Pith. sign in

REVIEW 2 major objections 4 minor 31 references

Analysis of ABC Frontend Audio Systems for the NIST-SRE24

T0 review · 2 major / 4 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 Pre-training a ResNet on VoxBlink2, resampled to 8 kHz with GSM augmentation, then fine-tuning it on the CTS Superset with long segments, produces the best SRE24 evaluation frontend tested in this paper, reaching Cprimary 0.562 and 7.59%…

desk verdict Useful telephony SV recipes on official NIST sets, but the VoxBlink2 attribution is confounded with pooling, recipe, and scoring changes and needs an ablation or softer claims. read the letter →

arxiv 2505.15320 v1 pith:DIIEUX47 submitted 2025-05-21 eess.AS cs.SD

Sara Barahona , Anna Silnova , Ladislav Mošner , Junyi Peng , Oldřich Plchot , Johan Rohdin , Lin Zhang , Jiangyu Han
show 9 more authors
Petr Palka Federico Landini Lukáš Burget Themos Stafylakis Sandro Cumani Dominik Boboš Miroslav Hlavaček Martin Kodovsky Tomáš Pavlíček
This is my paper · ORCID
classification eess.AScs.SD
keywords speakerverificationNISTSRE24embeddingextractorsVoxBlink2conversationaltelephonespeechAAM-Softmaxfine-tuningsegmentlengthResNet
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that the strongest speaker-embedding frontend for the SRE24 audio track comes not from the newest architectures but from pre-training a ResNet on a large out-of-domain corpus, VoxBlink2, adapted to the telephone channel by downsampling to 8 kHz and applying GSM codec to half of it, then fine-tuning on the conversational telephone speech (CTS) Superset. In the open condition this recipe yields the best evaluated system: a ResNet-152 with standard temporal statistics pooling reaches Cprimary 0.562 and 7.59% EER, compared with Cprimary 0.683 and 10.18% EER for the best fixed-condition system. The paper also establishes that lengthening the fine-tuning segments from 6 s up to 40 s steadily improves the evaluation metrics, with a 23.98% relative EER reduction on the evaluation set. A sympathetic reader would care because the result gives a practical, reproducible recipe for building strong telephone-speaker frontends from abundant wideband video data.

What carries the argument

The load-bearing mechanism is a two-stage training recipe: additive angular margin softmax (AAM-Softmax) pre-training on a domain-adapted external corpus, followed by large-margin fine-tuning on the target-domain CTS Superset. Domain adaptation of the pre-training data (8 kHz resampling plus GSM codec on half the data) is what aligns the YouTube-derived VoxBlink2 audio with telephone speech; the fine-tuning segment length acts as a second control, with longer segments (up to 40 s) producing monotonically better SRE24 results. The fixed-condition comparison is carried by xi-vector pooling, an uncertainty-aware replacement for temporal statistics pooling in ResNets, which gives the strongest fixed-condition numbers but is not what drives the leading open-condition result.

What would settle it

Train the exact ResNet-152-VB recipe (same augmentation, same AAM-Softmax stages, same 40 s fine-tuning) but initialize from random weights instead of VoxBlink2; if the SRE24 evaluation EER stays near 6.47%, VoxBlink2 pre-training is not the cause of the gain.

Watch

Extended reading notes

Core claim

The central claim is that a ResNet-152 trained first on the VoxBlink2 corpus (111,284 speakers from YouTube audio, downsampled to 8 kHz with GSM applied to 50% of segments) and then large-margin fine-tuned on the NIST CTS Superset with long segments is the best frontend in the ABC submission for SRE24. On the evaluation set this ResNet-152-VB model achieves Cprimary 0.562 and EER 7.59%, outperforming all fixed-condition systems, the ReDimNet-B3, the XLS-R foundation model, and even the larger XI-ResNet-221 used in the fixed track. The paper further reports a consistent improvement from increasing fine-tuning segment length to 6, 10, 20, and 40 seconds, culminating in a 23.98% relative EER reduction on the evaluation set and Cprimary 0.491. In contrast, the fixed-condition models using xi-vector pooling, an uncertainty-aware pooling approach, show better performance than ReDimNet and XLS-R in their setting, but still trail the VoxBlink2-initialized open-condition model.

