Siamese Capsule Network for End-to-End Speaker Recognition In The Wild

Ali Etemad; Amirhossein Hajavi

REVIEW 4 major objections 4 minor 37 references

Siamese Capsule Network for End-to-End Speaker Recognition In The Wild

T0 review · 4 major / 4 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a Siamese capsule network with dynamic routing, used as the back-end of an end-to-end speaker verification system, achieves a 3.14% equal error rate on VoxCeleb1 and outperforms existing benchmarks while training on…

desk verdict A plausible architecture with an unsupported SOTA claim; the paper needs controlled comparisons and protocol details before the headline EER can be trusted. read the letter →

arxiv 2009.13480 v1 pith:3ZWUUZK3 submitted 2020-09-28 eess.AS cs.LGcs.SD

Amirhossein Hajavi , Ali Etemad This is my paper · ORCID

classification eess.AScs.LGcs.SD

keywords speakerverificationrecognitionSiamesecapsulenetworkdynamicroutingembeddingsimilarityVoxCelebGhostVladequalerrorrate

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 tries to establish that replacing the standard cosine-distance back-end in speaker verification with a trainable Siamese capsule network improves accuracy, and that this improvement is large enough to surpass existing state-of-the-art systems even when training only on the smaller VoxCeleb1 dataset. It reports an equal error rate of 3.14% on VoxCeleb1, beating benchmark models including ones trained on the much larger VoxCeleb2 dataset. If correct, this would mean that a learned part-whole comparison of embedding coordinates, rather than a fixed geometric similarity, is a viable and data-efficient route to speaker verification in the wild.

What carries the argument

The load-bearing object is a Siamese capsule back-end. The enrollment and test embedding vectors of size 4096 are paired index-by-index into 4096 tuples, each tuple treated as a part; four higher capsules of dimension 128 use dynamic routing with three iterations to weight and combine these parts into a whole-utterance similarity representation, which is then passed through a sigmoid for the final score. Dynamic routing, an iterative mechanism that adjusts how strongly each low-level part contributes to each high-level capsule, is what carries the argument: it learns which coordinate tuples matter for same-speaker versus different-speaker decisions, replacing the fixed cosine similarity.

What would settle it

Run the proposed model and the Table 1 baselines on a single fixed VoxCeleb1 trial list with identical preprocessing and score normalization; if the reproduced EER does not fall below the best baseline's 3.22% under that shared protocol, the claimed advantage is not established.

Watch

Extended reading notes

Core claim

The central claim is that paired coordinate tuples of two utterance embeddings can be treated as parts, and dynamic routing between those tuples and a small set of higher capsules learns a similarity score that is more accurate than cosine distance or simple MLP back-ends. For speaker verification, the paper's model uses thin-ResNet34 with GhostVlad aggregation as a frozen front-end, then feeds 4096-dimensional embedding vectors from the aggregation module into four higher capsules of dimension 128 with three dynamic-routing iterations, and maps the aggregated capsule output through a sigmoid to produce the verification score. The reported result is an EER of 3.14% on VoxCeleb1, which the paper claims outperforms all benchmark systems in its comparison, including the same front-end with cosine scoring and systems trained on VoxCeleb2. The paper also claims that capsule back-ends work best when fed directly from the aggregation module rather than from the final fully connected layer, and that using primary capsules hurts performance because their nonlinearities destroy the coordinatewise correspondence between the two embeddings.

Load-bearing premise

The state-of-the-art claim rests on the assumption that all compared systems were evaluated under exactly the same VoxCeleb1 protocol (same trial list, segments, and score normalization), which the paper does not specify.

Editorial extensions

If this is right

Replacing the cosine-distance back-end with the Siamese capsule back-end lowers the equal error rate on VoxCeleb1 from the best compared baseline of 3.22% to 3.14%, according to the paper's reported protocol.
Feeding the back-end with GhostVlad aggregation outputs instead of embeddings from the final fully connected layer improves performance, from 3.65% EER to 3.14% EER with four capsules.
The capsule back-end can be attached to any front-end speaker representation learning model, making the approach a general drop-in replacement for cosine scoring.
Random selection of utterance triplets may make a smaller training set sufficient, since VoxCeleb1 can generate roughly 148,000 times 122 training combinations under the paper's sampling scheme.
The best configuration avoids primary capsules entirely, indicating that the coordinatewise pairing of embeddings is what the higher capsules need to see.

