Pith. sign in

REVIEW 4 major objections 5 minor 51 references

Memory-Efficient Training for Deep Speaker Embedding Learning in Speaker Verification

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper claims that replacing stored activations with reversible recomputation and compressing optimizer states to 8 bits lets deep speaker embedding extractors train on one or two consumer-grade GPUs with up to 16.2x memory savings…

desk verdict Reversible nets for speaker embedding are a credible idea, but the memory metric is undefined and internally inconsistent, so the 16.2x headline cannot be verified as reported. read the letter →

arxiv 2412.01195 v1 pith:MRVGG5YK submitted 2024-12-02 eess.AS cs.AIcs.SD

Bei Liu , Yanmin Qian This is my paper · ORCID
classification eess.AScs.AIcs.SD
keywords speakerverificationmemory-efficienttrainingreversibleneuralnetworkoptimizerstatequantizationdeepembeddingDF-ResNetVoxCeleb
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that deep speaker embedding extractors—networks like ResNet101, ResNet152, and DF-ResNet233 that normally need multiple 40GB A100 GPUs—can be trained on one or two consumer 11GB 2080Ti GPUs without losing accuracy. The authors argue that two memory costs dominate training: cached activations and optimizer states, and they attack both. Activations are handled by redesigning residual networks so intermediate activations can be recomputed in the backward pass; optimizer states are compressed from 32-bit to a block-wise 8-bit dynamic format. On VoxCeleb the combined recipe reports up to 16.2x memory savings with nearly unchanged equal error rates. If correct, this would make state-of-the-art speaker verification training accessible to researchers without high-end hardware.

What carries the argument

The load-bearing mechanism is the additive-coupling reversible block: input activations are split into $x_1$ and $x_2$, the block computes $y_1 = x_1 + F(x_2)$ and $y_2 = x_2 + G(y_1)$, and during back-propagation $x_2 = y_2 - G(y_1)$ and $x_1 = y_1 - F(x_2)$ recover the inputs, so intermediate activations need not be stored. Type II variants make downsampling reversible by reshaping $C \times F \times T$ features into $r^2 C \times F/r \times T/r$. For optimizer states, the paper uses a dynamic tree-based 8-bit quantization: the state tensor is chunked into blocks of 2048, each block is normalized by its absolute maximum, each value is mapped to the nearest representable 8-bit value by binary search, and only the integer index is stored; the state is dequantized for updates and re-quantized afterwards.

What would settle it

Measure peak GPU memory with a standard memory profiler for vanilla ResNet152 and RevNet197 at a fixed batch size on the same 11GB 2080Ti, and separately compute the float32 momentum-buffer size for a 6.7 million-parameter model; if the 16.2x compression or the roughly 0.02 GB optimizer-state saving does not appear, the paper's headline memory claim is not reproducible.

Watch

Extended reading notes

Core claim

The central claim is that reversible versions of ResNets and DF-ResNets can train deep speaker embedding extractors without caching activations, and that dynamically quantizing optimizer states to 8 bits cuts their memory by 75%, so the two techniques together let models such as DF-ResNet233 (normally four A100 GPUs) train on two 2080Ti GPUs with nearly identical EER. The paper reports that memory per utterance no longer grows with depth, that both Type I (partially reversible) and Type II (fully reversible, with reversible downsampling via tensor reshaping) variants match baseline accuracy, and that the 8-bit SGD and Adam versions add a small extra saving, bringing the largest reported compression to 16.2x.

Load-bearing premise

The quantitative claims rest on the assumption that the paper's 'memory (GB/utter)' measurement is a complete, consistent measure of training memory that actually includes the optimizer-state savings; if that metric is measured in a way that excludes optimizer states or varies with batch size, the reported 16.2x compression and 75% optimizer-state saving cannot be verified.