Load-bearing premise

The paper's leading result is credited to VoxBlink2 pre-training, but because the winning system differs from the fixed-condition systems in architecture, pooling, and recipe, the improvement cannot be pinned on the dataset alone.

Editorial extensions

If this is right

  • If the recipe is right, open-condition speaker verification systems can exploit 100k+ speaker wideband video corpora for telephony tasks after lightweight channel simulation (8 kHz plus GSM), without needing matched telephone data at pretraining scale.
  • Fine-tuning segment length is a meaningful hyperparameter for evaluation conditions with variable enrollment durations and short test segments; longer segments improve both EER and Cprimary across a broad range of operating points.
  • For CTS-style tasks, ResNet backbones with standard pooling remain competitive with newer specialized architectures (ReDimNet) and large self-supervised models (XLS-R), which do not overtake them here.
  • Releasing the VoxBlink2-pretrained models would let other teams skip a costly pretraining stage and fine-tune directly on their target domain.

Reading between the lines

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

  • A natural control that would strengthen the central claim is a same-recipe ablation: ResNet-152 trained with the paper's augmentation and segment schedule but initialized randomly versus from VoxBlink2, measuring the SRE24 evaluation EER gap.
  • The monotone improvement from 6 to 40 s suggests testing even longer segments or a short-to-long curriculum could push the evaluation Cprimary below 0.491.
  • Because mixing original 16 kHz and upsampled 8 kHz audio preserves SRE24 accuracy while restoring VoxCeleb1 generalization, extending the same idea to other codecs and channel simulations may yield a single robust frontend for both telephony and wideband domains.
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.

Referee Report

2 major / 4 minor

Summary. The paper reports the ABC team's speaker-embedding frontends for the NIST SRE24 audio track, covering both the fixed condition (CTS Superset only) and the open condition (additional public data). It compares XI-ResNet variants, ReDimNet-B3, ResNet-152 pre-trained on VoxBlink2 (with 8 kHz/GSM domain adaptation), and an XLS-R-based system. The main empirical claims are: (i) in the fixed condition, larger XI-ResNets outperform ReDimNet; (ii) in the open condition, ResNet-152-VB gives the best results (Cprimary=0.562, EER=7.59% on the evaluation set); and (iii) increasing fine-tuning segment length on the CTS Superset from 6s to 40s yields a 23.98% relative EER reduction on the evaluation set. The paper also studies resampling strategies for VoxBlink2 and shows that mixing 16 kHz and upsampled 8 kHz data improves generalization to VoxCeleb1 while staying competitive on SRE24.

Significance. If the central claims hold, the paper provides practically useful recipes for using large-scale out-of-domain data (VoxBlink2, 111k speakers) for telephony-domain speaker verification, and a clear demonstration that long fine-tuning segments help under the SRE24 enrollment/test duration variability. Strengths include evaluation on official NIST development and evaluation sets, a consistent cosine-scoring/LDA preprocessing framework for the fixed-condition comparisons, and the stated intention to release pre-trained models on WeSpeaker, which would support reproducibility. However, the paper's headline attribution of the open-condition gain to VoxBlink2's speaker diversity is not backed by a controlled comparison, because the open system differs from the fixed systems in architecture, pooling, training recipe, and scoring pipeline simultaneously. This is the load-bearing issue for the paper's main analytical claim and requires additional experiments or a substantially revised interpretation.

