Pith. sign in

REVIEW 3 major objections 5 minor 30 references

Effective Modeling of Critical Contextual Information for TDNN-based Speaker Verification

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read By making ECAPA-TDNN's Res2 blocks bidirectional and replacing their convolutions with Bi-LSTM, this paper cuts VoxCeleb1-O equal error rate to 0.67% from ECAPA-TDNN's 0.87% with only about one million extra parameters.

desk verdict A useful ECAPA-TDNN variant with consistent EER gains, but the headline 23% rests on quoted baselines that may not be comparable. read the letter →

arxiv 2509.09932 v1 pith:674MQLX6 submitted 2025-09-12 eess.AS

Shilong Weng , Liu Yang , Ji Mao This is my paper · ORCID
classification eess.AS
keywords speakerverificationECAPA-TDNNRes2NetBi-LSTMcontextualinformationbidirectionalmodelingequalerrorrateTDNN
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 ECAPA-TDNN's SE-Res2Block only looks backward in its multi-scale feature hierarchy, so it ignores future context and fails to model long-term dependencies effectively. To fix this, the authors propose three modified blocks: a bidirectional Res2Block, a dual-stream SE-Res2Block, and a variant that swaps the central convolutions for Bi-LSTM. On VoxCeleb1, the best architecture (SE-Res2Bi-LSTM-ECAPA) achieves a 23% relative EER improvement over ECAPA-TDNN on VoxCeleb1-O, with only about one million additional parameters. The same architecture also beats five recent TDNN-based speaker verification systems listed in the comparison table. A sympathetic reader would care because this is a practical, low-cost way to push TDNN speaker verification accuracy higher.

What carries the argument

The key machinery is the SE-Res2Bi-LSTM block, which replaces the central 1D convolutional layers in ECAPA-TDNN's Res2Net hierarchy with Bi-LSTM modules. Each feature-map subset is processed by a Bi-LSTM that reads both forward and backward along the sequence, so every subset incorporates context from both sides and can learn which parts of the context are relevant. The other two proposed blocks—SE-Bi-Res2Block (a reversed Res2Block added to the original) and Bi-SE-Res2Block (two SE-Res2Blocks on the original and channel-flipped maps)—are simpler bidirectional variants. Together they serve as drop-in replacements for SE-Res2Block, demonstrating that bidirectional multi-scale context is the a

What would settle it

Retrain ECAPA-TDNN and the five comparison systems under exactly the same pipeline used for the proposed models (same 2-second audio length, 80-dimensional Mel features, spec-augment settings, AAM-Softmax margin and scale, Adam optimizer schedule, and AS-norm) and then compare VoxCeleb1-O EER. If ECAPA-TDNN's EER drops to near 0.67% or the gap shrinks dramatically, the claimed benefit of the new blocks would be largely an artifact of recipe differences.

Watch

Extended reading notes

Core claim

The central claim is that the hierarchical one-way context aggregation in ECAPA-TDNN's SE-Res2Block is the main bottleneck, not the overall TDNN structure. By processing feature-map subsets in both temporal directions—either by adding a reversed Res2Block, by running two SE-Res2Blocks on the original and channel-flipped feature maps, or by replacing the inner convolutions with Bi-LSTM—the model can exploit both preceding and following context and focus on relevant information. The SE-Res2Bi-LSTM variant is the strongest: it reaches 0.67% EER on VoxCeleb1-O versus 0.87% for ECAPA-TDNN, and also improves on VoxCeleb1-E and H, while adding only about one million parameters. On the more challeng

Load-bearing premise

The load-bearing assumption is that the baseline numbers in Table I are directly comparable to the proposed models, even though the paper does not say the baselines were retrained under the identical pipeline; if those baselines used different training recipes, the reported 23% improvement could be inflated by recipe differences rather than by the proposed blocks.

Editorial extensions

If this is right

  • If SE-Res2Bi-LSTM-ECAPA is correct, TDNN-based speaker verification can reach 0.67% EER on VoxCeleb1-O with about 15.7 million parameters, a strong operating point for practical verification systems.
  • The 23% relative improvement over ECAPA-TDNN at only about one million extra parameters suggests that bidirectional context modeling is a cost-effective way to close the gap between TDNN and 2D-convolution-based models.
  • Because the proposed blocks are drop-in replacements for SE-Res2Block, they can be inserted into other TDNN-based architectures without changing the rest of the pipeline.
  • On CN-Celeb.E, Bi-SE-Res2Block-ECAPA wins under mixed-genre training, indicating that the optimal context mechanism may depend on dataset complexity and training data diversity.
  • The MinDCF results show that the proposed architectures improve EER but not consistently MinDCF on VoxCeleb1-O and H, so score normalization or calibration would need retuning to fully exploit the EER gain.

