Pith. sign in

REVIEW 4 major objections 5 minor 49 references

SEED: Speaker Embedding Enhancement Diffusion Model

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A diffusion model applied directly to speaker embeddings can refine noisy embeddings into clean ones, reducing speaker recognition errors by up to 19.6% under environmental mismatch while preserving conventional-scenario performance.

desk verdict A practical embedding-level denoiser that works on the reported sets, but the diffusion framing is stronger than the mechanism; the missing regression baseline is the one thing to insist on. read the letter →

arxiv 2505.16798 v1 pith:JEET5DUQ submitted 2025-05-22 eess.AS cs.AI

KiHyun Nam , Jungwoo Heo , Jee-weon Jung , Gangin Park , Chaeyoung Jung , Ha-Jin Yu , Joon Son Chung This is my paper · ORCID
classification eess.AScs.AI
keywords speakerrecognitiondiffusionprobabilisticmodelembeddingenhancementenvironmentalmismatchrepresentationDDIMlabel-freetrainingreal-worldrobustness
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

SEED is a diffusion-based post-processor that operates directly on speaker embeddings rather than on raw audio. The paper's central claim is that by training the model to reconstruct clean embeddings from both clean and Gaussian-corrupted noisy embeddings of the same utterance, the same network can refine real noisy embeddings at inference time. This matters because it offers a label-free, model-agnostic way to make existing speaker recognition systems robust to environmental mismatch, with reported gains up to 19.6% in recognition accuracy on mismatched evaluation sets and no loss on conventional ones. The paper positions this as the first application of a diffusion model at the speaker-embedding level.

What carries the argument

The load-bearing object is the cross-reconstruction objective of Eqs. (6) and (9): a sample-prediction diffusion network $f_\theta$ that maps corrupted noisy embeddings $y_t$ back to the clean embedding $x_0$. The network is a small stack of residual fully-connected blocks with timestep embedding, and inference uses a single DDIM step at $t=50$. The multi-pair augmentation that produces noisy variants of each clean utterance, the absence of speaker labels, and the plug-in compatibility with arbitrary encoders all exist to make this reconstruction path stable and transferable.

What would settle it

Measure the average same-utterance clean/noisy embedding distance on a held-out mismatched corpus, such as low-SNR or far-field recordings, and compare it with the Gaussian corruption levels used at training time; if the real gap exceeds the training noise scale, SEED's single-step reconstruction at $t=50$ will not close it and the equal-error-rate gain will vanish.

Watch

Extended reading notes

Core claim

SEED treats a clean speaker embedding $x_0$ and a noisy embedding $y_0$ extracted from the same utterance as two observations of one identity. In the forward process it corrupts both with the same Gaussian noise schedule, producing $x_t$ and $y_t$, and the trainable network $f_\theta$ is asked to predict the clean embedding $x_0$ directly from either corrupted input through the sample-prediction objective $L_{\mathrm{SEED}} = \mathbb{E}[\|x_0 - f_\theta(x_t,t)\| + \sum_{k=0}^{N-1}\|x_0 - f_\theta(y_t^k,t)\|]$. The cross-reconstruction term $y_t \to x_0$ is what makes the model an enhancer rather than a generator: it removes the component that distinguishes the noisy embedding from its clean counterpart, with the same-utterance gap $\|x_0 - y_0\|$ acting as a scaled noise term that the network learns to regress. At inference, every incoming embedding is treated as $y_t$ and regenerated in a single DDIM step at $t=50$. On the two environment-mismatch evaluation sets, SEED reduces equal error rate relative to the baseline recognizers by up to 19.6%, while leaving conventional-scenario performance essentially unchanged.

Load-bearing premise

The method relies on the clean and noisy embeddings of the same utterance being close enough in embedding space that a model trained on Gaussian-corrupted pairs can learn to close the gap for real mismatches.

Editorial extensions

If this is right

  • Any pretrained speaker recognition model can be wrapped by SEED without retraining or speaker labels, gaining robustness to environmental mismatch.
  • The approach retains or slightly improves performance on clean and generalization benchmarks, so it can be deployed in conventional scenarios without sacrificing accuracy.
  • SEED matches or beats a disentangled-representation baseline that requires speaker and environment labels, while training only on unlabeled clean speech.
  • Compared with waveform-domain speech enhancement, SEED needs fewer parameters, less memory, and a lower real-time factor, and it avoids the distortion that enhancement can introduce into speaker-specific information.
  • Because SEED works directly on embeddings, it can be stacked on any upstream encoder, including raw-waveform models, without modifying the encoder.