Reading between the lines

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

The paper freezes the front-end to isolate the gain to the back-end, so a natural extension is joint end-to-end training; the margin could grow or shrink once the embedding is also optimized for the capsule score.
If the reported gain transfers to other pair-comparison tasks, the same index-wise tuple routing idea could be applied to face verification, language identification, or text similarity, where cosine or MLP back-ends dominate.
Because the number of capsules and routing iterations was chosen empirically and fixed at four and three, the result does not establish a general recipe; reproducing it across speaker domains or with VoxCeleb2 training would test how much of the gain is dataset-specific.
The paper does not state its exact trial protocol, so the claim cannot be independently checked from the paper alone; publishing the trial list and score normalization would be the cheapest way to test it.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

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,

The one thing you should know: this is a reasonable architecture paper that overclaims. The index-by-index pairing of embedding tuples as parts for a capsule back-end is a distinct design, and the authors were sensible to freeze the front-end and test embedding layers and capsule counts separately. But the headline 3.14% EER is not a demonstrated state of the art. The comparison table pits a VoxCeleb1-trained model against systems trained on VoxCeleb2, with no cosine-similarity baseline using the same front-end and same data. So the claimed advantage of capsules over cosine is confounded by training set and protocol differences. And the protocol is never specified — no trial list, no segment handling, no score normalization — which makes the cross-paper EER comparisons in Table 1 uninterpretable.

The paper does have real merit. The idea of treating each aligned index pair as a part, skipping primary capsules, and routing directly to high-level capsules is clean and does not appear in the cited face or text Siamese capsule papers. The experimental probe of FC vs. GhostVlad embeddings and the effect of primary capsules is informative. If the result holds under a controlled comparison, it is a modest but useful contribution to speaker verification.

Soft spots, in order of severity. First, no same-front-end cosine baseline on VoxCeleb1. The authors compare to Xie et al.'s Thin-ResNet+GhostVlad, but that model was trained on VoxCeleb2; the 3.14 vs 3.22 gap could be due to data, trial list, or implementation details. Second, hyperparameters: the choice of 4 capsules is made by scanning 2/4/6 on the test set, which optimistically biases the reported number. Third, the "end-to-end" and "less training data" claims are undersupported — the front-end is frozen, and the paper never discloses whether the thin-ResNet was initialized from a VoxCeleb2 pretrained checkpoint, which would undermine the "less data" narrative. No code or weights are released. The "found empirically" line in Section 3.2 is honest as far as it goes, but without a validation split it is test-set selection.

Who gets value from this: researchers working on trainable back-ends for speaker verification, and anyone thinking about capsule networks for metric learning. The writing is clear and the architecture is reproducible in principle.

My recommendation: send it to peer review. The submission is not ready as-is, but the core idea is worth a referee's time. A reviewer should require a controlled comparison against cosine on identical embeddings and data, a precise description of the evaluation protocol, and either a validation-based hyperparameter selection or a sensitivity analysis. If the authors can do that, the paper becomes a decent conference contribution.

Referee Report

4 major / 4 minor

Summary. The paper proposes a speaker verification pipeline consisting of a frozen thin-ResNet34/GhostVlad front-end and a Siamese capsule back-end with dynamic routing. The back-end receives paired index-aligned embeddings of dimension 4096, routes them to four higher capsules of dimension 128, and outputs a sigmoid similarity score. Training uses triplet-like utterance pairs from VoxCeleb1 with binary cross-entropy. The authors report an EER of 3.14%, claim that this outperforms all benchmark systems including models trained on VoxCeleb2, and perform ablations on the number of capsules, the embedding layer, and the use of primary capsules.

