On the influence of language similarity in non-target speaker verification trials

Michael Jessen; Paul M. Reuter

REVIEW 3 major objections 5 minor 32 references

On the influence of language similarity in non-target speaker verification trials

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

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

Pith's one-line read Speaker scores track language similarity for untrained languages, while trained-language comparisons cluster tightly.

desk verdict A useful empirical study with a genuinely new operationalization, but the headline effects need uncertainty bounds before I'd trust the numbers. read the letter →

arxiv 2506.02777 v1 pith:4JRHNRUB submitted 2025-06-03 eess.AS cs.SD

Paul M. Reuter , Michael Jessen This is my paper

classification eess.AScs.SD

keywords speakerverificationcross-lingualtrialslanguagesimilarityscoredistributionsnon-targetforensicrecognitionECAPA-TDNNVoxCeleb

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper aims to show that in non-target (different-speaker) speaker verification trials, the effect of language on scores depends on whether a language was seen in training. When one side of the comparison is a training language, the scores cluster tightly and barely move with the choice of comparison language; when both sides are untrained languages, speaker-similarity scores rise with language similarity, and this correlation is stronger for the multilingual system than for the English-only system. The authors argue this matters for forensic speaker recognition, where score normalisation and calibration sets often do not match the language of the case recording.

What carries the argument

The central machinery is the ECAPA-TDNN speaker embedding model trained on multilingual VoxCeleb1+2 and English-only VoxCeleb2-en data, with embeddings compared by cosine similarity. Language similarity is operationalized by a ResNet34 classifier trained on VoxLingua107, where the cosine similarity between the last-layer embeddings of two utterances is the language-similarity score. The trained-versus-untrained split, English, German and French as trained for Vox1+2 and English only for Vox2-en, is the categorization that separates the clustering effect from the language similarity effect. t-SNE plots of averaged embeddings are used to visualize how speakers of training languages form their own clusters.

What would settle it

Retrain or reconfigure the systems with a stated training-share threshold, for example counting Hindi, Italian, Spanish and Dutch as trained languages, and check whether the lower standard deviations for trained-vs-untrained comparisons and the higher correlations for untrained-vs-untrained comparisons in Tables 2 and 3 persist or disappear.

Watch

Extended reading notes

Core claim

The paper's central claim is a two-part pattern in cross-lingual speaker verification score distributions. First, a clustering effect: for different-speaker comparisons that involve a language present in the system's training data, the comparison language has only a small influence on cosine similarity scores, producing tightly grouped distributions. Second, a language similarity effect: for comparisons between two languages the system was not trained on, speaker similarity scores correlate with language similarity as measured by a separate language classifier, with the effect more pronounced for the multilingual VoxCeleb1+2 system than for the English-only VoxCeleb2 system. The same-language trials consistently produce the highest scores, and comparisons with a training language produce the lowest scores.

Load-bearing premise

The results depend on an arbitrary binary split of languages into trained and untrained, with Spanish, Italian, Dutch and Hindi present in the VoxCeleb training data at 1.6 to 2.7 percent but treated as untrained, so the clustering and correlation contrasts would change if that threshold moved.

Editorial extensions

If this is right

Score normalization and calibration sets in forensic casework can be expected to behave differently for trained versus untrained reference languages, so language-matched norm sets matter most when the test language is untrained.
A multilingual system's scores carry more language information, meaning cross-lingual different-speaker trials with untrained languages will show predictable language-distance-driven score shifts.
Same-language different-speaker trials can be expected to score higher than cross-lingual trials regardless of training composition.
Comparisons involving a training language give compressed score distributions, so choosing among several comparison languages will barely change the score distribution.
Language similarity measured by a separate classifier is a usable predictor for score shifts in untrained-language comparisons.

Reading between the lines

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

Beyond the paper: if the language-similarity effect is causal, score normalization could be adapted by subtracting a language-distance-dependent offset for untrained-language trials, not just by choosing an equilingual cohort.
Beyond the paper: the binary trained/untrained split is likely a graded quantity, so a system trained with intermediate language proportions may show a smooth transition between clustering and correlation rather than a sharp switch.
Beyond the paper: the same mechanism predicts that adding a new language to training data would remove that language's similarity effect and pull it into the clustering pattern, a testable prediction with existing fine-tuning recipes.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague, here's my read of arXiv:2506.02777.

