Pith. sign in

REVIEW 4 major objections 7 minor 44 references

Beyond Residual Connections: Manifold-Constrained Hyper-Connections for Robust Speaker Representation Learning

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that replacing standard residual shortcuts with manifold-constrained hyper-connections reduces speaker-verification error across four backbones at no added parameter cost.

desk verdict Borrowed mechanism, honest attribution, plausible but under-evidenced gains; needs multiple seeds and a corrected energy-conservation claim before publication. read the letter →

arxiv 2608.05549 v1 pith:MBVYLNKH submitted 2026-08-06 cs.SD cs.MM

Zezhong Jin , Xiaoyu Wang , Zhe Li , Chong-Xin Gan , Zilong Huang , Man-Wai Mak , Kong Aik Lee This is my paper · ORCID
classification cs.SDcs.MM
keywords speakerverificationrecognitionresidualconnectionshyper-connectionsmanifold-constraineddoublystochasticmatrixECAPA-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

The paper is trying to establish that the residual connection, the identity shortcut used in deep speaker embedding networks, can be improved by replacing it with manifold-constrained hyper-connections (mHC), a multi-stream mixing operation whose mixing matrix is constrained to be doubly stochastic. Because the matrix has unit row and column sums, the shortcut preserves overall signal scale the way an identity mapping does, while letting every stream combine information from all previous streams. The authors test this by swapping mHC into four standard backbones and report lower equal error rates and detection costs on VoxCeleb1 and VoxSRC21-val with no increase in parameters. If the claim holds, mHC is a drop-in architectural upgrade for speaker embedding networks that costs almost nothing at inference time.

What carries the argument

The load-bearing object is the stream-wise update $h_i^{l+1} = \sum_j W_{ij} h_j^l + f(x_l)_i$, where W is a learnable $N \times N$ doubly stochastic matrix produced by exponentiating a parameter matrix and then alternating row and column normalization. The formula does two jobs at once: the weighted sum over all N previous streams is the new inter-channel information exchange, and the doubly stochastic constraint preserves total signal scale, so gradients do not blow up or vanish as the network deepens. The paper's static parameterization of W, rather than generating W from the input features, keeps the added cost at $O(N^2)$.

What would settle it

Retrain each backbone with the stream count chosen on a held-out split, for example VoxSRC21-val, instead of on VoxCeleb1-O/E/H, and report mean and standard deviation over several random seeds. If mHC no longer beats the corresponding residual baseline on VoxCeleb1, or if the gains shrink to within run-to-run noise, the paper's consistent-enhancement claim would be falsified.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's terms, is that constraining the mixing matrix of Hyper-Connections to the doubly stochastic manifold makes the multi-stream shortcut stable enough to improve speaker recognition. In mHC, the hidden state at each layer is split into N parallel streams; after the transformation block, stream i is updated as $h_i^{l+1} = \sum_j W_{ij} h_j^l + f(x_l)_i$, with W nonnegative and with all row and column sums equal to one, obtained by alternating row and column normalization. That constraint restores the scale-preserving property of an identity shortcut, which the paper calls energy conservation, while the dense W provides the inter-channel mixing that identity addition lacks. The authors report that this single change lowers equal error rate on VoxCeleb1-O/E/H for all four tested backbones, with the largest relative gains on the large ECAPA-TDNN variant, and also improves VoxSRC21-val.

Load-bearing premise

The headline gains depend on choosing the stream count by lowest error on the same test sets where the results are reported, and the paper reports single runs without error bars, so selection rather than the mHC mechanism could explain part of the improvement.

Editorial extensions

If this is right

  • mHC can be dropped into existing speaker backbones by replacing only the residual shortcut, leaving input and output dimensions, layer count, and parameter count unchanged.
  • Speaker embedding networks gain lower equal error rate and minimum detection cost on VoxCeleb1-O/E/H and VoxSRC21-val without meaningful added computation.
  • The doubly stochastic constraint is the reason the gain is stable: it preserves signal scale, so multi-stream mixing does not reintroduce the gradient explosion or vanishing that unconstrained hyper-connections suffer.
  • The effect appears across architectures with different internal designs, including CNN residual blocks and TDNN SE-Res2Blocks, suggesting the mechanism is not tied to one backbone family.
  • In the tested range, a smaller stream count (N=4) works better than larger counts, giving a practical default for future applications.