Significance. If the claimed result is reliable, the contribution is meaningful: it demonstrates that a learned similarity back-end based on dynamic routing can improve on cosine similarity for speaker verification, and it provides useful direct evidence that aggregation-layer embeddings (GhostVlad output) are better than bottleneck FC embeddings for this back-end. The ablation of primary capsules and the empirical comparison of capsule counts are informative. However, no evaluation protocol, code, trained models, or uncertainty estimates are provided, so the headline state-of-the-art claim is currently not independently verifiable.

major comments (4)

[§4.2, Table 1] The state-of-the-art claim is unsupported because the evaluation protocol is not specified. Section 4.1 only says that VoxCeleb is used for training and evaluation, and Section 4.2 reports EER values without stating the trial list, the original versus cleaned VoxCeleb1 evaluation set, the number of trial pairs, the segment duration, or any score normalization. The baseline EERs in Table 1 are taken from different publications (Cai et al., Okabe et al., Xie et al.) that may use different trial sets and back-end scoring; comparing them against 3.14% without a controlled common protocol is not evidence of superiority. The authors should specify the protocol in full and, ideally, evaluate at least the closest baselines under the same protocol.
[§3.2, Table 2] The headline configuration is selected on the test set. Table 2 reports EER for nine configurations, and the best value (4 capsules, no primary capsules, aggregation-layer embeddings) is then reported as the final result. Since the selection criterion is the test-set EER itself, the reported 3.14% is optimistically biased. Moreover, the differences among the top aggregation-layer configurations (3.14, 3.16, 3.18) are small and no error bars, confidence intervals, or significance tests are given, so the paper does not establish that 4 capsules is better than 2 or 6. A validation split should be used for architecture selection, and test-set results should be reported with uncertainty or multiple runs.
[§3.3, §4.2] The claim of using 'substantially less training data' is conditional on undisclosed pretraining of the front-end. Section 3.3 states that the Front-end model remains frozen during training, but the initialization of the thin-ResNet34/GhostVlad front-end is not reported. Because this architecture is borrowed from Xie et al. [24], whose model is trained on VoxCeleb2, it is essential to disclose whether the frozen front-end was pretrained on VoxCeleb2. If it was, the comparison against VoxCeleb2-trained baselines in Table 1 would not demonstrate a data-efficiency advantage for the proposed method.
[§3.3, title and abstract] The description 'end-to-end' is an overstatement as written. The front-end is frozen and only the Siamese capsule back-end is trained, so the model is not trained end-to-end in the usual sense. This contradicts the abstract's 'end-to-end deep model' and the introduction's 'thoroughly end-to-end pipeline.' The authors should either train the full pipeline jointly or qualify the claim by stating explicitly that the front-end is a fixed feature extractor and only the back-end is learned.

minor comments (4)

[§4.1] In the sentence 'the scale of possible triplet combinations for training adds up to 148, 000× 1222,' the multiplier '1222' appears to be a typo for '122,' and the combinatorial count should be stated more precisely (e.g., number of positive and negative utterance pairs).
[§3.2, Eq. (2)] Equation (2) reuses the symbol V_i for both the input tuple and its normalized version; please introduce a distinct normalized notation and clarify that V_i denotes the tuple (v1_i, v2_i) rather than the full embedding vector.
[§4.2] The sentence 'This may be due to the random selection of utterance triplets' is a speculation that is not tested; an ablation or a quantitative comparison of training-set sample counts would be needed to support this explanation.
[§3.3, §4.2] The paper omits several training details needed for reproducibility: the number of training epochs, the exact cyclical learning-rate schedule, data augmentation (if any), and the number of test trials used to compute the reported EER.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical and benchmarked against external baselines; the test-set architecture selection is a validity concern, not circularity.

full rationale

This paper contains no derivation chain of the kind the circularity analysis targets: there is no equation that is claimed to predict X from Y while X is defined in terms of Y, and no parameter is fitted to the target result by construction. The headline claim (EER 3.14% on VoxCeleb1, outperforming prior systems) is an experimental measurement obtained by training the proposed Siamese capsule back-end and comparing it with published baselines, most of which are external (Xie et al., Chung et al., Okabe et al., Cai et al.). The paper's self-citations ([6], [7]) are used as related work or as a baseline, but they are not load-bearing: removing them would not alter the architecture or the claimed EER measurement, and they do not supply a uniqueness theorem or a forced ansatz. The only notable methodological concern is that Table 2 selects the number of capsules (4), the choice of embedding layer (GhostVlad aggregation), and whether to include primary capsules using the same VoxCeleb test set on which the final 3.14% EER is reported. This is a test-set-selection / multiple-comparisons issue that can optimistically bias the reported number, but it is not circularity in the defined sense: the selected configuration's EER is not equal by construction to a fitted function of the inputs, and the comparison against external benchmarks is not mathematically forced by the design choice. The paper should have disclosed the exact evaluation protocol and ideally used a separate validation set, but those are correctness and reporting concerns rather than self-referential reductions. Accordingly, no specific circular step can be quoted and exhibited, and the score is 0.