Reading between the lines

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

  • The paper does not state that the baseline numbers in Table I were retrained under the same 2-second audio, 80-dim Mel, spec-augment, AAM-Softmax, and AS-norm pipeline; if they came from different recipes, part of the reported 23% improvement could be due to training conditions rather than the new blocks. Retraining all baselines identically would settle this.
  • The paper reports only parameter counts, not inference FLOPs or latency; Bi-LSTM is typically more compute-intensive per step than dilated convolutions, so real-time deployment cost remains an open question.
  • The bidirectional multi-scale context idea should transfer to other TDNN-based speech tasks such as language identification, emotion recognition, or voice activity detection, where long-range dependencies matter.
  • The consistent MinDCF degradation on VoxCeleb1-O and H across all three proposed blocks hints that the EER improvement may come with worse score calibration; this could be addressed by tuning the AS-norm enrollment or using a different score normalization scheme.
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

3 major / 5 minor

Summary. The manuscript proposes three architectural variants of ECAPA-TDNN's SE-Res2Block aimed at modeling bidirectional contextual information: SE-Bi-Res2Block (parallel forward and reversed Res2Blocks), Bi-SE-Res2Block (two SE-Res2Blocks on the original and channel-flipped feature maps), and SE-Res2Bi-LSTM (Bi-LSTM replacing the central dilated convolutions). Each variant is inserted into ECAPA-TDNN with channel widths 512 and 1024. Experiments on VoxCeleb1-O/E/H report consistent EER reductions relative to the ECAPA-TDNN numbers in Table I, with SE-Res2Bi-LSTM-ECAPA (C=1024) at 0.67% EER on VoxCeleb1-O, claimed as a 23% relative improvement over the quoted 0.87%. Additional experiments on CN-Celeb.E compare three training types (SG, MG, M). Source code is provided.

Significance. If the comparisons are valid, this is a useful incremental contribution: three clearly described architectural variants, consistent EER improvements across VoxCeleb test sets, and public code. The main caveat is that the baseline values in Table I appear to be quoted from prior publications rather than produced under the paper's own training recipe. Because the central claim is an architectural superiority claim, the absence of same-pipeline baselines makes the quantitative '23% improvement' and 'surpasses all five SOTA' statements unsubstantiated as currently presented. The paper also explicitly acknowledges a MinDCF degradation on VoxCeleb1-O and VoxCeleb1-H, which tempers the 'effectiveness' claim.

