 8.40, showing that simple back-end language normalization can rival more complex systems.

Figures

Figures reproduced from arXiv: 2607.22923 by Nina Hosseini-Kivanani.

Figure 1
Figure 1. Figure 1: Development set results. (a) EER (%) for all systems, sorted best (top) to worst (bottom). Colors denote encoder fam￾ily, and hatching indicates the scoring back-end (cosine, AS￾Norm, or NAP+AS). (b) Champion system (SimAM-ResNet34) EER under three scoring configurations; annotations show ab￾solute improvement over cosine scoring. 2.4. Submitted systems All official submissions build on SimAM ResNet34 embe… view at source ↗
Figure 2
Figure 2. Figure 2: NAP sweep and DET curves for the champion encoder. low-dimensional language subspace estimated from cross￾language same-speaker differences reduces development EER from 2.97% with cosine and 2.70% with AS-Norm to 2.18%. This is consistent with earlier work on NAP and front-end factor analysis, where low-rank nuisance subspaces capture channel or language variation [6]. The Codabench evaluation scores con￾f… 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

26 extracted references · 2 linked inside Pith

  1. [1]

    In such scenarios, language-specific phonotactics and prosodic patterns act as nuisance factors that can dominate the representation space and degrade speaker discrimination

    Introduction Deep speaker verification systems based on x-vectors, ECAPA- TDNN, and attention-enhanced ResNets achieve very low error rates on benchmarks such as V oxCeleb when train and test con- ditions are matched [1, 2, 3], but performance remains frag- ile under cross-lingual mismatch, where enrollment and test utterances of the same speaker are prod...

    Pith/arXiv arXiv 2026

  2. [2]

    wavlm base plus

    Methodology 2.1. Task and data The TidyV oice2026 task [13] is text-independent speaker ver- ification under cross-lingual mismatch. The TidyV oice corpus contains 3,666 training and 808 development speakers across 40 languages, and an evaluation set with 2,200 speakers in 38 un- seen languages. Audio is sampled at 16kHz. The development set comprises 59,...

  3. [3]

    Development set performance Figure 1 summarizes results on the TidyV oice development set, which contains 12 million trials from 808 speakers and 40 lan- guages

    Results 3.1. Development set performance Figure 1 summarizes results on the TidyV oice development set, which contains 12 million trials from 808 speakers and 40 lan- guages. The best single system without explicit language com- pensation is the champion with AS-Norm at 2.70% EER and 0.64 minDCF. Applying NAP withk=30to champion embed- dings and re-runnin...

  4. [4]

    Projecting out a Figure 2:NAP sweep and DET curves for the champion encoder

    Discussion Our results show that a simple NAP-based back-end on top of a strong pre-trained SimAM ResNet34 baseline is highly effective in the TidyV oice2026 setting. Projecting out a Figure 2:NAP sweep and DET curves for the champion encoder. low-dimensional language subspace estimated from cross- language same-speaker differences reduces development EER...

  5. [5]

    Conclusion We presented our TidyV oice2026 submission on cross-lingual verification. Using a SimAM ResNet34 pretrained on V oxCeleb and V oxBlink2 and fine-tuned on TidyV oice [3, 1, 13, 20], we showed NAP-based language normalization with AS-Norm re- duces EER from 2.70 to 2.18 and yields a Codabench score of 8.40. More complex alternatives, including EC...

  6. [6]

    The LLMs were not involved in the development of ideas, methodology design, experimental procedures, data analysis, or interpretation of results

    Generative AI Use Disclosure LLMs were used exclusively for language editing, including rephrasing and grammatical refinement, to improve clarity and readability. The LLMs were not involved in the development of ideas, methodology design, experimental procedures, data analysis, or interpretation of results. All scientific content was developed and verifie...