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

The central claim rests on a frozen pretrained front-end, an index-aligned pairing assumption, and an empirically chosen capsule count. The free parameters are architectural choices not derived from any theory. No new physical, mathematical, or conceptual entities are introduced, so the invented-entities list is empty. The main risks are the unverified protocol assumption and test-set-driven model selection.

free parameters (5)

number_of_higher_capsules = 4
Section 3.2: 'The number of capsules is found empirically'; Table 2 tests 2, 4, and 6 capsules and 4 yields the lowest EER.
capsule_dimension = 128
Section 3.3: 'four higher capsules with a capsule dimension of 128'. No ablation is reported.
dynamic_routing_iterations = 3
Section 3.2: 'dynamic routing [16] with 3 iterations'. No ablation is reported.
initial_learning_rate = 0.01 (cyclical)
Section 3.3: 'learning rate is initially set to 0.01, but is changed with cyclical learning rate pattern'. The schedule is not fully specified.
batch_size = 64
Section 3.3: 'For batch size, 64 utterance triplets are selected at each step of the training'.

assumptions (4)

domain assumption The same VoxCeleb1 evaluation protocol was used as in the cited prior works
The paper does not specify the trial list, score normalization, or utterance segmentation, yet the state-of-the-art claim in Section 4.2 depends on direct comparability of EER numbers.
ad hoc to paper Index-aligned embedding dimensions are semantically comparable parts
Section 3.2: 'Each tuple (v1i, v2i), where v1i is the ith index in the enrollment embedding, and v2i is the same index in the test embedding, is considered a single part'. This is a modeling assumption without independent evidence.
domain assumption The pretrained thin-ResNet+GhostVlad front-end produces speaker-discriminative embeddings
Section 3.1 adopts the architecture from [24] but does not state the training data for the pretrained model or provide weights.
domain assumption Random triplet sampling gives effective training diversity
Section 3.3 claims random selection of triplets reduces overfitting and increases effective samples; no theoretical or empirical support is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Siamese Capsule Network for End-to-End Speaker Recognition In The Wild." pith.science (2026). https://pith.science/paper/3ZWUUZK3

@misc{pith2026200913480,
  author       = {Pith},
  title        = {Pith review of: Siamese Capsule Network for End-to-End Speaker Recognition In The Wild},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3ZWUUZK3}},
  note         = {Machine review of arXiv:2009.13480}
}

read the original abstract

We propose an end-to-end deep model for speaker verification in the wild. Our model uses thin-ResNet for extracting speaker embeddings from utterances and a Siamese capsule network and dynamic routing as the Back-end to calculate a similarity score between the embeddings. We conduct a series of experiments and comparisons on our model to state-of-the-art solutions, showing that our model outperforms all the other models using substantially less amount of training data. We also perform additional experiments to study the impact of different speaker embeddings on the Siamese capsule network. We show that the best performance is achieved by using embeddings obtained directly from the feature aggregation module of the Front-end and passing them to higher capsules using dynamic routing.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

[24]

Stacked capsule autoencoders,

A. Kosiorek, S. Sabour, Y . W. Teh, and G. E. Hinton, “Stacked capsule autoencoders,” Advances in Neural Information Pro- cessing Systems, pp. 15 512–15 522, 2019

work page 2019
[1]

INTRODUCTION Speaker veriﬁcation models are comprised of two main parts. The Front-end component which encodes an utterance into ﬁxed sized embedding vectors [1], and the Back-end compo- nent which measures the similarity of two given vectors in the form of a similarity score [2]. Most commonly in previ- ous studies, the two components of the speaker veri...

work page Pith review arXiv 2009
[2]