major comments (2)
  1. [Section 3.2 and Table 1 (also Section 2.2)] The claim that VoxBlink2 pre-training has 'an enormous positive impact' and that the improvement is 'primarily attributed to the dataset's extensive speaker diversity' is not supported by the presented comparisons. The open-condition ResNet-152-VB differs from the fixed-condition XI-ResNet systems in at least three non-ablated respects: (1) pooling is standard TSTP rather than xi-vector; (2) the training recipe differs (e.g., 2s/10s segments and 150+10 epochs for ResNet-152-VB versus 3s segments and 130+5 epochs for XI-ResNet-152, with speed perturbation disabled in the latter); and (3) the scoring pipeline omits LDA for the VoxBlink2 models while fixed systems use LDA. Section 2.2 states that TSTP was employed 'to systematically isolate and evaluate the impact of incorporating the VoxBlink2 data corpus,' but no fixed-condition ResNet-152-TSTP baseline is reported. The authors should add a same-recipe ResNet-152-TSTP model trained on the CTS Superset only (with and without LDA), or otherwise vary one factor at a time, to separate the contribution of the VoxBlink2 data from the pooling, recipe, and scoring changes.
  2. [Section 3.4 and Table 3] The reported '23.98% reduction in EER' from extending the fine-tuning segment length lacks an explicit baseline in the text. From Table 3, the value corresponds to the comparison between the 6s row (eval EER 8.51%) and the 40s row (eval EER 6.47%). The paper should state this baseline explicitly and also clarify that the intermediate 10s row is the same model reported in Table 1. Without this clarification, the abstract's wording 'extending fine-tuning segment length to 40s gives a 23.98% relative EER reduction' could be misread as a reduction from the 10s system (which would be only 14.8% relative). This is a presentation issue in a headline number, but it affects the interpretability of the central segment-length result, so it should be fixed.
minor comments (4)
  1. [Table 3 caption] The caption's statement that 'VoxCeleb1 results are only shown for models trained during the first stage' is ambiguous: the SimAM rows are described in the text as including fine-tuned versions on the CTS Superset, and the ResNet152-VB rows also undergo CTS Superset fine-tuning. Please clarify exactly which training stage each VoxCeleb1 column corresponds to, and define the 'FT length' notation (e.g., '2 10s' for SimAM rows).
  2. [Section 2.2 / References] Reference [8] is listed as 'A. Undisclosed' with a withheld team name. If this citation is used to justify the claim that diverse frontends are complementary, the authors should either provide the full citation or replace it with a publicly available reference; a self-citation placeholder is not verifiable by the reader.
  3. [Section 3.2] The XLS-R discussion ends with the vague statement that 'further fine-tuning strategies should be studied.' Given that the paper reports only one fine-tuning configuration, it would be more informative to state what was tried (e.g., no VoxBlink2 pre-training, no different segment lengths, no additional data mixing) so the negative result is interpretable.
  4. [Throughout] There are several formatting inconsistencies: 'V oxBlink2' appears with a spurious space in the abstract and elsewhere, 'V AD' appears instead of 'VAD', and 'C primary' is sometimes typeset with a space. These should be cleaned up before final submission.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all central results are empirical measurements on the held-out NIST SRE24 evaluation set, and the only self-citation is a peripheral motivational reference.

full rationale

This paper is an experimental systems report rather than a mathematical derivation. The central claims, such as ResNet-152-VB achieving the best results on the SRE24 evaluation set, are empirical measurements on an external, held-out benchmark that was not used for training or parameter fitting. The VoxBlink2 pre-training and CTS Superset fine-tuning recipes are described explicitly, and performance is reported with cosine scoring on fixed evaluation trials; there is no fitted parameter that is later renamed as a prediction. The segment-length study (Section 3.4) similarly reports measured improvements on the evaluation set, not values obtained by construction. The only self-citation of note is reference [8], a withheld companion paper used in the introduction to motivate that diverse frontends are complementary for fusion. This is peripheral to the paper's frontend analysis and is not load-bearing for any reported result. The absence of a fixed-condition TSTP ResNet baseline is a real experimental confound when attributing gains to VoxBlink2, but confoundedness is not circularity: the results are still independent measurements against an external benchmark. No circular reduction, self-definitional step, or fitted-input-called-prediction pattern is present.