  7. [7]

    X-vectors: Robust dnn embeddings for speaker recognition,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudan- pur, “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

  8. [8]

    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,”Interspeech 2020, 2020

    2020

  9. [9]

    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,” inICASSP 2022-2022 IEEE international conference on acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6722–6726

    2022

  10. [10]

    Spoken language mismatch in speaker verification: An investigation with nist-sre and crss bi- ling corpora,

    A. Misra and J. H. Hansen, “Spoken language mismatch in speaker verification: An investigation with nist-sre and crss bi- ling corpora,” in2014 IEEE spoken language technology work- shop (SLT). IEEE, 2014, pp. 372–377

    2014

  11. [11]

    Modelling and compensation for language mismatch in speaker verification,

    ——, “Modelling and compensation for language mismatch in speaker verification,”Speech Communication, vol. 96, pp. 58–66, 2018

    2018

  12. [12]

    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 Trans- actions on Audio, Speech, and Language Processing, vol. 19, no. 4, pp. 788–798, 2010

    2010

  13. [13]

    Semi-supervised nuisance-attribute networks for domain adaptation,

    W. Lin, M.-W. Mak, Y . Tu, and J.-T. Chien, “Semi-supervised nuisance-attribute networks for domain adaptation,” inICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 6236–6240

    2019

  14. [14]

    Super- vised domain adaptation for text-independent speaker verification using limited data,

    S. S. Sarfjoo, S. Madikeri, P. Motlicek, and S. Marcel, “Super- vised domain adaptation for text-independent speaker verification using limited data,” inProc. Interspeech 2020, 2020, pp. 3815– 3819

    2020

  15. [15]

    Cross-lingual text- independent speaker verification using unsupervised adversarial discriminative domain adaptation,

    W. Xia, J. Huang, and J. H. Hansen, “Cross-lingual text- independent speaker verification using unsupervised adversarial discriminative domain adaptation,” inICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP). IEEE, 2019, pp. 5816–5820

    2019

  16. [16]

    Speaker verification using end-to-end adversar- ial language adaptation,

    J. Rohdin, T. Stafylakis, A. Silnova, H. Zeinali, L. Burget, and O. Plchot, “Speaker verification using end-to-end adversar- ial language adaptation,” inICASSP 2019-2019 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2019, pp. 6006–6010

    2019

  17. [17]

    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. Xiaoet al., “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

  18. [18]

    Superb: Speech processing universal performance benchmark,

    S.-w. Yang, P.-H. Chi, Y .-S. Chuang, C.-I. J. Lai, K. Lakhotia, Y . Y . Lin, A. T. Liu, J. Shi, X. Chang, G.-T. Linet al., “Superb: Speech processing universal performance benchmark,” inProc. Interspeech 2021, 2021, pp. 1194–1198

    2021

  19. [19]

    Tidyvoice 2026 challenge evaluation plan,

    A. Farhadipour, J. Marquenie, S. Madikeri, T. Vukovic, V . Dellwo, K. Reid, F. M. Tyers, I. Siegert, and E. Chodroff, “Tidyvoice 2026 challenge evaluation plan,”arXiv preprint arXiv:2601.21960, 2026

    arXiv 2026

  20. [20]

    Analysis of score normalization in multilingual speaker recognition,

    P. Mat ˇejka, O. Novotn `y, O. Plchot, L. Burget, M. D. S ´anchez, and J. ˇCernock`y, “Analysis of score normalization in multilingual speaker recognition,” inProc. Interspeech 2017, 2017, pp. 1567– 1571

    2017

  21. [21]

    V oxblink2: A 100k+ speaker recognition corpus and the open- set speaker-identification benchmark,

    Y . Lin, M. Cheng, F. Zhang, Y . Gao, S. Zhang, and M. Li, “V oxblink2: A 100k+ speaker recognition corpus and the open- set speaker-identification benchmark,” inProc. Interspeech 2024, 2024, pp. 4263–4267

    2024

  22. [22]

    V oxceleb2: Deep speaker recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: Deep speaker recognition,”Interspeech 2018, 2018

    2018

  23. [23]

    Domain-adversarial training of neural networks,

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

    2016

  24. [24]

    Nuisance at- tribute projection,

    A. Solomonoff, W. M. Campbell, and C. Quillen, “Nuisance at- tribute projection,”Speech Communication, pp. 1–73, 2007

    2007

  25. [25]

    To train or not to train adversarially: A study of bias mitigation strategies for speaker recognition,

    R. Peri, K. Somandepalli, and S. Narayanan, “To train or not to train adversarially: A study of bias mitigation strategies for speaker recognition,”arXiv preprint arXiv:2203.09122, 2022

    Pith/arXiv arXiv 2022

  26. [26]

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

    A. Farhadipour, J. Marquenie, S. Madikeri, and E. Chodroff, “Tidyvoice: A curated multilingual dataset for speaker verification derived from common voice,”arXiv preprint arXiv:2601.16358, 2026

    arXiv 2026