The genuinely new thing here is the operationalization: language similarity is measured continuously via embeddings from a separate language classifier, then related to non-target speaker-verification score distributions, with the analysis split by whether the reference language appeared in the training data. That is a real step beyond the usual matched/mismatched language dichotomy, and the paper deserves credit for it. The main qualitative patterns — a clustering effect for well-represented training languages and a language-similarity effect for untrained languages, stronger with multilingual training — are consistent across Globalphone and LDC and line up with the t-SNE visualization. The forensic motivation is concrete: score shifts matter for normalization and calibration set selection.

The soft spots are real and mostly in the quantitative section. Tables 2 and 3 report standard deviations and Pearson correlations with no confidence intervals, bootstrap, or permutation tests. The trials are pairwise across speakers, so they are strongly non-independent; with only 20–40 speakers per language, the effective sample size is far smaller than the number of trials. The gap between trained and untrained languages could be sampling noise, especially in LDC where English also shows a fairly high correlation. The binary trained/untrained split is also arbitrary: Spanish, Italian, Dutch, and Hindi appear at 1.5–2.7% in VoxCeleb but are binned as untrained, so the quantitative separation depends on that threshold. A continuous training-proportion variable or sensitivity analysis would help. LDC's telephone data may also confound language with channel, though the fact that the pattern appears in Globalphone mitigates that.

I don't see circularity: speaker and language similarity come from independently trained systems, so the reported correlations are not definitional. The citation pattern looks appropriate, and the writing is straightforward.

Bottom line: this is a useful empirical study for forensic speaker recognition and for the broader AI-bias discussion, but the central quantitative claims are not yet established. I'd send it to peer review rather than desk reject, and ask the authors to add speaker-level bootstrap/permutation intervals, a more defensible treatment of training proportion, and ideally release code or at least detailed per-language trial counts. With those additions it would be a solid contribution.

Referee Report

3 major / 5 minor

Summary. The paper investigates how language similarity affects non-target (different-speaker) speaker verification scores in cross-lingual trials. Using an ECAPA-TDNN speaker embedding system trained on VoxCeleb1+2 (multilingual) and on English-only VoxCeleb2, and a ResNet34 language classifier trained on VoxLingua107 to measure language similarity, the authors analyze score distributions on Globalphone and LDC Multilingual CTS. They report two main effects: a 'clustering effect', where cross-lingual comparisons involving a training language produce scores that vary little with the choice of comparison language (Table 2), and a 'language similarity effect', where cross-lingual comparisons between untrained languages produce speaker similarity scores that correlate positively with language similarity (Table 3), especially for the multilingual system. The paper argues these findings are relevant to forensic speaker recognition, particularly score normalization and calibration.

Significance. If the reported effects are real, they add a useful nuance to the literature on language mismatch in speaker verification: the influence of the comparison language depends on whether a language was present in the system's training data. The study uses a standard state-of-the-art architecture, two complementary multilingual corpora, and a language-similarity measure derived from an independently trained classifier, so the speaker-language correlation is not definitionally circular. The paper also addresses an underexplored outcome measure (score distributions rather than EER/Cllr alone), which is directly relevant to forensic Tippett-style reporting. However, the quantitative foundations of the two central claims—the clustering and language-similarity effects—are weakened by the absence of uncertainty quantification and by an ad hoc binary split of languages into trained/untrained. These issues are addressable but require additional analysis.

major comments (3)

[§3.2, Table 2] The clustering effect is supported only by point estimates of standard deviations across languages. With 40 speakers per language in Globalphone and 20 in LDC, and with pairwise different-speaker trials that are non-independent (each speaker contributes multiple files and appears in many trials), the differences between languages (e.g., English 0.086 vs. Swedish 0.099 for Vox1+2) may not be statistically reliable. The authors should provide bootstrap confidence intervals, permutation tests, or a mixed-effects model that accounts for speaker-level clustering. Without this, the claim that 'comparisons involving a training language... have a smaller score variation' is not quantitatively established.
[§3.2, Table 3] The language-similarity effect rests on Pearson correlations between speaker and language similarity scores. These correlations are computed over non-independent trials, and no confidence intervals, significance tests, or corrections for multiple comparisons are reported. The numerical differences that drive the conclusion—e.g., untrained languages showing higher correlations (Mandarin r=0.53 in Globalphone) than trained languages (English r=0.08, German r=0.05)—could shift considerably under speaker resampling. The authors should report CIs (e.g., bootstrap by speaker) and, ideally, a test for the difference between correlation coefficients across systems or language groups.
[§3.2, Section 3.2 and Fig. 1] The binary split into 'trained' (English, German, French for Vox1+2; English for Vox2-en) and 'untrained' (all others) is based on an implicit threshold on the VoxCeleb language distribution. Spanish, Italian, Dutch, and Hindi are present in the training data at 1.5–2.7% but are treated as untrained, while German and French at 5–7% are treated as trained. The quantitative separation in Tables 2 and 3 depends on this arbitrary categorization. The authors should either justify the threshold, treat the training proportion as a continuous variable, or at minimum perform a sensitivity analysis excluding/adjusting borderline languages.