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

The paper is empirical; its central claims rest on experimental choices (data augmentation ratios, model selection, segment lengths) and on assumptions that the evaluation protocol and development set are reliable. No new theoretical entities are introduced.

free parameters (3)
  • VoxBlink2 GSM augmentation probability = 50% of segments
    The paper applies GSM codec to half of the downsampled VoxBlink2 data to simulate telephone channels; this ratio is chosen by hand and not swept.
  • ReDimNet configuration = B3
    Selected as best among B0, B2, B3, B6 based on SRE24 development set Cprimary/EER; a model selection choice that depends on the evaluation metric.
  • Fine-tuning segment length = 40 seconds
    The paper sweeps 6, 10, 20, 40 seconds and reports consistent improvement; the final model uses the longest tested length, making the result partly data-driven.
assumptions (3)
  • domain assumption The SRE24 development set is representative of the evaluation set distribution.
    Model and hyperparameter choices (e.g., ReDimNet-B3, segment length) are made based on dev set scores; the paper does not analyze dev-eval correlation or uncertainty.
  • domain assumption The WeSpeaker default training recipe (AAM-Softmax scale 32, margin schedules, learning rate) is near-optimal for the telephony domain.
    The paper adopts the VoxCeleb recipe without re-tuning for 8 kHz telephone data, assuming it transfers.
  • domain assumption The NIST-provided trial lists and scoring scripts are used without error.
    All results depend on the correctness of the official Cprimary and EER computation; the paper provides no verification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analysis of ABC Frontend Audio Systems for the NIST-SRE24." pith.science (2026). https://pith.science/paper/DIIEUX47

@misc{pith2026250515320,
  author       = {Pith},
  title        = {Pith review of: Analysis of ABC Frontend Audio Systems for the NIST-SRE24},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DIIEUX47}},
  note         = {Machine review of arXiv:2505.15320}
}
read the original abstract

We present a comprehensive analysis of the embedding extractors (frontends) developed by the ABC team for the audio track of NIST SRE 2024. We follow the two scenarios imposed by NIST: using only a provided set of telephone recordings for training (fixed) or adding publicly available data (open condition). Under these constraints, we develop the best possible speaker embedding extractors for the pre-dominant conversational telephone speech (CTS) domain. We explored architectures based on ResNet with different pooling mechanisms, recently introduced ReDimNet architecture, as well as a system based on the XLS-R model, which represents the family of large pre-trained self-supervised models. In open condition, we train on VoxBlink2 dataset, containing 110 thousand speakers across multiple languages. We observed a good performance and robustness of VoxBlink-trained models, and our experiments show practical recipes for developing state-of-the-art frontends for speaker recognition.

Figures

Figures reproduced from arXiv: 2505.15320 by the authors.

Figure 1
Figure 1. DCF plots for ResNet152-VB fine-tuned on segments of different lengths. 3.4. Impact of segment length in the fine-tuning process Given that one of the key challenges introduced in SRE24 was the presence of shorter test segments and variability in enroll￾ment durations, we investigated the impact of different segment lengths during the fine-tuning stage on the CTS Superset us￾ing our ResNet152-VB model. We systematic… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 19 canonical work pages

  1. [8]

    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,” inInterspeech 2024, 2024, pp. 4263–4267

    2024

  2. [1]

    Analysis of ABC Frontend Audio Systems for the NIST-SRE24

    Introduction The major body of speaker verification (SV) research focuses on 16 kHz datasets with audio excerpts extracted from (e.g., YouTube) video clips. The V oxCeleb datasets [1, 2] represented a great milestone in pushing the verification performance for- ward by providing training data comprising thousands of speak- ers. This initiative has been re...