RELA TED WORK 2.1. Siamese Networks for Speaker Recognition Deep learning approaches for speaker recognition have gained a lot of attention given the advancements in compu- tational capacity and availability of large in-the-wild datasets [22, 23]. A large number of studies using DNN models for speaker embedding extraction have been performed in the past f...

work page
[3]

The majority of DNN models used in speaker recogni- tion take a single utterance as input and provide a ﬁxed-size vector as the speaker embedding for the utterance

have used TDNN to obtain reliable speaker embeddings using MFCC features. The majority of DNN models used in speaker recogni- tion take a single utterance as input and provide a ﬁxed-size vector as the speaker embedding for the utterance. Another process is then used to calculate the similarity of the two em- beddings obtained from an enrolment utterance ...

work page
[4]

Through the follow- ing sub-sections, we describe the different components of our model

PROPOSED NETWORK We propose an architecture based on a Siamese network with a Back-end that utilizes capsules for directly measuring a sim- ilarity score for speaker veriﬁcation. Through the follow- ing sub-sections, we describe the different components of our model. An overview of the model is presented in Figure 1. 3.1. Front-end We utilize a model with...

work page
[5]

Dataset In this paper, the V oxCeleb dataset is used for both training and evaluation

EXPERIMENT SETUP AND RESULTS 4.1. Dataset In this paper, the V oxCeleb dataset is used for both training and evaluation. The training set consists of approximately 148k utterances spoken by 1,211 speakers. Using the random selection of utterances for triplets, the scale of possible triplet combinations for training adds up to 148, 000× 1222, where the num...

work page
[6]

CONCLUSION AND FUTURE WORK In this paper a novel Siamese network using capsules and dy- namic routing was proposed for speaker veriﬁcation in the wild. Our end-to-end pipeline used thin-ResNet as its Front- end component for speech representation learning, while cap- sules were used in its Back-end to extract part-whole rela- tions of the embeddings later...

work page
[7]

ACKNOWLEDGEMENTS The authors would like to thank IMRSV Data Labs for their support of this work and also acknowledge the Natural Sci- ences and Engineering Research Council of Canada (NSERC) for supporting this research (grant no.: CRDPJ 533919-18)

work page

Show all 37 references

[8]

Front-end factor analysis for speaker veriﬁcation,

N. Dehak, P. J. Kenny, R. Dehak, P. Dumouchel, and P. Ouel- let, “Front-end factor analysis for speaker veriﬁcation,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 19, no. 4, pp. 788–798, 2010

2010
[9]

Bayesian speaker veriﬁcation with, heavy tailed pri- ors,

P. Kenny, “Bayesian speaker veriﬁcation with, heavy tailed pri- ors,”Odyssey, 2010

2010
[10]

Speaker recognition for multi-speaker con- versations using x-vectors,

D. Snyder, D. Garcia-Romero, G. Sell, A. McCree, D. Povey, and S. Khudanpur, “Speaker recognition for multi-speaker con- versations using x-vectors,”IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), pp. 5796– 5800, 2019

2019
[11]

Centroid-based deep metric learning for speaker recognition,

J. Wang, K.-C. Wang, M. Law, F. Rudzicz, and M. Brudno, “Centroid-based deep metric learning for speaker recognition,” IEEE International Conference on Acoustics, Speech, and Sig- nal Processing (ICASSP), pp. 3652–3656, 2019

2019
[12]

Deep speaker repre- sentation using orthogonal decomposition and recombination for speaker veriﬁcation,

I. Kim, K. Kim, J. Kim, and C. Choi, “Deep speaker repre- sentation using orthogonal decomposition and recombination for speaker veriﬁcation,” IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), pp. 6126– 6130, 2019

2019
[13]

A deep neural network for short- segment speaker recognition,

A. Hajavi and A. Etemad, “A deep neural network for short- segment speaker recognition,” INTERSPEECH, pp. 2878– 2882, 2019

2019
[14]

Knowing what to listen to: Early attention for deep speech representation learning,

——, “Knowing what to listen to: Early attention for deep speech representation learning,” arXiv preprint arXiv:2009.01822, 2020

2009 arXiv
[15]