minor comments (5)

[Table 3 title] The table header reads 'systems Vox1+2 and Vox-en'; the second system is elsewhere called 'Vox2-en'. Please make the abbreviation consistent.
[Table 3, Bengali row] The Bengali row in the Globalphone columns shows a stray slash before the LDC values; this formatting error should be corrected.
[Section 3.1, Fig. 2] The caption for Fig. 2 reports 'German (males)' and 'Russian (females)', but Table 1 shows female averages; please clarify whether the qualitative patterns were checked for both sexes and whether the sex shown is representative.
[Section 3.1, Fig. 4] The t-SNE interpretation is presented without quantitative support. It would be useful to add a clustering metric or at least state that the visualization is illustrative only, since t-SNE can create apparent clusters even in random data.
[Section 2.3.1] The sentence 'Since English is the main language in VoxCeleb but not contained in Globalphone, we included 40 random male and female speakers from Librispeech' implies 40 male and 40 female, but earlier the text says '40 male and/or female speakers were randomly selected'; please be consistent about the exact numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the paper's correlations are empirical associations between outputs of independently trained systems.

full rationale

The paper's central results are empirical correlations between two sets of embeddings. Speaker similarity scores come from an ECAPA-TDNN trained on VoxCeleb1/2, while language similarity scores come from a ResNet34 trained on VoxLingua107; these are independent systems with different architectures, training objectives, and training corpora, so the Pearson correlations in Table 3 are not identities or fits. The trained/untrained language split (Section 3.2) is defined by the estimated language distribution of the VoxCeleb training data (Fig. 1) and not by the resulting score standard deviations or correlations, so the clustering effect and language-similarity effect are empirical outcomes rather than consequences of the categorization. References [9]-[11], which include one of the present authors, are cited only as background for the forensic importance of score distributions and calibration; they do not provide a load-bearing premise or an imported uniqueness theorem. No equation in the paper reduces a predicted quantity to an input by construction.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new numeric parameters or entities. The main implicit postulates are the validity of the language-similarity metric and the binary trained/untrained language split.

assumptions (4)

domain assumption Cosine similarity of ECAPA-TDNN embeddings is a valid speaker-similarity score for non-target trials.
Standard practice in speaker verification, assumed without justification here.
domain assumption Cosine similarity of ResNet34 language-classifier embeddings measures language similarity relevant to speaker verification.
The paper operationalizes language similarity this way, citing VoxLingua107 training, but provides no external validation against other language-distance measures.
ad hoc to paper Only English, German and French count as trained languages for the Vox1+2 system; other VoxCeleb languages are untrained.
The threshold is implicit in the VoxCeleb distribution (Fig. 1); Spanish, Italian, Dutch, Hindi are present at 1.6-2.7 percent but treated as untrained.
standard math Pearson correlation appropriately captures the association between speaker and language similarity.
Used without checking normality or monotonicity; acceptable as a descriptive tool.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On the influence of language similarity in non-target speaker verification trials." pith.science (2026). https://pith.science/paper/4JRHNRUB

@misc{pith2026250602777,
  author       = {Pith},
  title        = {Pith review of: On the influence of language similarity in non-target speaker verification trials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4JRHNRUB}},
  note         = {Machine review of arXiv:2506.02777}
}

read the original abstract

In this paper, we investigate the influence of language similarity in cross-lingual non-target speaker verification trials using a state-of-the-art speaker verification system, ECAPA-TDNN, trained on multilingual and monolingual variants of the VoxCeleb dataset. Our analysis of the score distribution patterns on multilingual Globalphone and LDC CTS reveals a clustering effect in speaker comparisons involving a training language, whereby the choice of comparison language only minimally impacts scores. Conversely, we observe a language similarity effect in trials involving languages not included in the training set of the speaker verification system, with scores correlating with language similarity measured by a language classification system, especially when using multilingual training data.

Figures

Figures reproduced from arXiv: 2506.02777 by the authors.