Editorial extensions

If this is right

  • Models of the size of DF-ResNet233, previously trained on four 40GB A100 GPUs, can be trained on two 11GB 2080Ti GPUs with Vox1-H EER changing from 1.44 to 1.46.
  • Memory per training utterance becomes nearly independent of depth for reversible variants, for example roughly 0.03-0.04 GB per 2-second utterance for RevNets regardless of ResNet depth.
  • 8-bit SGD and 8-bit AdamW each cut optimizer-state memory by 75% while keeping EER differences small (about 0.01-0.03 absolute).
  • For a total batch size of 256, the required number of 11GB GPUs drops from 9-22 for the vanilla models to 1-2 for the reversible variants, with training time increasing from roughly 2-4 days to 3-6 days.
  • Maximum batch size on a single 2080Ti rises by up to 13.4x, for example ResNet152 from 22 to 295 samples.

Reading between the lines

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

  • The same reversible-block and 8-bit-optimizer recipe could plausibly transfer to other deep speech models such as ECAPA-TDNN variants and to fine-tuning large pre-trained models, since the activation-plus-optimizer-state memory profile is generic; the paper does not test this.
  • The reported 'memory (GB/utter)' metric is not defined tightly enough to audit: the 0.001 GB difference between 32-bit and 8-bit optimizer rows is far smaller than the roughly 0.02 GB that a float32 momentum buffer for a 6.7 million-parameter model would occupy, so the 75% optimizer-state saving is not visible in the headline per-utterance numbers.
  • Because Type II reversible downsampling is a lossless reshape, it may also serve as a general drop-in operation for any convolutional speech or audio backbone that wants activation memory independent of depth.
  • The 16.2x figure combines activation elimination with optimizer-state compression; a user adopting only the reversible networks would see up to 15.7x, so the incremental gain of 8-bit optimizers on top is small when activations dominate.
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 / 5 minor

Summary. The paper proposes memory-efficient training methods for deep speaker embedding extractors in speaker verification. It first analyzes GPU memory allocation and identifies activations and optimizer states as the main consumers. It then designs two families of reversible neural networks (RevNets and DF-RevNets) that avoid storing intermediate activations during back-propagation, and introduces 8-bit dynamic quantization for SGD and AdamW optimizer states. Experiments on VoxCeleb1/VoxCeleb2 report comparable EER between the reversible variants and the original ResNets and DF-ResNets, with claimed memory savings up to 16.2x and the ability to train deep models on one or two consumer 2080Ti GPUs instead of multiple high-end GPUs.

Significance. If the quantitative memory claims are substantiated, the paper would make deep speaker embedding training substantially more accessible by removing the need for multiple A100/V100 GPUs, while preserving verification performance. The reversible architecture formulation is clearly specified and the EER comparisons across ResNet34/101/152 and DF-ResNet56/110/179/233 are extensive and largely consistent. The 8-bit optimizer-state quantization follows an established dynamic tree approach and is in principle sound. However, the headline memory savings rest on an undefined and internally inconsistent per-utterance memory metric, so the central quantitative claim is currently unverifiable. The paper does not provide machine-checked proofs or code, but the architectural design and experimental comparisons are reproducible in principle.