End-to-end dnn based speaker recognition in- spired by i-vector and plda,

J. Rohdin, A. Silnova, M. Diez, O. Plchot, P. Mat ˇejka, and L. Burget, “End-to-end dnn based speaker recognition in- spired by i-vector and plda,”IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 4874– 4878, 2018

2018
[16]

End-to- end attention based text-dependent speaker veriﬁcation,

S.-X. Zhang, Z. Chen, Y . Zhao, J. Li, and Y . Gong, “End-to- end attention based text-dependent speaker veriﬁcation,”IEEE Spoken Language Technology Workshop (SLT) , pp. 171–178, 2016

2016
[17]

Deep siamese architecture based replay detection for secure voice biometric

K. Sriskandaraja, V . Sethu, and E. Ambikairajah, “Deep siamese architecture based replay detection for secure voice biometric.”INTERSPEECH, pp. 671–675, 2018

2018
[18]

Siamese capsule networks,

J. O. Neill, “Siamese capsule networks,” arXiv preprint arXiv:1805.07242, 2018

2018 arXiv
[19]

Siamese neural networks: An overview,

D. Chicco, “Siamese neural networks: An overview,” Artiﬁcial Neural Networks, pp. 73–94, 2020

2020
[20]

Convolutional neural network and language embeddings for end-to-end dialect recognition,

S. Shon, A. Ali, and J. Glass, “Convolutional neural network and language embeddings for end-to-end dialect recognition,” Odyssey, The Speaker and Language Recognition Workshop , pp. 98–104, 2018

2018
[21]

Siamese x-vector reconstruction for domain adapted speaker recognition,

S. Rozenberg, H. Aronowitz, and R. Hoory, “Siamese x-vector reconstruction for domain adapted speaker recognition,” arXiv preprint arXiv:2007.14146, 2020

2007 arXiv
[22]

Sampling strategies in siamese networks for unsu- pervised speech representation learning,

R. Riad, C. Dancette, J. Karadayi, N. Zeghidour, T. Schatz, and E. Dupoux, “Sampling strategies in siamese networks for unsu- pervised speech representation learning,” INTERSPEECH, pp. 2658–2662, 2018

2018
[23]

Dynamic routing be- tween capsules,

S. Sabour, N. Frosst, and G. E. Hinton, “Dynamic routing be- tween capsules,” inAdvances in neural information processing systems, 2017, pp. 3856–3866

2017
[25]

Speech emotion recognition using capsule networks,

X. Wu, S. Liu, Y . Cao, X. Li, J. Yu, D. Dai, X. Ma, S. Hu, Z. Wu, X. Liu et al. , “Speech emotion recognition using capsule networks,” IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP), pp. 6695–6699, 2019

2019
[26]

Residual+ capsule networks (rescap) for simultaneous single-channel overlapped keyword recognition

Y . Xiong, V . Berisha, and C. Chakrabarti, “Residual+ capsule networks (rescap) for simultaneous single-channel overlapped keyword recognition.”INTERSPEECH, pp. 3337–3341, 2019

2019
[27]

Capsule attention for multi- modal eeg and eog spatiotemporal representation learning with application to driver vigilance estimation,

G. Zhang and A. Etemad, “Capsule attention for multi- modal eeg and eog spatiotemporal representation learning with application to driver vigilance estimation,”arXiv preprint arXiv:1912.07812, 2019

1912 arXiv
[28]

Matrix capsules with em routing,

G. E. Hinton, S. Sabour, and N. Frosst, “Matrix capsules with em routing,” inInternational conference on learning represen- tations, 2018

2018
[29]

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, p. 101027, 2020

2020
[30]

The speak- ers in the wild (sitw) speaker recognition database,

M. McLaren, L. Ferrer, D. Castan, and A. Lawson, “The speak- ers in the wild (sitw) speaker recognition database,” INTER- SPEECH, pp. 818–822, 2016

2016
[31]

Utterance-level aggregation for speaker recognition in the wild,

W. Xie, A. Nagrani, J. S. Chung, and A. Zisserman, “Utterance-level aggregation for speaker recognition in the wild,” IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), pp. 5791–5795, 2019

2019
[32]

Seq2seq attentional siamese neural networks for text-dependent speaker veriﬁcation,