Reading between the lines

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

  • The same shortcut replacement could be tested in non-speech residual networks, such as image classifiers or Transformers, where identity shortcuts are also standard; similar relative gains there would support the energy-conservation mechanism beyond speaker embeddings.
  • Because mHC adds no parameters and negligible FLOPs, its reported value is concentrated in representation quality; a natural extension is to check whether the gains persist with much larger training data or longer utterances, where baseline models also improve.
  • Combining mHC with training-time techniques such as knowledge distillation or self-supervised pretraining is untested but plausible, since those settings stress gradient stability and representation diversity.
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

4 major / 7 minor

Summary. The paper adapts Manifold-Constrained Hyper-Connections (mHC), originally proposed by Xie et al. [24], to speaker embedding networks. It replaces the additive identity shortcut in residual blocks with an update that maintains N parallel streams and mixes them with a learnable doubly stochastic matrix W, obtained by Sinkhorn-Knopp projection of exp(Theta). The authors integrate mHC into ResNet-34, Res2Net, ECAPA-TDNN-S, and ECAPA-TDNN-L, and report EER/MinDCF on VoxCeleb1-O/E/H and VoxSRC21-val, an N=4 ablation on ECAPA-L, a comparison with unconstrained HC, and a complexity analysis. The central claim is that mHC is a universal drop-in replacement that consistently improves speaker verification at negligible parameter and computational overhead, with an energy-conservation property that stabilizes training.

Significance. An architecture-level drop-in replacement that improves EER across all four backbones without changing model size would be practically valuable. The paper uses standard external benchmarks, publicly available training data (VoxCeleb2), an established toolkit, and a useful comparison against unconstrained HC. The method's theoretical framing is not proven, however, and the main empirical claim is currently supported only by single-run numbers that were partly tuned on the evaluation sets. The contribution is therefore promising but needs a substantially stronger evidence base before the paper's conclusions can be accepted.

major comments (4)
  1. [§4.2 (Fig. 2) and Table 1] The number of streams is selected as N=4 because it gives the lowest EER on the VoxCeleb1-O/E/H test sets, and Table 1 then reports the mHC results on those same test sets. All systems are single runs without error bars. This makes the claimed consistent improvements vulnerable to selection and run-to-run noise. Please fix N in advance on a development split, report results for all N values for every backbone, and provide multiple seeds with means and confidence intervals, at least for the headline ECAPA-L and ResNet-34 comparisons.
  2. [§2.1 (Eq. 1) and Abstract] The paper states that the Sinkhorn projection of W onto the doubly stochastic manifold ensures energy conservation by preserving signal intensity and feature mean. A doubly stochastic matrix has row and column sums equal to one, so it preserves the sum of the stream vectors in Eq. (1), but it does not preserve the L2 norm, the energy, or the feature mean, particularly after the additive term f(x_l)_i. This theoretical claim is unproven as stated. Please define the invariant precisely and prove it under Eq. (1), or replace the claim with the weaker column-sum conservation property.
  3. [Table 1 and Section 5] The conclusions claim consistent EER and MinDCF reductions, but two MinDCF cells in Table 1 move in the wrong direction: mHC-Res2Net on VoxCeleb1-O (0.154 vs 0.150) and mHC-ECAPA-S on VoxCeleb1-O (0.107 vs 0.106). Several EER gains are also very small, for example mHC-Res2Net on VoxCeleb1-E (1.40 vs 1.41). Please qualify the consistency claim, or show with multi-run statistics that these discrepancies are within noise.
  4. [Abstract and Section 3.2] The abstract and the introduction list E-Res2Net among the validated backbones, but Table 1 and Table 2 contain no E-Res2Net experiment; the four evaluated backbones are ResNet34, Res2Net, ECAPA-TDNN-S, and ECAPA-TDNN-L. The claim should be corrected or the E-Res2Net experiment should be added, otherwise the paper overstates the tested scope of the universal drop-in claim.
minor comments (7)
  1. [Index Terms] The index term 'doubly-stochostic' should be 'doubly-stochastic'.
  2. [§3.2 and §2.1] The paper does not explicitly state that N=4 was used for all backbones; Fig. 2 tunes N only on ECAPA-TDNN-L. Please state the value of N for every architecture and explain how the feature channels are partitioned into N streams at each stage.
  3. [Eq. (1)] The notation f(x_l)_i and the operators H_pre and H_post are described informally; please provide a precise definition of how the unified feature map x_l is formed from the streams and how the output is partitioned back into f(x_l)_1, ..., f(x_l)_N, including the treatment of channel counts at stage transitions.
  4. [Figure 2] The figure would benefit from explicit axis labels and a statement that each point is a single run; the current caption does not make the absence of error bars clear.
  5. [Table 3] The comparison between HC and mHC reports only one EER value per method; please state the training protocol and add error bars or multiple runs, since the 0.07-point gap is within the range of plausible run-to-run variation.
  6. [Table 1] The parameter counts are rounded to two decimals; since W adds learnable parameters, please state the exact parameter increase or clarify that it is below the displayed precision.
  7. [§2.1] The paper says Sinkhorn-Knopp iterations are applied with k=3, but it does not state whether exactly three unrolled iterations are used in the forward/backward computation and whether gradients flow through the iterations; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: mHC is a borrowed external method evaluated on external benchmarks, and the energy-conservation property follows algebraically from the doubly stochastic constraint rather than from a fitted value.