major comments (4)
  1. [Section VI-A, Tables V and VI] The metric "Memory (GB/utter)" is never defined. The reported values are inconsistent with the stated GPU capacity: DF-ResNet233 is listed at 1.034 GB/utter with maximum batch size 12, which implies 12.4 GB total, exceeding an 11 GB 2080Ti; DF-ResNet110 implies 0.523×23=12.0 GB, DF-ResNet179 implies 0.858×14=12.0 GB, and DF-ResNet56 implies 0.355×33=11.7 GB. If the value is total memory at maximum batch divided by batch size, these rows are impossible. If it is a per-sample marginal measurement at batch size 1, the value mixes framework, weight, and optimizer overhead with activation memory, and the cross-model compression ratios do not measure trainability as claimed. This is load-bearing for the headline 16.2x memory-savings claim; the authors must provide an explicit measurement protocol (which allocations are counted, batch size, whether optimizer states and gradient buffers are included) and report values that are internally consistent.
  2. [Section IV-B and VI-B] The claimed 75% optimizer-state memory saving is not reflected in the per-utterance memory numbers. For a 12.3M-parameter model, AdamW's two float32 states occupy about 12.3M×8 bytes ≈ 98 MB, and 8-bit quantization should save roughly 74 MB (0.074 GB). Yet the reported difference between DF-RevNet354 and DF-RevNet354+AdamW(8-bit) is only 0.001 GB/utter, and similar tiny differences appear throughout Tables V and VI. The reason appears to be that the memory metric is an aggregate that does not break out optimizer states, so the 75% saving is not observable in the reported results. Please report activation/optimizer/weight/other memory allocations separately for at least one reversible model with and without 8-bit optimizer states, so that both the 75% claim and the 16.2x claim can be independently verified.
  3. [Section III-B, Eqs. (7)-(8), Algorithm 1] The treatment of batch normalization inside the reversible residual functions F and G is not specified. Batch normalization uses batch statistics computed during the forward pass, and the paper does not state whether those statistics are cached (which consumes memory) or recomputed during the backward pass. If they are recomputed, the exact recovery of the inputs to F and G must be shown to suffice for reproducing the batch statistics; if they are cached, the memory claim needs to account for that storage. This issue is central to the assertion that intermediate activations need not be stored, and it deserves either a concrete implementation explanation or an ablation demonstrating the memory/performance impact of the chosen BN handling.
  4. [Tables I and V] The reported per-utterance memory for ResNet34 is inconsistent across the two tables. Table I reports 5.14 GB of activations and 0.455 GB of other memory at batch size 64, which implies a total of about 5.65 GB, or 0.088 GB/utter, while Table V lists ResNet34 at 0.06 GB/utter. The discrepancy may originate from different counting conventions, but it underscores the need for a single, defined memory metric used consistently across all tables and figures. Without this consistency, the reader cannot determine whether the memory reductions come from the proposed methods or from a change in how memory is counted.
minor comments (5)
  1. [Abstract and Section IV] The wording "8-bit versions of SGD and Adam save 75% of memory costs" describes a theoretical reduction in optimizer-state size, not an observed reduction in total training memory; please clarify this distinction in the abstract and contribution list.
  2. [Section VI-B] Several EER differences between the 32-bit and 8-bit optimizer runs are in the range of 0.01–0.03 (e.g., RevNet46 Vox1-O 0.85 vs. 0.87; DF-RevNet66 Vox1-O 0.84 vs. 0.89). Since no variance or repeated runs are reported, the statement that 8-bit optimizers "maintain model performance" should be softened to "no consistent degradation beyond run-to-run variability," unless error bars are provided.
  3. [Section IV-B.1] The sentence "the tensor is flattened into a one-dimensional sequence and then segmented into B blocks" should read "segmented into blocks of size B," since B is the block size, not the number of blocks.
  4. [Eq. (14)] The argmin expression has a misplaced subscript and the notation n is not defined; the quantization mapping is described as "8-bit" but the formula uses a generic n-bit integer range. Please define n and fix the equation formatting.
  5. [Table I and Fig. 1] The "Others" category is described as framework overhead independent of network architecture and input data, but no evidence is provided for that independence. Please specify what the category includes (e.g., CUDA context, workspace, data loader buffers) and how it was measured.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: memory and EER claims are measured against external VoxCeleb baselines; the explicit self-citation to prior work [41] is transparent and non-load-bearing.

full rationale

The paper's central claims are empirical, not derived from their own outputs. The activation-memory saving follows from the reversible-block equations (Eq. 7-8 and Alg. 1): activations are recomputed in the backward pass rather than stored, and the memory reductions are measured on a 2080Ti against standard ResNet/DF-ResNet baselines. The 75% optimizer-state saving is the arithmetic consequence of replacing 32-bit storage with 8-bit storage (Section IV), and the reported per-utterance changes are small because optimizer states are only ~0.6-0.7% of total memory (Table I). EER results are benchmarked on VoxCeleb1-O/E/H against the original systems, so nothing is fitted and renamed as a prediction. The only direct self-citation, [41] ('This is an extended work of our previous paper'), is declared and does not carry the argument: the reversible backpropagation principle is attributed to [43] and dynamic tree quantization to [44], and the DF-ResNet baselines are previously published architectures used as comparison systems. The undefined 'Memory (GB/utter)' metric and the apparent inconsistency between per-utterance memory and maximum batch sizes in Tables V-VI are correctness/verifiability concerns, not circularity: the numbers are presented as measurements, and no equation or definition is constructed in terms of the claim it supports. Thus no circular step can be exhibited.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