**Figure 1.** Language distribution in VoxCeleb1 and 2 2.2. Speaker verification ECAPA-TDNN [17] was adopted as the speaker verification model architecture as it is commonly used in state-of-the-art speaker verification [18, 19, 20]. It is built upon the original xvector architecture [21] and employs several enhancements regarding channel attention, propagation and aggregation. In our experiments, we used two models of the same… view at source ↗

**Figure 2.** shows the score distributions of different-speaker trials in Globalphone involving German and Russian, respectively, for the system trained on VoxCeleb1+2. For both reference languages, speaker similarity scores are highest on average for the same-language case (German-German, Russian-Russian) (H1 fulfilled). However, there are noticeable differences in the distributions of different-language scores. While scores … view at source ↗

**Figure 3.** Score distributions of different-speaker trials involving a) German (males) and b) Russian (females) in Globalphone set with English-only VoxCeleb2 training To further understand the two different score distribution patterns in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Language and speaker similarity score distribution contours of male different-speaker cross-lingual trials involving Punjabi in LDC set with VoxCeleb1+2 training Secondly, we performed a correlation analysis between speaker similarity and language similarity. For every…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 30 canonical work pages

[1]

A common research paradigm to address this question has been to study mismatch between the language spoken by the questioned speaker and known speaker [1, 2, 3, 4]

Introduction There is continued interest over the years in how and to what extent automatic speaker recognition is affected by language. A common research paradigm to address this question has been to study mismatch between the language spoken by the questioned speaker and known speaker [1, 2, 3, 4]. Some researchers ap- proached the language effect by st...

work page
[2]

On the influence of language similarity in non-target speaker verification trials

Methods 2.1. Language classification For language classification, a ResNet34 [13] was trained on V oxLingua107 [14] using cross-entropy loss. The model was used to analyze the distribution of languages spoken in the V ox- Celeb1 and 2 datasets [15, 16] which served as training data for the speaker verification models (see Fig. 1). 66.3% 13.6% 7.3% 5.4% 2....

work page Pith review arXiv 2025
[3]

Qualitative analysis Fig

Results 3.1. Qualitative analysis Fig. 2 shows the score distributions of different-speaker trials in Globalphone involving German and Russian, respectively, for the system trained on V oxCeleb1+2. For both reference lan- guages, speaker similarity scores are highest on average for the same-language case (German-German, Russian-Russian) (H1 fulfilled). Ho...

work page
[4]

Re- sults show that in comparisons involving a training language the choice of comparison language has only a small effect on the generated scores (clustering effect)

Conclusion In this paper, we focussed particularly on the influence of lan- guage similarity in non-target cross-lingual speaker verification trials for a state-of-the-art speaker verification system. Re- sults show that in comparisons involving a training language the choice of comparison language has only a small effect on the generated scores (clusteri...

work page
[5]

Acknowledgements This work was partially funded by the Federal Ministry of Ed- ucation and Research of Germany (BMBF) in the VIKING project (13N16239)

work page
[6]

Fusion of heterogeneous speaker recognition sys- tems in the stbu submission for the nist speaker recognition evalu- ation 2006,

N. Brummer, L. Burget, J. Cernocky, O. Glembek, F. Grezl, M. Karafiat, D. A. van Leeuwen, P. Matejka, P. Schwarz, and A. Strasheim, “Fusion of heterogeneous speaker recognition sys- tems in the stbu submission for the nist speaker recognition evalu- ation 2006,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 15, no. 7, pp. 2072–2084, 2007

work page 2006
[7]

Modelling and compensation for lan- guage mismatch in speaker verification,

A. Misra and J. H. Hansen, “Modelling and compensation for lan- guage mismatch in speaker verification,”Speech Communication, vol. 96, pp. 58–66, 2018

work page 2018
[8]

The impact of mismatched recordings on an automatic-speaker-recognition system and human listeners,

T. Nechansk ´y, T. Boˇril, A. Houzar, and R. Skarnitzl, “The impact of mismatched recordings on an automatic-speaker-recognition system and human listeners,” Acta Universitatis Carolinae Philo- logica 1 / Phonetica Pragensia, pp. 11–22, 2022

work page 2022

Show all 32 references

[9]

Tackling the score shift in cross-lingual speaker verification by exploiting lan- guage information,

J. Thienpondt, B. Desplanques, and K. Demuynck, “Tackling the score shift in cross-lingual speaker verification by exploiting lan- guage information,” in IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP) , 2022, pp. 7187– 7191

2022
[10]

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,” in Interspeech, 2017, pp. 1567–1571