full rationale

The paper's central derivation is Eq. (1), h_i^{l+1} = sum_j W_ij h_j^l + f(x_l)_i, with W projected to a doubly stochastic matrix via Sinkhorn-Knopp iterations. The claimed energy conservation follows directly from the unit row and column sums of W, so it is a mathematical consequence of the construction, not a prediction that reduces to a fitted input. The method itself is attributed to external prior work: 'Xie et al. [24] introduce Manifold-Constrained Hyper-Connections (mHC)' (Section 1), and the in-paper contribution is an efficient static parameterization plus application to speaker backbones. Evaluation uses external benchmarks, VoxCeleb1-O/E/H and VoxSRC21-val, against baselines from other groups. The numerous author self-citations appear in the introduction and reference list as background related work, not as load-bearing premises for the mHC design or for the experimental claims. Two caveats are worth noting but are not circularity: (1) Section 4.2 selects N=4 by lowest EER on the same VoxCeleb1 test sets later reported in Table 1, a benchmark-selection and statistical-reliability concern, not a reduction by the paper's own equations; and (2) the energy-conservation statement in Section 2.1 is asserted without a formal proof, an omitted derivation/soundness gap. Also, Table 1 shows MinDCF worsening for mHC-Res2Net on VoxCeleb1-O (0.150 to 0.154) and for mHC-ECAPA-S on VoxCeleb1-O (0.106 to 0.107), which weakens the 'consistent reductions' claim empirically but is not circular. No step in the paper equates an output to an input by construction or relies on a self-citation chain to force the conclusion.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central mechanism is borrowed from ref [24], so the paper's own ledger is light: the only hand-set hyperparameters are the stream count N and the Sinkhorn iteration count k. The load-bearing domain assumptions are that double stochasticity provides the claimed 'energy conservation' and gradient stability, and that a static input-independent mixing matrix preserves the benefits of the original input-dependent mHC.

free parameters (2)
  • Number of streams N = 4
    Set to 4 because it gave the lowest EER among {4,8,16,32} on the VoxCeleb1 test sets (Fig. 2); the headline results use this test-set-tuned value.
  • Sinkhorn-Knopp iterations k = 3
    Chosen as a small fixed number (Section 2.1, 'typically k=3'); not ablated, and it controls how closely W approximates exact double stochasticity.
assumptions (3)
  • standard math Sinkhorn-Knopp alternating normalization converges to a unique doubly stochastic matrix
    Section 2.1 relies on this classical result to build W from exp(Theta) without proving convergence.
  • domain assumption A doubly stochastic mixing matrix preserves signal scale and thereby stabilizes gradients
    Section 2.1 claims 'energy conservation' and stability from unit row and column sums; a doubly stochastic matrix preserves the sum of stream entries, but the paper does not show it preserves L2 energy or that this stabilizes gradients in deep speaker networks.
  • domain assumption Replacing identity shortcuts with a learned, input-independent W is an effective drop-in change for all four tested backbones
    Section 2.2 assumes the swap keeps optimization and convergence behavior comparable; the only evidence is final EER, with no analysis of training curves or convergence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Residual Connections: Manifold-Constrained Hyper-Connections for Robust Speaker Representation Learning." pith.science (2026). https://pith.science/paper/MBVYLNKH

@misc{pith2026260805549,
  author       = {Pith},
  title        = {Pith review of: Beyond Residual Connections: Manifold-Constrained Hyper-Connections for Robust Speaker Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MBVYLNKH}},
  note         = {Machine review of arXiv:2608.05549}
}
read the original abstract