No fundamentally new entities are introduced; the 8-bit data type is adopted from [44] and reversible coupling from [43]. The main assumptions are architectural invertibility, BN compatibility, and the chosen hyperparameters.

free parameters (2)
  • Quantization block size B = 2048
    Chosen by hand in Section IV-B1 for 8-bit optimizer state quantization; not swept, and quantization error depends on it.
  • Reversible network channel widths and block counts per stage = C=[48,96,...] and B per architecture in Tables II, III, IV; e.g., C=[48,96,192,384]
    Hand-selected to keep parameter counts close to baseline ResNets and DF-ResNets; these choices affect the memory/performance comparison that underlies the claim.
assumptions (5)
  • standard math Standard backpropagation with stored activations uses memory linear in depth.
    Section II-A reviews the algorithm; the memory analysis relies on this.
  • domain assumption Residual blocks with stride-1 convolutions are exactly invertible under additive coupling (Eqs. 7 and 8).
    Section III-B1 states reversibility holds only when F/G use stride 1; all reversible blocks in the tables are designed this way and irreversible downsampling is kept or made invertible.
  • ad hoc to paper Batch normalization inside reversible F/G blocks remains compatible with exact or sufficiently accurate input recovery during backward passes.
    The paper does not discuss how BN batch statistics are handled in Algorithm 1; exact inversion during training is assumed without proof or experiment.
  • domain assumption The dynamic tree-based 8-bit representation covers [-1,1] with enough precision that quantized optimizer states do not change convergence.
    Section IV-B2 normalizes each block by its max absolute value and relies on the prior tree quantization format; performance parity is empirical only.
  • domain assumption The VoxCeleb training and evaluation protocol (200-frame chunks, AAM-Softmax, AS-Norm) is representative for speaker verification.
    Section V sets this protocol; all EER conclusions depend on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Memory-Efficient Training for Deep Speaker Embedding Learning in Speaker Verification." pith.science (2026). https://pith.science/paper/MRVGG5YK

@misc{pith2026241201195,
  author       = {Pith},
  title        = {Pith review of: Memory-Efficient Training for Deep Speaker Embedding Learning in Speaker Verification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MRVGG5YK}},
  note         = {Machine review of arXiv:2412.01195}
}
read the original abstract

Recent speaker verification (SV) systems have shown a trend toward adopting deeper speaker embedding extractors. Although deeper and larger neural networks can significantly improve performance, their substantial memory requirements hinder training on consumer GPUs. In this paper, we explore a memory-efficient training strategy for deep speaker embedding learning in resource-constrained scenarios. Firstly, we conduct a systematic analysis of GPU memory allocation during SV system training. Empirical observations show that activations and optimizer states are the main sources of memory consumption. For activations, we design two types of reversible neural networks which eliminate the need to store intermediate activations during back-propagation, thereby significantly reducing memory usage without performance loss. For optimizer states, we introduce a dynamic quantization approach that replaces the original 32-bit floating-point values with a dynamic tree-based 8-bit data type. Experimental results on VoxCeleb demonstrate that the reversible variants of ResNets and DF-ResNets can perform training without the need to cache activations in GPU memory. In addition, the 8-bit versions of SGD and Adam save 75% of memory costs while maintaining performance compared to their 32-bit counterparts. Finally, a detailed comparison of memory usage and performance indicates that our proposed models achieve up to 16.2x memory savings, with nearly identical parameters and performance compared to the vanilla systems. In contrast to the previous need for multiple high-end GPUs such as the A100, we can effectively train deep speaker embedding extractors with just one or two consumer-level 2080Ti GPUs.

Figures

Figures reproduced from arXiv: 2412.01195 by the authors.