2017
[11]

An investiga- tion of domain adaptation in speaker embedding space for speaker recognition,

F. Bahmaninezhad, C. Zhang, and J. H. Hansen, “An investiga- tion of domain adaptation in speaker embedding space for speaker recognition,” Speech Communication, vol. 129, pp. 7–16, 2021

2021
[12]

Effects of language mismatch in au- tomatic forensic voice comparison using deep learning embed- dings,

D. Sztah ´o and A. Fejes, “Effects of language mismatch in au- tomatic forensic voice comparison using deep learning embed- dings,” Journal of Forensic Sciences, vol. 68, no. 3, pp. 871–883, 2023

2023
[13]

Tuning the perfor- mance of automatic speaker recognition in different conditions: effects of language and simulated voice disguise,

R. Skarnitzl, M. Asiaee, and M. Nourbakhsh, “Tuning the perfor- mance of automatic speaker recognition in different conditions: effects of language and simulated voice disguise,” International Journal of Speech, Language and the Law , vol. 26, no. 2, p. 209–229, 2020

2020
[14]

Experiments with two forensic automatic speaker comparison systems using refer- ence populations that (mis)match the test language,

D. van der Vloed, M. Jessen, and S. Gfroerer, “Experiments with two forensic automatic speaker comparison systems using refer- ence populations that (mis)match the test language,” inAES Inter- national Conference on Audio Forensics, 2017

2017
[15]

Drygajlo, M

A. Drygajlo, M. Jessen, S. Gfroerer, I. Wagner, J. Vermeulen, and T. Niemi, Methodological Guidelines for Best Practice in Forensic Semiautomatic and Automatic Speaker Recognition . Frankfurt: Verlag f ¨ur Polizeiwissenschaft, 2015. [Online]. Avail- able: https://enfsi.eu/wp-c...

2015
[16]

Con- sensus on validation of forensic voice comparison,

G. S. Morrison, E. Enzinger, V . Hughes, M. Jessen, D. Meuwly, C. Neumann, S. Planting, W. C. Thompson, D. van der Vloed, R. J. Ypma, C. Zhang, A. Anonymous, and B. Anonymous, “Con- sensus on validation of forensic voice comparison,” Science & Justice, vol. 61, no. 3, pp. 299–...

2021
[17]

Multi-laboratory evaluation of forensic voice comparison systems under conditions reflect- ing those of a real forensic case (forensic eval 01) – conclusion,

G. S. Morrison and E. Enzinger, “Multi-laboratory evaluation of forensic voice comparison systems under conditions reflect- ing those of a real forensic case (forensic eval 01) – conclusion,” Speech Communication, vol. 112, pp. 37–39, 2019

2019
[18]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

2016
[19]

V oxlingua107: A dataset for spoken lan- guage recognition,

J. Valk and T. Alum ¨ae, “V oxlingua107: A dataset for spoken lan- guage recognition,” in IEEE Spoken Language Technology Work- shop (SLT), 2021, pp. 652–658

2021
[20]

V oxceleb: A large- scale speaker identification dataset,

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

2017
[21]

V oxceleb2: Deep speaker recognition,

J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: Deep speaker recognition,” in Interspeech, 2018, pp. 1086–1090

2018
[22]

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,” in Interspeech, 2020, pp. 3830– 3834

2020
[23]

Integrating frequency translational invariance in tdnns and frequency posi- tional information in 2d resnets to enhance speaker verification,

J. Thienpondt, B. Desplanques, and K. Demuynck, “Integrating frequency translational invariance in tdnns and frequency posi- tional information in 2d resnets to enhance speaker verification,” in Interspeech, 2021, pp. 2302–2306

2021
[24]

Pcf: Ecapa-tdnn with progressive channel fusion for speaker verification,

Z. Zhao, Z. Li, W. Wang, and P. Zhang, “Pcf: Ecapa-tdnn with progressive channel fusion for speaker verification,” in IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP), 2023, pp. 1–5

2023
[25]

Build a sre challenge system: Lessons from voxsrc 2022 and cnsrc 2022,

Z. Chen, B. Han, X. Xiang, H. Huang, B. Liu, and Y . Qian, “Build a sre challenge system: Lessons from voxsrc 2022 and cnsrc 2022,” in Interspeech, 2023, pp. 3202–3206

2022
[26]

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

2018
[27]

Speechbrain: A general-purpose speech toolkit,

M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell, L. Lugosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong, J.-C. Chou, S.-L. Yeh, S.-W. Fu, C.-F. Liao, E. Rastorgueva, F. Grondin, W. Aris, H. Na, Y . Gao, R. D. Mori, and Y . Bengio, “Speechbrain: A general-purp...