Residual connections are fundamental to deep speaker recogni- tion models, such as ECAPA-TDNN and ResNet. However, standard identity mapping limits information flow to a sin- gle path, constraining representation capacity. We introduce Manifold-Constrained Hyper-Connections (mHC), reformulat- ing residual paths as a multi-stream evolution where informa- tion is mixed through a doubly stochastic matrix. By employing Sinkhorn-Knopp iterations, mHC ensures energy conservation by preserving signal intensity and feature mean, which stabi- lizes gradients and mitigates signal degradation in complex net- works. We evaluate mHC by replacing standard residual con- nections in backbones including ECAPA-TDNN, ResNet-34, Res2Net, and E-Res2Net. Extensive experiments on VoxCeleb1 demonstrate that mHC connections consistently enhance per- formance across all architectures, highlighting its effectiveness for robust speaker representation learning.

Figures

Figures reproduced from arXiv: 2608.05549 by the authors.

Figure 1
Figure 1. Architectural comparison between the standard residual shortcut and mHC: (a) standard residual connection, (b) high-level mHC overview, and (c) detailed mHC design, where Hpre aggregates streams before F(·) and Hpost splits the output back into streams. fectively reduced from O(nCn2 ) to O(n 2 ). This refinement is vital for lightweight backbones like ECAPA-TDNN, enabling the model to enjoy the benefits of multi-str… view at source ↗
Figure 3
Figure 3. Computational cost (GFLOPs) comparison between baseline backbones and their mHC variants. complexity analysis comparing the baseline models with their mHC-enhanced versions across two scales: ECAPA-TDNN-S and ECAPA-TDNN-L. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Effect of the number of parallel streams N in mHC [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 29 canonical work pages

  1. [24]

    mHC: Manifold-constrained hyper- connections,

    Z. Xie, Y . Wei, H. Cao, C. Zhao, C. Deng, J. Li, D. Dai, H. Gao, J. Chang, L. Zhaoet al., “mHC: Manifold-constrained hyper- connections,”arXiv preprint arXiv:2512.24880, 2025

    arXiv 2025

  2. [1]

    Beyond Residual Connections: Manifold-Constrained Hyper-Connections for Robust Speaker Representation Learning

    Introduction Speaker recognition has become an indispensable technology in various security-sensitive domains, including biometric authen- tication, secure financial transactions, and forensic analysis. In recent years, the rapid advancement of speaker recognition has been primarily driven by the development of powerful deep learning architectures [1, 2, ...

    work page Pith review arXiv 2026

  3. [2]

    Energy Conservation

    Methodology 2.1. Manifold-Constrained Hyper-Connections Fig. 1a illustrates the standard residual connection. At layer l, the output is computed asx l+1 =x l +F(x l), wherex l is passed through the shortcut unchanged and then added to the transformation blockF(x l). This identity shortcut helps to train deep networks, but the shortcut itself ischannel-wis...

    work page

  4. [3]

    Datasets Training data.We used the V oxCeleb2 development set [28] (5,994 speakers) for training

    Experiment Setting 3.1. Datasets Training data.We used the V oxCeleb2 development set [28] (5,994 speakers) for training. V oxCeleb2 is multilingual, while the majority of utterances are in English. Evaluation data.We report results on the standard V ox- Celeb1 test sets [29], namely V ox-O, V ox-E, and V ox-H. V ox-O contains 37,611 trials from 40 speake...

    work page 2021

  5. [4]

    Performance on VoxCeleb and VoxSRC21-val Table 1 summarizes the results on the V oxCeleb1 test sets (V oxCeleb1-O/E/H)

    Experiment Results 4.1. Performance on VoxCeleb and VoxSRC21-val Table 1 summarizes the results on the V oxCeleb1 test sets (V oxCeleb1-O/E/H). Replacing the standard residual short- cut with mHC consistently improves performance across all four backbones without increasing model parameters. For Res2Net, mHC reduces EER from 1.56% to 1.41% on V oxCeleb1-O...

    work page

  6. [5]

    Conclusions In this paper, we investigated the application of Manifold- Constrained Hyper-Connections (mHC) to enhance speaker recognition systems. By integrating mHC into ResNet, Res2Net, and ECAPA-TDNN, we demonstrated that replac- ing conventional residual shortcuts with manifold-constrained multi-stream mixing improves speaker embeddings while pre- se...

    work page

  7. [6]