major comments (3)
  1. [Section III.B / Table I] The paper specifies a particular pipeline (2s utterances, 80-dim Mel, spec-augment, AAM-Softmax margin 0.2/scale 30, Adam, cyclic LR, AS-norm) but never states that the ECAPA-TDNN or SOTA rows in Table I were retrained under this pipeline. If these rows are taken from their original papers, differences in utterance duration, augmentation, loss, or scoring are conflated with the architectural changes. The headline 23% relative EER gain and the 'surpasses all five SOTA' claim therefore require either (a) same-recipe retraining of the baselines or (b) a clearly restricted claim and an explicit statement of which comparisons are directly comparable.
  2. [Section IV / Table I] The proposed models' MinDCF is systematically worse than ECAPA-TDNN on VoxCeleb1-O and VoxCeleb1-H. For example, SE-Res2Bi-LSTM-ECAPA (C=1024) gives MinDCF 0.1108 vs 0.1066 on VoxCeleb1-O and 0.2212 vs 0.2101 on VoxCeleb1-H, and the other proposed variants show the same pattern. The paper acknowledges this but frames the results as a general improvement. A claim of 'effectiveness' should address this metric-specific trade-off, including its operational consequences and whether the EER gain is robust across operating points.
  3. [Section IV / Table I] Several claimed improvements are small in absolute terms (e.g., SE-Bi-Res2Block C=1024: 1.05% vs 1.12% EER on VoxCeleb1-E; Bi-SE-Res2Block C=512: 0.85% vs 0.87% EER on VoxCeleb1-O). Without repeated runs, error bars, or a significance test, these differences are within typical run-to-run variability for speaker verification systems. At least for the headline result, a repeated-seed or confidence-interval analysis would materially strengthen the claim.
minor comments (5)
  1. [Table I, PCF-ECAPA row] The MinDCF value for VoxCeleb1-O appears as '0.892', which is outside the feasible range for MinDCF and is likely a typo for 0.0892. Please correct and verify all numeric entries in the table.
  2. [Abstract / Section IV] The '23%' and '23%, 11.6%, 14.2%' statements should be explicitly labeled as relative EER reductions, not absolute percentage-point improvements, to avoid ambiguity.
  3. [Section III.B] The cyclical learning rate description says the rate 'undergoes 65k iterations of up and down steps'; please specify the total number of training iterations/epochs and batch repetition schedule for reproducibility.
  4. [Section II.A / Fig. 1] Equation (1) writes the block as SE{[f2(...)]} + X, but Fig. 1 shows the SE block after a Conv1D. Make the order of f2 and the SE block explicit and ensure the equation and figure are consistent.
  5. [Table II] The CN-Celeb baseline rows (Branch-ECAPA, CAM++) are listed under 'M' training, but it is not stated whether these numbers were obtained by retraining under the same recipe or quoted from prior work. Please clarify, as this affects the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical benchmark comparisons; no fitted input is renamed as prediction and no load-bearing self-citation chain is present.

full rationale

The paper's central contribution is the proposal of three TDNN blocks (SE-Bi-Res2Block, Bi-SE-Res2Block, SE-Res2Bi-LSTM) and their empirical evaluation on VoxCeleb and CN-Celeb. The reported improvements (e.g., 23% lower EER on VoxCeleb1-O) are direct experimental outcomes of training and evaluation, not quantities derived from a fitted parameter or from a normalization identity. There is no equation in which a prediction equals an input by construction: the block definitions (Eqs. 1, 4 and the SE-Res2Bi-LSTM description) are architectural designs, not self-referential derivations of the reported EER numbers. The paper cites prior work, but none of the cited results are produced by the present authors, and no uniqueness theorem or ansatz is imported from the authors' own prior papers. The strongest concern—that baseline numbers in Table I may come from different training recipes—is an experimental comparability issue, not a circularity issue, because the quoted baselines are external published results rather than quantities constructed from the paper's own fitted values. The paper also openly acknowledges that MinDCF worsens on some VoxCeleb subsets, further indicating that the claims are presented as observed benchmark results rather than as forced consequences of an internal argument. Therefore no circular step can be identified under the stated criteria.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No explanatory entities are introduced; the three blocks are engineering architecture changes. All quantitative claims rest on the chosen training recipe and on external baseline comparability.

free parameters (5)
  • AAM-Softmax margin = 0.2
    Set in Section III.B; standard choice, but the numeric EER depends on it.
  • AAM-Softmax scale = 30
    Set in Section III.B; standard ECAPA recipe value.
  • SpecAugment frequency/time mask sizes = 8 (freq), 10 (time)
    Chosen hyperparameters in Section III.B; affect regularization and final scores.
  • Utterance duration = 2 seconds
    Input length fixed in Section III.B; impacts the context available to Bi-LSTM.
  • Res2Block scale dimension N = not stated
    N is defined in Section II.A but never specified in the experimental section, so replication requires guessing it.
assumptions (4)
  • domain assumption Published baseline results from earlier papers are comparable under the same evaluation protocol despite different training pipelines.
    Used in Table I; if false, the claimed relative improvements are not attributable to the proposed blocks.
  • domain assumption 80-dimensional Mel-spectrogram features with 25ms Hamming window preserve the speaker information needed for verification.
    Input representation is fixed in Section III.B; standard for the field but not argued here.
  • domain assumption Bi-directional processing along Res2Net channel subsets improves 'contextual information' relevant to speaker identity.
    Core motivation in Section II.A/B; assumed, not empirically isolated from raw capacity increase.
  • standard math Backpropagation and stochastic optimization can train the proposed recurrent blocks effectively.
    Standard training assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Effective Modeling of Critical Contextual Information for TDNN-based Speaker Verification." pith.science (2026). https://pith.science/paper/674MQLX6