    work page Pith review arXiv 2016

  3. [2]

    Training data and augmentations For the fixed condition, we used the NIST CTS Superset [4] to train the embedding extractors

    Proposed method 2.1. Training data and augmentations For the fixed condition, we used the NIST CTS Superset [4] to train the embedding extractors. To enhance the robustness of the model, we implemented acoustic data enhancement us- ing the Kaldi toolkit [15], incorporating noise from the MU- SAN database [16] and room impulse responses from the RIR databa...

    work page

  4. [3]

    Experimental Results Focusing on comparing and analyzing the effects of different embedding extractors, we employed cosine scoring as our clas- sifier, a natural choice given our optimization of the AAM loss. To isolate the effects of different frontend systems, we ap- plied a consistent preprocessing pipeline consisting of center- ing, dimensionality red...

    work page arXiv 2024

  5. [4]

    With the aim of building a strong frontend, we explored various architectures, pooling methods, and models pre-trained in a self-supervised way

    Conclusion This paper targeted speaker verification in a challenging mix- ture of telephony and audio-from-video speech. With the aim of building a strong frontend, we explored various architectures, pooling methods, and models pre-trained in a self-supervised way. What we consider the most prominent contribution is the analysis of using large-scale (not ...

    work page

  6. [5]

    Acknowledgements This work was partly supported by project PID2021- 125943OB-I00, MCIN/AEI/10.13039/501100011033/FEDER, UE from the Spanish Ministerio de Ciencia e Innovacion, Fondo Europeo de Desarrollo Regional

    doi:10.13039/501100011033/feder

  7. [6]

    V oxceleb: A large- scale speaker identification dataset,

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

    2017

  8. [7]

    V oxceleb2: Deep speaker recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxceleb2: Deep speaker recognition,”arXiv preprint arXiv:1806.05622, 2018

    arXiv 2018

Show all 31 references
  1. [9]

    Nist sre cts superset: A large-scale dataset for telephony speaker recognition,

    O. Sadjadi, “Nist sre cts superset: A large-scale dataset for telephony speaker recognition,” 2021-08-16 04:08:00 2021. [Online]. Available: https://tsapps.nist.gov/publication/get_pdf. cfm?pub_id=933116

    2021

  2. [10]

    The 2016 nist speaker recogni- tion evaluation,

    S. O. Sadjadi, T. Kheyrkhah, A. Tong, C. S. Green- berg, D. A. Reynolds, E. Singer, L. P. Mason, and J. Hernandez-Cordero, “The 2016 nist speaker recogni- tion evaluation,” inInterspeech, 2017. [Online]. Available: https://api.semanticscholar.org/CorpusID:263893047

    2016

  3. [11]

    The 2021 nist speaker recognition evaluation,

    S. O. Sadjadi, C. Greenberg, E. Singer, L. Mason, and D. Reynolds, “The 2021 nist speaker recognition evaluation,” inThe Speaker and Language Recognition Workshop (Odyssey 2022), 2022, pp. 322–329

    2021

  4. [12]

    Audio-visual person recognition in multimedia data from the iarpa janus program,

    G. Sell, K. Duh, D. Snyder, D. Etter, and D. Garcia-Romero, “Audio-visual person recognition in multimedia data from the iarpa janus program,” in2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018, pp. 3031–3035

    2018

  5. [13]

    Analysis of the <withheld team name> classifi- cation backends for nist sre24,

    A. Undisclosed, “Analysis of the <withheld team name> classifi- cation backends for nist sre24,” inSubmitted to Interspeech 2025, 2025

    2025

  6. [14]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  7. [15]

    The V oxCeleb Speaker Recognition Challenge: A Retrospective,

    J. Huh, J. S. Chung, A. Nagrani, A. Brown, J.-w. Jung, D. Garcia- Romero, and A. Zisserman, “The V oxCeleb Speaker Recognition Challenge: A Retrospective,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, p. 3850–3866,