    15228223), and The Hong Kong Polytechnic University, Project ID P0049192

    Acknowledgement This work was supported in part by the Research Grants Council of the Hong Kong SAR (Grant No. 15228223), and The Hong Kong Polytechnic University, Project ID P0049192

    work page

  8. [7]

    All scientific content, experiments, anal- yses, and conclusions were produced and verified by the au- thors

    Generative AI Use Disclosure Generative AI tools were used only for language polishing and formatting assistance. All scientific content, experiments, anal- yses, and conclusions were produced and verified by the au- thors

    work page

Show all 44 references
  1. [8]

    Distilling attention knowledge for speaker verifica- tion,

    Z. Jin, S. Liu, Z. Li, C.-X. Gan, Z. Huang, M.-W. Mak, and K. A. Lee, “Distilling attention knowledge for speaker verifica- tion,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 16 447–16 451

    2026

  2. [9]

    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

  3. [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,” inInterspeech 2020, 2020, pp. 3830–3834

    2020

  4. [11]

    Uncertainty- aware multi-head multi-mode knowledge distillation for self- supervised speaker verification,

    Z. Jin, Y . Tu, Z. Li, M.-W. Mak, and K.-A. Lee, “Uncertainty- aware multi-head multi-mode knowledge distillation for self- supervised speaker verification,”IEEE Transactions on Audio, Speech and Language Processing, vol. 34, pp. 2101–2114, 2026

    2026

  5. [12]

    Adversar- ially adaptive temperatures for decoupled knowledge distillation with applications to speaker verification,

    Z. Jin, Y . Tu, C.-X. Gan, M.-W. Mak, and K.-A. Lee, “Adversar- ially adaptive temperatures for decoupled knowledge distillation with applications to speaker verification,”Neurocomputing, vol. 624, p. 129481, 2025

    2025

  6. [13]

    W-GVKT: Within-global-view knowledge transfer for speaker verification,

    Z. Jin, Y . Tu, and M.-W. Mak, “W-GVKT: Within-global-view knowledge transfer for speaker verification,” inInterspeech 2024, 2024, pp. 3779–3783

    2024

  7. [14]

    De- noising student features with diffusion models for knowledge dis- tillation in speaker verification,

    Z. Jin, Y . Tu, Z. Li, Z. Huang, C.-X. Gan, and M.-W. Mak, “De- noising student features with diffusion models for knowledge dis- tillation in speaker verification,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing, 2025, pp. 1–5

    2025

  8. [15]

    Disentangling speech representations learning with latent dif- fusion for speaker verification,

    Z. Li, M.-W. Mak, J.-T. Chien, M. Pilanci, Z. Jin, and H. Meng, “Disentangling speech representations learning with latent dif- fusion for speaker verification,”IEEE Transactions on Audio, Speech and Language Processing, 2025

    2025

  9. [16]

    Residual networks be- have like ensembles of relatively shallow networks,

    A. Veit, M. J. Wilber, and S. Belongie, “Residual networks be- have like ensembles of relatively shallow networks,”Advances in Neural Information Processing Systems, vol. 29, 2016

    2016

  10. [17]

    Vision transformer: ViT and its derivatives,

    Z. Fu, “Vision transformer: ViT and its derivatives,”arXiv preprint arXiv:2205.11239, 2022

    2022 arXiv

  11. [18]

    Varia- tional regularization for end-to-end speech deepfake detection,

    S. Qin, K. A. Lee, M.-W. Mak, P. Lisena, and M. Todisco, “Varia- tional regularization for end-to-end speech deepfake detection,” in 2025 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). IEEE, 2025, pp. 2241–2246

    2025

  12. [19]

    Self-supervised learning with multi-head multi-mode knowledge distillation for speaker verifi- cation,

    Z. Jin, Y . Tu, and M.-W. Mak, “Self-supervised learning with multi-head multi-mode knowledge distillation for speaker verifi- cation,” inProc. Interspeech 2024, 2024, pp. 4723–4727

    2024

  13. [20]

    An enhanced res2net with local and global feature fusion for speaker verification,

    Y . Chen, S. Zheng, H. Wang, L. Cheng, Q. Chen, and J. Qi, “An enhanced res2net with local and global feature fusion for speaker verification,” inInterspeech 2023, 2023, pp. 2228–2232

    2023

  14. [21]

    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

  15. [22]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 4700–4708

    2017

  16. [23]