@misc{pith2026250909932,
  author       = {Pith},
  title        = {Pith review of: Effective Modeling of Critical Contextual Information for TDNN-based Speaker Verification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/674MQLX6}},
  note         = {Machine review of arXiv:2509.09932}
}
read the original abstract

Today, Time Delay Neural Network (TDNN) has become the mainstream architecture for speaker verification task, in which the ECAPA-TDNN is one of the state-of-the-art models. The current works that focus on improving TDNN primarily address the limitations of TDNN in modeling global information and bridge the gap between TDNN and 2-Dimensional convolutions. However, the hierarchical convolutional structure in the SE-Res2Block proposed by ECAPA-TDNN cannot make full use of the contextual information, resulting in the weak ability of ECAPA-TDNN to model effective context dependencies. To this end, three improved architectures based on ECAPA-TDNN are proposed to fully and effectively extract multi-scale features with context dependence and then aggregate these features. The experimental results on VoxCeleb and CN-Celeb verify the effectiveness of the three proposed architectures. One of these architectures achieves nearly a 23% lower Equal Error Rate compared to that of ECAPA-TDNN on VoxCeleb1-O dataset, demonstrating the competitive performance achievable among the current TDNN architectures under the comparable parameter count.

Figures

Figures reproduced from arXiv: 2509.09932 by the authors.

Figure 2
Figure 2. The proposed Bi-SE-Res2Block. size k = 1 to recover the number of channels. Finally, an SE-Block is used for channel attention processing. The whole process can be expressed by the following formula, X′ = SE{f2 [Res2 (f1 (X)) + Res2 Rev (f1 (X))]}+X (1) where X′ represents output feature map of the SE-Bi￾Res2Block. f1 and f2 represent two 1D convolution layers, while Res2 and Res2 Rev signify Res2Block and reversed … view at source ↗
Figure 1
Figure 1. The proposed SE-Bi-Res2Block that consists of Res2Block and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. The proposed SE-Res2Bi-LSTM block. registered and test speeches, yielding 3,484,292 test pairs that are independent of gender. Furthermore, the MUSAN and RIR datasets are employed for online augmentation, introducing noise and reverberation, respectively. To verify the effectiveness of the proposed architectures on complicated multi-scale datasets, different types of training experiments are conducted, including sin… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 6 linked inside Pith

  1. [1]

    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, 2010

    2010

  2. [2]

    Unsupervised speaker adaptation based on the cosine similarity for text-independent speaker verification

    S. Shum, N. Dehak, R. Dehak, and J. R. Glass, “Unsupervised speaker adaptation based on the cosine similarity for text-independent speaker verification.” inOdyssey, 2010, p. 16

    2010

  3. [3]

    Probabilistic linear discriminant analysis,

    S. Ioffe, “Probabilistic linear discriminant analysis,” inECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part IV 9. Springer, 2006, pp. 531–542

    2006

  4. [4]

    X- vectors: robust DNN embeddings for speaker recognition,

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

    2018

  5. [5]

    Speaker recognition for multi-speaker conversations using x-vectors,

    D. Snyder, D. Garcia-Romero, G. Sell, A. McCree, D. Povey, and S. Khudanpur, “Speaker recognition for multi-speaker conversations using x-vectors,” inICASSP 2019-2019 IEEE International conference on acoustics, speech and signal processing (ICASSP). IEEE, 2019, pp. 5796–5800

    2019

  6. [6]

    A time delay neural net- work architecture for efficient modeling of long temporal contexts,

    V . Peddinti, D. Povey, and S. Khudanpur, “A time delay neural net- work architecture for efficient modeling of long temporal contexts,” in Sixteenth annual conference of the international speech communication association, 2015

    2015

  7. [7]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  8. [8]

    Deep speaker embedding extraction with channel-wise feature responses and additive supervision softmax loss function,

    J. Zhou, T. Jiang, Z. Li, L. Li, and Q. Hong, “Deep speaker embedding extraction with channel-wise feature responses and additive supervision softmax loss function,” inInterspeech, 2019, pp. 2883–2887

    2019

Show all 30 references
  1. [9]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7132–7141

    2018

  2. [10]

    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, 2020, pp. 3830– 3834

    2020

  3. [11]