Figure 1
Figure 1. GPU memory allocation during the training process for [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The comparison between reversible and non-reversible [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. illustration of 8-bit dynamic quantization and dequan [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: 8-bit dynamic tree quantization data type. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: GPU Memory Usage vs. Parameter Number for ResNets and DF-ResNets. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Maximum Batch Size vs. Parameter Number for ResNets and DF-ResNets. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 43 canonical work pages

  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, 2011

    work page 2011

  2. [2]

    Probabilistic linear discriminant analysis,

    S. Ioffe, “Probabilistic linear discriminant analysis,” in European Con- ference on Computer Vision (ECCV) , 2006, pp. 531–542

    work page 2006

  3. [3]

    Deep feature for text-dependent speaker verification,

    Y . Liu, Y . Qian, N. Chen, T. Fu, Y . Zhang, and K. Yu, “Deep feature for text-dependent speaker verification,”Speech Communication, vol. 73, pp. 1–13, 2015

    work page 2015

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

    work page 2018

  5. [5]

    Deep neural network embeddings for text-independent speaker verification,

    D. Snyder, D. Garcia-Romero, D. Povey, and S. Khudanpur, “Deep neural network embeddings for text-independent speaker verification,” in Proc. Interspeech, 2017, pp. 999–1003

    work page 2017

  6. [6]

    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,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2019, pp. 5796–5800

    work page 2019

  7. [7]

    But system description to voxceleb speaker recognition challenge 2019,

    H. Zeinali, S. Wang, A. Silnova, P. Mat ˇejka, and O. Plchot, “But system description to voxceleb speaker recognition challenge 2019,” arXiv preprint arXiv:1910.12592 , 2019. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

    arXiv 2019

  8. [8]

    Densely connected time delay neural network for speaker verification,

    Y . Yu and W. Li, “Densely connected time delay neural network for speaker verification,” in Proc. Interspeech, 2020, pp. 921–925

    work page 2020

Show all 51 references
  1. [9]

    Ecapa-tdnn: Em- phasized channel attention, propagation and aggregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Em- phasized channel attention, propagation and aggregation in tdnn based speaker verification,” in Proc. Interspeech, 2020, pp. 3830–3834

    2020

  2. [10]

    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,” in Proc. Interspeech, 2023

    2023

  3. [11]

    An effective deep embedding learning method based on dense-residual networks for speaker verification,

    Y . Liu, Y . Song, I. McLoughlin, L. Liu, and L. Dai, “An effective deep embedding learning method based on dense-residual networks for speaker verification,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2021, pp. 6668–6672

    2021

  4. [12]

    Self-knowledge distillation via feature enhancement for speaker verification,

    B. Liu, H. Wang, Z. Chen, S. Wang, and Y . Qian, “Self-knowledge distillation via feature enhancement for speaker verification,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 7542–7546

    2022

  5. [13]

    Adaptive convolutional neural network for text-independent speaker recognition,

    S.-H. Kim and Y .-H. Park, “Adaptive convolutional neural network for text-independent speaker recognition,” in Proc. Interspeech, 2021, pp. 66–70

    2021

  6. [14]

    Df-resnet: Boosting speaker verification performance with depth-first design,

    B. Liu, Z. Chen, S. Wang, H. Wang, B. Han, and Y . Qian, “Df-resnet: Boosting speaker verification performance with depth-first design,” in Proc. Interspeech, 2022, pp. 296–300

    2022

  7. [15]

    Depth-first neural architecture with atten- tive feature fusion for efficient speaker verification,

    B. Liu, Z. Chen, and Y . Qian, “Depth-first neural architecture with atten- tive feature fusion for efficient speaker verification,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 31, pp. 1825–1838, 2023

    2023

  8. [16]

    Mlp-svnet: A multi-layer per- ceptrons based network for speaker verification,

    B. Han, Z. Chen, B. Liu, and Y . Qian, “Mlp-svnet: A multi-layer per- ceptrons based network for speaker verification,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 7522–7526

    2022

  9. [17]