    Reducing overfitting in deep networks by decorrelating representations,

    M. Cogswell, F. Ahmed, R. Girshick, L. Zitnick, and D. Batra, “Reducing overfitting in deep networks by decorrelating representations,” 2016. [Online]. Available: https://arxiv.org/abs/ 1511.06068

    2016 arXiv

  17. [25]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, 2017

    2017

  18. [26]

    Trink: Ink generation with trans- former network,

    Z. Jin, S. Desai, X. Chen, B. Fang, Z. Huang, Z. Li, C.-X. Gan, X. Tu, M.-W. Mak, Y . Luet al., “Trink: Ink generation with trans- former network,” inProceedings of the 2025 Conference on Em- pirical Methods in Natural Language Processing, 2025, pp. 4857– 4864

    2025

  19. [27]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language models are few-shot learners,” inAdvances in Neural Information Processing Systems (NeurIPS), 2020. [Online]. Available: https://arxiv.org/abs/2...

    2020 arXiv

  20. [28]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” inProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2019. [Online]. Av...

    2019 arXiv

  21. [29]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azharet al., “Llama: Open and efficient foundation language models,”arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  22. [30]

    Mm-nodeformer: Node transformer multimodal fusion for emotion recognition in conver- sation,

    Z. Huang, M.-W. Mak, K. A. Leeet al., “Mm-nodeformer: Node transformer multimodal fusion for emotion recognition in conver- sation,” 2024

    2024

  23. [31]

    Residual: Transformer with dual residual connec- tions,

    Y . Baiet al., “Residual: Transformer with dual residual connec- tions,”arXiv preprint arXiv:2304.13044, 2023

    2023 arXiv

  24. [32]

    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), 2017, pp. 5220–5224

    2017

  25. [33]

    The small model (ECAPA-S) uses a channel size of 512 and contains 6.19M parameters

    speaker encoder with enhanced channel attention and fea- ture aggregation. The small model (ECAPA-S) uses a channel size of 512 and contains 6.19M parameters. ECAPA-TDNN-L: The large model (ECAPA-L) increases the channel size to 1024 and contains 20.76M parameters. Fol- lowing...

  26. [34]

    DeepCrossAttention: Supercharging transformer residual connections,

    M. Heddes, A. Javanmard, K. Axiotis, G. Fu, M. Bateni, and V . Mirrokni, “DeepCrossAttention: Supercharging transformer residual connections,” inProceedings of the 42nd International Conference on Machine Learning (ICML). PMLR, 2025

    2025

  27. [35]

    Hyper-connections,

    D. Zhu, H. Huang, Z. Huang, Y . Zeng, Y . Mao, B. Wu, Q. Min, and X. Zhou, “Hyper-connections,” inInternational Conference on Learning Representations (ICLR), 2025

    2025

  28. [36]

    Concerning nonnegative matrices and doubly stochastic matrices,

    R. Sinkhorn and P. Knopp, “Concerning nonnegative matrices and doubly stochastic matrices,”Pacific Journal of Mathematics, vol. 21, no. 2, pp. 343–348, 1967

    1967

  29. [37]

    V oxceleb2: Deep speaker recognition,

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

    2018

  30. [38]

    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

  31. [39]

    The Kaldi speech recog- nition toolkit,

    D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motl ´ıˇcek, Y . Qian, P. Schwarz, J. Silovsk´y, G. Stemmer, and K. Vesel´y, “The Kaldi speech recog- nition toolkit,” inProceedings of ASRU 2011. IEEE Signal Pro- cessing Society, 2011, pp. 1–4

    2011

  32. [40]

    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

  33. [41]

    A time delay neural network architecture for efficient modeling of long temporal con- texts,

    V . Peddinti, D. Povey, and S. Khudanpur, “A time delay neural network architecture for efficient modeling of long temporal con- texts,” inInterspeech 2015, 2015, pp. 3214–3218

    2015

  34. [42]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 7132–7141

    2018

  35. [43]

    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 (CVPR), 2019, pp. 4690–4699

    2019

  36. [44]

    3d-speaker-toolkit: An open-source toolkit for multimodal speaker verification and di- arization,

    Y . Chen, S. Zheng, H. Wang, L. Cheng, T. Zhu, R. Huang, C. Deng, Q. Chen, S. Zhang, W. Wanget al., “3d-speaker-toolkit: An open-source toolkit for multimodal speaker verification and di- arization,” inICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Sig...

    2025

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.