    Res2net: a new multi-scale backbone architecture,

    S.-H. Gao, M.-M. Cheng, K. Zhao, X.-Y . Zhang, M.-H. Yang, and P. Torr, “Res2net: a new multi-scale backbone architecture,”IEEE transactions on pattern analysis and machine intelligence, vol. 43, no. 2, pp. 652–662, 2019

    2019

  4. [12]

    Branch-ECAPA- TDNN: a parallel branch architecture to capture local and global features for speaker verification,

    J. Yao, C. Liang, Z. Peng, B. Zhang, and X.-L. Zhang, “Branch-ECAPA- TDNN: a parallel branch architecture to capture local and global features for speaker verification,” inProc. Interspeech 2023, 2023, pp. 1943– 1947

    2023

  5. [13]

    Dual-stream time-delay neural network with dynamic global filter for speaker verification,

    Y . Li and X. Lin, “Dual-stream time-delay neural network with dynamic global filter for speaker verification,”arXiv preprint arXiv:2303.11020, 2023

    2023 arXiv

  6. [14]

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

    2023

  7. [15]

    ECAPA++: fine-grained deep embedding learning for TDNN based speaker verification,

    B. Liu and Y . Qian, “ECAPA++: fine-grained deep embedding learning for TDNN based speaker verification,” inProc. Interspeech 2023, 2023, pp. 3132–3136. [Online]. Available: https://api.semanticscholar. org/CorpusID:260910169

    2023

  8. [16]

    PhaseDCN: a phase-enhanced dual-path dilated convolutional network for single- channel speech enhancement,

    L. Zhang, M. Wang, Q. Zhang, X. Wang, and M. Liu, “PhaseDCN: a phase-enhanced dual-path dilated convolutional network for single- channel speech enhancement,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 2561–2574, 2021

    2021

  9. [17]

    V oxCeleb: a large-scale speaker identification dataset,

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

    2017

  10. [18]

    V oxCeleb: large- scale speaker verification in the wild,

    A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “V oxCeleb: large- scale speaker verification in the wild,”Computer Speech & Language, vol. 60, pp. 101 027.1–101 027.15, 2020

    2020

  11. [19]

    V oxceleb2: deep speaker recognition,

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

    2018 arXiv

  12. [20]

    CN-Celeb: a challenging chinese speaker recog- nition dataset,

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

    2020

  13. [21]

    CN-Celeb: multi-genre speaker recognition,

    L. Li, R. Liu, J. Kang, Y . Fan, H. Cui, Y . Cai, R. Vipperla, T. F. Zheng, and D. Wang, “CN-Celeb: multi-genre speaker recognition,”Speech Communication, 2022

    2022

  14. [22]

    Musan: a music, speech, and noise corpus,

    D. Snyder, G. Chen, and D. Povey, “Musan: a music, speech, and noise corpus,”arXiv preprint arXiv:1510.08484, 2015

    2015 arXiv

  15. [23]

    A study on data augmentation of reverberant speech for robust speech recognition,

    T. Ko, V . Peddinti, D. Povey, M. L. Seltzer, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,” in2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2017, pp. 5220–5224

    2017

  16. [24]

    Specaugment: a simple data augmentation method for automatic speech recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “Specaugment: a simple data augmentation method for automatic speech recognition,”arXiv preprint arXiv:1904.08779, 2019

    1904 arXiv

  17. [25]

    Arcface: additive angular margin loss for deep face recognition,

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

    2019

  18. [26]

    Margin matters: to- wards more discriminative deep neural network embeddings for speaker recognition,

    X. Xiang, S. Wang, H. Huang, Y . Qian, and K. Yu, “Margin matters: to- wards more discriminative deep neural network embeddings for speaker recognition,” in2019 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). IEEE, 2019, pp...

    2019

  19. [27]

    Adam: a method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: a method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  20. [28]

    Cyclical learning rates for training neural networks,

    L. N. Smith, “Cyclical learning rates for training neural networks,” in2017 IEEE winter conference on applications of computer vision (WACV). IEEE, 2017, pp. 464–472

    2017

  21. [29]

    Duality temporal-channel-frequency attention enhanced speaker representation learning,

    L. Zhang, Q. Wang, and L. Xie, “Duality temporal-channel-frequency attention enhanced speaker representation learning,” in2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021, pp. 206–213

    2021

  22. [30]

    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,”arXiv preprint arXiv:2303.00332, 2023

    2023 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.