    Attentive statistics pooling for deep speaker embedding,

    K. Okabe, T. Koshinaka, and K. Shinoda, “Attentive statistics pooling for deep speaker embedding,” in Proc. Interspeech, 2018, pp. 2252–2256

    2018

  10. [18]

    Self-attentive speaker embeddings for text-independent speaker verification,

    Y . Zhu, T. Ko, D. Snyder, B. Mak, and D. Povey, “Self-attentive speaker embeddings for text-independent speaker verification,” in Proc. Interspeech, 2018, pp. 3573–3577

    2018

  11. [19]

    Self multi-head attention for speaker recognition,

    M. India, P. Safari, and J. Hernando, “Self multi-head attention for speaker recognition,” in Proc. Interspeech, 2019, pp. 4305–4309

    2019

  12. [20]

    Improving aggregation and loss function for better embedding learning in end-to- end speaker verification system,

    Z. Gao, Y . Song, I. McLoughlin, P. Li, Y . Jiang, and L. Dai, “Improving aggregation and loss function for better embedding learning in end-to- end speaker verification system,” in Proc. Interspeech, 2019, pp. 361– 365

    2019

  13. [21]

    Revisiting the statistics pooling layer in deep speaker embedding learning,

    S. Wang, Y . Yang, Y . Qian, and K. Yu, “Revisiting the statistics pooling layer in deep speaker embedding learning,” in International Symposium on Chinese Spoken Language Processing (ISCSLP) , 2021, pp. 1–5

    2021

  14. [22]

    End-to-end text-independent speaker verifi- cation with triplet loss on short utterances

    C. Zhang and K. Koishida, “End-to-end text-independent speaker verifi- cation with triplet loss on short utterances.” in Proc. Interspeech, 2017, pp. 1487–1491

    2017

  15. [24]

    Margin matters: Towards more discriminative deep neural network embeddings for speaker recognition,

    X. Xiang, S. Wang, H. Huang, Y . Qian, and K. Yu, “Margin matters: Towards more discriminative deep neural network embeddings for speaker recognition,” arXiv preprint arXiv:1906.07317 , 2019

    1906 arXiv

  16. [25]

    Real additive margin softmax for speaker verification,

    L. Li, R. Nai, and D. Wang, “Real additive margin softmax for speaker verification,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2022, pp. 7527–7531

    2022

  17. [26]

    Deep residual learning for image recognition,

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

    2016

  18. [27]

    Attentive feature fusion for robust speaker verification,

    B. Liu, Z. Chen, and Y . Qian, “Attentive feature fusion for robust speaker verification,” in Proc. Interspeech, 2022, pp. 286–290

    2022

  19. [28]

    Simple attention module based speaker verification with iterative noisy label detection,

    X. Qin, N. Li, C. Weng, D. Su, and M. Li, “Simple attention module based speaker verification with iterative noisy label detection,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 6722–6726

    2022

  20. [29]

    Dual path embedding learning for speaker verification with triplet attention,

    B. Liu, Z. Chen, and Y . Qian, “Dual path embedding learning for speaker verification with triplet attention,” in Proc. Interspeech, 2022, pp. 291– 295

    2022

  21. [30]

    Repvgg: Making vgg-style convnets great again,

    X. Ding, X. Zhang, N. Ma, J. Han, G. Ding, and S. Jian, “Repvgg: Making vgg-style convnets great again,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2021, pp. 13 733–13 742

    2021

  22. [31]

    The sjtu x-lance lab system for cnsrc 2022,

    Z. Chen, B. Liu, B. Han, L. Zhang, and Y . Qian, “The sjtu x-lance lab system for cnsrc 2022,” arXiv preprint arXiv:2206.11699 , 2022

    2022 arXiv

  23. [32]

    Sjtu- aispeech system for voxceleb speaker recognition challenge 2022,

    Z. Chen, B. Han, X. Xiang, H. Huang, B. Liu, and Y . Qian, “Sjtu- aispeech system for voxceleb speaker recognition challenge 2022,”arXiv preprint arXiv:2209.09076, 2022

    2022 arXiv

  24. [33]