2021
[28]

Arcface: Additive an- gular margin loss for deep face recognition,

J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive an- gular margin loss for deep face recognition,” in IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) , 2019, pp. 4685–4694

2019
[29]

Globalphone: a multilingual speech and text database developed at karlsruhe university,

T. Schultz, “Globalphone: a multilingual speech and text database developed at karlsruhe university,” in 7th International Confer- ence on Spoken Language Processing (ICSLP 2002) , 2002, pp. 345–348

2002
[30]

Lib- rispeech: An asr corpus based on public domain audio books,

V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- rispeech: An asr corpus based on public domain audio books,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

2015
[31]

New resources for recognition of confusable linguistic varieties: the LRE11 corpus,

S. Strassel, K. Walker, K. Jones, D. Graff, and C. Cieri, “New resources for recognition of confusable linguistic varieties: the LRE11 corpus,” inThe Speaker and Language Recognition Work- shop (Odyssey 2012), 2012, pp. 202–208

2012
[32]

Visualizing data using t-sne,

L. van der Maaten and G. Hinton, “Visualizing data using t-sne,” Journal of Machine Learning Research, vol. 9, no. 86, pp. 2579– 2605, 2008

2008

Pith tools

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

[1] [1]

A common research paradigm to address this question has been to study mismatch between the language spoken by the questioned speaker and known speaker [1, 2, 3, 4]

Introduction There is continued interest over the years in how and to what extent automatic speaker recognition is affected by language. A common research paradigm to address this question has been to study mismatch between the language spoken by the questioned speaker and known speaker [1, 2, 3, 4]. Some researchers ap- proached the language effect by st...

work page

[2] [2]

On the influence of language similarity in non-target speaker verification trials

Methods 2.1. Language classification For language classification, a ResNet34 [13] was trained on V oxLingua107 [14] using cross-entropy loss. The model was used to analyze the distribution of languages spoken in the V ox- Celeb1 and 2 datasets [15, 16] which served as training data for the speaker verification models (see Fig. 1). 66.3% 13.6% 7.3% 5.4% 2....

work page Pith review arXiv 2025

[3] [3]

Qualitative analysis Fig

Results 3.1. Qualitative analysis Fig. 2 shows the score distributions of different-speaker trials in Globalphone involving German and Russian, respectively, for the system trained on V oxCeleb1+2. For both reference lan- guages, speaker similarity scores are highest on average for the same-language case (German-German, Russian-Russian) (H1 fulfilled). Ho...

work page

[4] [4]

Re- sults show that in comparisons involving a training language the choice of comparison language has only a small effect on the generated scores (clustering effect)

Conclusion In this paper, we focussed particularly on the influence of lan- guage similarity in non-target cross-lingual speaker verification trials for a state-of-the-art speaker verification system. Re- sults show that in comparisons involving a training language the choice of comparison language has only a small effect on the generated scores (clusteri...

work page

[5] [5]

Acknowledgements This work was partially funded by the Federal Ministry of Ed- ucation and Research of Germany (BMBF) in the VIKING project (13N16239)

work page

[6] [6]

Fusion of heterogeneous speaker recognition sys- tems in the stbu submission for the nist speaker recognition evalu- ation 2006,

N. Brummer, L. Burget, J. Cernocky, O. Glembek, F. Grezl, M. Karafiat, D. A. van Leeuwen, P. Matejka, P. Schwarz, and A. Strasheim, “Fusion of heterogeneous speaker recognition sys- tems in the stbu submission for the nist speaker recognition evalu- ation 2006,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 15, no. 7, pp. 2072–2084, 2007

work page 2006

[7] [7]

Modelling and compensation for lan- guage mismatch in speaker verification,

A. Misra and J. H. Hansen, “Modelling and compensation for lan- guage mismatch in speaker verification,”Speech Communication, vol. 96, pp. 58–66, 2018

work page 2018

[8] [8]

The impact of mismatched recordings on an automatic-speaker-recognition system and human listeners,

T. Nechansk ´y, T. Boˇril, A. Houzar, and R. Skarnitzl, “The impact of mismatched recordings on an automatic-speaker-recognition system and human listeners,” Acta Universitatis Carolinae Philo- logica 1 / Phonetica Pragensia, pp. 11–22, 2022

work page 2022