Y . Zhang, M. Yu, N. Li, C. Yu, J. Cui, and D. Yu, “Seq2seq attentional siamese neural networks for text-dependent speaker veriﬁcation,” ICASSP, pp. 6131–6135, 2019

2019
[33]

Siamese capsule networks with global and local features for text classiﬁcation,

Y . Wu, J. Li, J. Wu, and J. Chang, “Siamese capsule networks with global and local features for text classiﬁcation,” Neuro- computing, 2020

2020
[34]

Analysis of Length Normalization in End-to-End Speaker Veriﬁcation System,

W. Cai, J. Chen, and M. Li, “Analysis of Length Normalization in End-to-End Speaker Veriﬁcation System,” INTERSPEECH, pp. 3618–3622, 2018

2018
[35]

V oxceleb2: Deep speaker recognition,

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

2018
[36]

Attentive Statistics Pooling for Deep Speaker Embedding,

K. Okabe, T. Koshinaka, and K. Shinoda, “Attentive Statistics Pooling for Deep Speaker Embedding,” INTERSPEECH, pp. 2252–2256, 2018

2018
[37]

Cyclical learning rates for training neural net- works,

L. N. Smith, “Cyclical learning rates for training neural net- works,”IEEE Winter Conference on Applications of Computer Vision (WACV), pp. 464–472, 2017

2017

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [24]

Stacked capsule autoencoders,

A. Kosiorek, S. Sabour, Y . W. Teh, and G. E. Hinton, “Stacked capsule autoencoders,” Advances in Neural Information Pro- cessing Systems, pp. 15 512–15 522, 2019

work page 2019

[2] [1]

INTRODUCTION Speaker veriﬁcation models are comprised of two main parts. The Front-end component which encodes an utterance into ﬁxed sized embedding vectors [1], and the Back-end compo- nent which measures the similarity of two given vectors in the form of a similarity score [2]. Most commonly in previ- ous studies, the two components of the speaker veri...

work page Pith review arXiv 2009

[3] [2]

RELA TED WORK 2.1. Siamese Networks for Speaker Recognition Deep learning approaches for speaker recognition have gained a lot of attention given the advancements in compu- tational capacity and availability of large in-the-wild datasets [22, 23]. A large number of studies using DNN models for speaker embedding extraction have been performed in the past f...

work page

[4] [3]

The majority of DNN models used in speaker recogni- tion take a single utterance as input and provide a ﬁxed-size vector as the speaker embedding for the utterance

have used TDNN to obtain reliable speaker embeddings using MFCC features. The majority of DNN models used in speaker recogni- tion take a single utterance as input and provide a ﬁxed-size vector as the speaker embedding for the utterance. Another process is then used to calculate the similarity of the two em- beddings obtained from an enrolment utterance ...

work page

[5] [4]

Through the follow- ing sub-sections, we describe the different components of our model

PROPOSED NETWORK We propose an architecture based on a Siamese network with a Back-end that utilizes capsules for directly measuring a sim- ilarity score for speaker veriﬁcation. Through the follow- ing sub-sections, we describe the different components of our model. An overview of the model is presented in Figure 1. 3.1. Front-end We utilize a model with...

work page

[6] [5]

Dataset In this paper, the V oxCeleb dataset is used for both training and evaluation

EXPERIMENT SETUP AND RESULTS 4.1. Dataset In this paper, the V oxCeleb dataset is used for both training and evaluation. The training set consists of approximately 148k utterances spoken by 1,211 speakers. Using the random selection of utterances for triplets, the scale of possible triplet combinations for training adds up to 148, 000× 1222, where the num...

work page

[7] [6]

CONCLUSION AND FUTURE WORK In this paper a novel Siamese network using capsules and dy- namic routing was proposed for speaker veriﬁcation in the wild. Our end-to-end pipeline used thin-ResNet as its Front- end component for speech representation learning, while cap- sules were used in its Back-end to extract part-whole rela- tions of the embeddings later...

work page

[8] [7]

ACKNOWLEDGEMENTS The authors would like to thank IMRSV Data Labs for their support of this work and also acknowledge the Natural Sci- ences and Engineering Research Council of Canada (NSERC) for supporting this research (grant no.: CRDPJ 533919-18)

work page