    Unisound system for voxceleb speaker recognition challenge 2023,

    Y . Zheng, Y . Zhang, C. Niu, Y . Zhan, Y . Long, and D. Xu, “Unisound system for voxceleb speaker recognition challenge 2023,” arXiv preprint arXiv:2308.12526, 2023

    2023 arXiv

  25. [34]

    Learning representations by back-propagating errors,

    D. Rumelhart, G. Hinton, and R. Williams, “Learning representations by back-propagating errors,” Nature, vol. 323, no. 6088, pp. 533–536, 1986

    1986

  26. [35]

    On the momentum term in gradient descent learning algo- rithms,

    N. Qian, “On the momentum term in gradient descent learning algo- rithms,” Neural Networks, vol. 12, no. 1, pp. 145–151, 1999

    1999

  27. [36]

    Adam: A method for stochastic optimization,

    D. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in International Conference on Learning Representations , 2015

    2015

  28. [37]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in International Conference on Learning Representations , 2019

    2019

  29. [38]

    Tensorflow: A system for large-scale machine learning,

    M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, and et al., “Tensorflow: A system for large-scale machine learning,” in Operating Systems Design and Implementation (OSDI) , 2016, pp. 265– 283

    2016

  30. [39]

    Automatic differentiation in pytorch,

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer, “Automatic differentiation in pytorch,” in Advances in Neural Information Processing Systems (NIPS) Autodiff Workshop, 2017

    2017

  31. [40]

    Ai and memory wall,

    A. Gholami, Z. Yao, S. Kim, C. Hooper, M. Mahoney, and K. Keutzer, “Ai and memory wall,” arXiv preprint arXiv:2403.14123 , 2024

    2024 arXiv

  32. [41]

    Reversible neural networks for memory-efficient speaker verification,

    B. Liu and Y . Qian, “Reversible neural networks for memory-efficient speaker verification,” in Proc. Interspeech, 2023, pp. 3127–3131

    2023

  33. [42]

    Compiling machine learning programs via high-level tracing,

    R. Frostig, M. Johnson, and C. Leary, “Compiling machine learning programs via high-level tracing,” in Machine Learning and Systems (MLSys), 2018

    2018

  34. [43]

    The reversible residual network: Backpropagation without storing activations,

    A. N. Gomez, M. Ren, R. Urtasun, and R. B. Grosse, “The reversible residual network: Backpropagation without storing activations,” in Ad- vances in Neural Information Processing Systems (NIPS) , 2017, pp. 2211–2221

    2017

  35. [44]

    8-bit approximations for parallelism in deep learning,

    T. Dettmers, “8-bit approximations for parallelism in deep learning,” in International Conference on Learning Representations , 2016

    2016

  36. [45]

    V oxceleb: A large-scale speaker identification dataset,

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

    2017

  37. [46]

    V oxceleb2: Deep speaker recognition,

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

    2018

  38. [47]

    The dku-dukeece systems for voxceleb speaker recognition challenge 2020,

    W. Wang, D. Cai, X. Qin, and M. Li, “The dku-dukeece systems for voxceleb speaker recognition challenge 2020,” arXiv preprint arXiv:2010.12731, 2020

    2020 arXiv

  39. [48]

    On-the-fly data loader and utterance-level aggregation for speaker and language recognition,

    W. Cai, J. Chen, J. Zhang, and M. Li, “On-the-fly data loader and utterance-level aggregation for speaker and language recognition,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 28, pp. 1038–1051, 2020

    2020

  40. [49]

    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

  41. [50]

    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,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2017, pp. 5220–5224

    2017

  42. [51]

    Large margin softmax loss for speaker verification,

    Y . Liu, L. He, and J. Liu, “Large margin softmax loss for speaker verification,” in Proc. Interspeech, 2019, pp. 2873–2877

    2019

  43. [52]

    Towards reduced false- alarms using cohorts,

    Z. N. Karam, W. M. Campbell, and N. Dehak, “Towards reduced false- alarms using cohorts,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2011, pp. 4512–4515

    2011

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.