  8. [16]

    MUSAN: A Music, Speech, and Noise Corpus,

    D. Snyder, G. Chen, and D. Povey, “MUSAN: A Music, Speech, and Noise Corpus,” 2015, arXiv:1510.08484v1

    2015 arXiv

  9. [17]

    The 2021 NIST Speaker Recognition Evaluation,

    S. O. Sadjadi, C. Greenberg, E. Singer, L. Mason, and D. Reynolds, “The 2021 NIST Speaker Recognition Evaluation,” inThe Speaker and Language Recognition Workshop (Odyssey 2022), 2022, pp. 322–329

    2021

  10. [18]

    Reshape Dimensions Network for Speaker Recognition,

    I. Yakovlev, R. Makarov, A. Balykin, P. Malov, A. Okhotnikov, and N. Torgashov, “Reshape Dimensions Network for Speaker Recognition,” inInterspeech 2024, 2024, pp. 3235–3239

    2024

  11. [19]

    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

  12. [20]

    An Attention-Based Backend Allowing Efficient Fine- Tuning of Transformer Models for Speaker Verification,

    J. Peng, O. Plchot, T. Stafylakis, L. Mošner, L. Burget, and J. ˇCer- nocký, “An Attention-Based Backend Allowing Efficient Fine- Tuning of Transformer Models for Speaker Verification,” in2022 IEEE Spoken Language Technology Workshop (SLT), 2023, pp. 555–562

    2023

  13. [21]

    The kaldi speech recognition toolkit,

    D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek, Y . Qian, P. Schwarzet al., “The kaldi speech recognition toolkit,” inIEEE 2011 workshop on automatic speech recognition and understanding. IEEE Signal Processing Society, 2011

    2011

  14. [22]

    NIST 2024 Speaker Recognition Evaluation Plan,

    “NIST 2024 Speaker Recognition Evaluation Plan,” Online,

    2024

  15. [23]

    A Study on Data Augmentation of Reverberant Speech for Ro- bust Speech Recognition,

    T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A Study on Data Augmentation of Reverberant Speech for Ro- bust Speech Recognition,” in2017 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 5220–5224

    2017

  16. [24]

    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,” inIEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  17. [25]

    Advancing speaker embedding learning: Wespeaker toolkit for research and produc- tion,

    S. Wang, Z. Chen, B. Han, H. Wang, C. Liang, B. Zhang, X. Xi- ang, W. Ding, J. Rohdin, A. Silnovaet al., “Advancing speaker embedding learning: Wespeaker toolkit for research and produc- tion,”Speech Communication, vol. 162, p. 103104, 2024

    2024

  18. [26]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2019, pp. 4690–4699

    2019

  19. [27]

    XLS-R: Self-supervised Cross-lingual Speech Rep- resentation Learning at Scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Conneau, and M. Auli, “XLS-R: Self-supervised Cross-lingual Speech Rep- resentation Learning at Scale,” inInterspeech 2022, 2022, pp. 2278–2282

    2022

  20. [29]

    Available: https://www.nist.gov/itl/iad/mig/ nist-2024-speaker-recognition-evaluation-sre24

    [Online]. Available: https://www.nist.gov/itl/iad/mig/ nist-2024-speaker-recognition-evaluation-sre24

    2024

  21. [30]

    Simam: A sim- ple, parameter-free attention module for convolutional neural net- works,

    L. Yang, R.-Y . Zhang, L. Li, and X. Xie, “Simam: A sim- ple, parameter-free attention module for convolutional neural net- works,” inProceedings of the 38th International Conference on Machine Learning, ser. Proceedings of Machine Learning Re- search, M. Meila and T. Zhang, E...

    2021

  22. [31]

    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

  23. [2024]

    Available: http://dx.doi.org/10.1109/TASLP

    [Online]. Available: http://dx.doi.org/10.1109/TASLP. 2024.3444456

    2024

Pith tools

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