Reading between the lines

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

  • If the embedding-level correction generalizes beyond speaker recognition, the same recipe could be applied to other embedding-based speech tasks, such as language or emotion recognition, where domain mismatch also widens same-class embedding distances.
  • The paper's single-step $t=50$ inference suggests SEED largely learns a deterministic correction; explicitly modeling the residual domain gap, rather than Gaussian corruption, could extend the gains to extreme mismatches.
  • A natural test is to stack SEED on top of a front-end speech enhancer: waveform enhancement first, then embedding refinement, to see whether the gains compound or saturate.
  • Because SEED needs no labels and trains on ordinary clean speech corpora, scaling it to more data is limited by clean-audio availability rather than by annotation effort.
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. This manuscript proposes SEED, a diffusion-based post-processor for speaker embeddings. Clean and noisy embeddings are extracted from the same utterance by a frozen speaker encoder, corrupted by the DDPM forward process, and a network f_theta is trained with the sample-prediction loss L_SEED of Eq. (9) to reconstruct the clean embedding from both the clean and noisy corrupted embeddings. At inference the raw embedding is passed through f_theta at a fixed timestep t=50 with a single DDIM step, and for WavLM-ECAPA the original and enhanced embeddings are summed. Experiments compare three baselines, a DRL method, and waveform-domain speech enhancement on VoxSRC23, VC-Mix, and VoxCeleb1, reporting EER and minDCF; the headline result is up to 19.6% relative EER reduction on mismatch sets while retaining performance on clean sets.

Significance. The practical goal is worthwhile and the paper is clearly written. The design is label-free, add-on, and computationally light; the evaluation uses external benchmarks; code is promised. The authors compare against prior DRL and an audio-enhancement baseline, and they include computational cost (params, RTF, memory). The reported EER numbers on VC-Mix and VoxSRC23 do support an improvement over the baselines. However, the paper's central attribution of the gains to the diffusion formulation is not yet supported because the inference procedure is a single deterministic regression step and no non-diffusion regression control is reported. The WavLM feature ensemble introduces an unexplained per-model modification. With additional controlled experiments, the approach could be an interesting contribution; at present the evidence supports a weaker claim: L2-based embedding denoising improves robustness.

major comments (4)
  1. [Section 4.2 and Eq. (9)] The inference scheme is a single application of f_theta(., t=50), and the training loss is an L2 sample-reconstruction loss. Nothing in this pipeline requires the diffusion forward process or multi-step sampling: a plain MLP or linear regression trained on the same noisy/clean embedding pairs with the same L2 objective would produce a deterministic map of the same functional form. Since no such non-diffusion control is reported, the 'up to 19.6%' improvement cannot currently be attributed to the diffusion mechanism; the gains may be due to the L2 regression objective. Please add a matched non-diffusion baseline (same architecture without timestep conditioning, and a linear regressor) and an ablation that isolates the role of timestep conditioning.
  2. [Section 4.2 vs. Eq. (4)] At inference the authors state that 'all input speaker embeddings will be y_t' and set t=50, but the network was trained on y_t = sqrt(alpha_t) y_0 + sqrt(1-alpha_t) epsilon with Gaussian epsilon, whereas at inference the raw embedding y_0 is fed directly (or, if not, the scaling/corruption is unspecified). This is a distribution shift between training and inference. Please specify exactly which quantity is input to f_theta at test time. If it is y_0, the model must generalize from Gaussian-corrupted inputs to uncorrupted ones; provide an ablation over t and over corruption at inference. If it is a corrupted version of y_0, describe how the corruption is computed.
  3. [Section 4.2, Tables 1-2] For WavLM-ECAPA, the final embedding is the sum of the original and SEED outputs, while for ResNet34 and ECAPA-TDNN the SEED output alone is used. This feature ensemble is not in the method and not applied to the other models. The WavLM-ECAPA to SEED row therefore measures a different enhancement pipeline. Please report WavLM-ECAPA with SEED output alone, and the same summation ensemble applied to the other baselines, to make the comparison uniform.
  4. [Section 3.3.3 and Eq. (8)] The method rests on the assumption that ||x_0 - y_0|| is moderate, and the conclusion concedes that training instability may occur, but the paper never measures this gap or its distribution. Please report the empirical distribution of ||x_0 - y_0|| for the training augmentations and for the evaluation sets, and show how performance varies with SNR and timestep. This would also test whether the synthetic augmentation covers the real mismatch in VoxSRC23 and VC-Mix.
minor comments (5)
  1. [Eqs. (3)-(4)] The same epsilon is used for x_t and y_t but this is not stated; please make explicit, since the derivation of Eq. (8) depends on it.
  2. [Abstract] The abstract's 'up to 19.6%' should specify that it is a relative EER improvement; otherwise it can be misread as an absolute improvement.
  3. [Table 1] Differences of 0.01-0.03 EER on Vox1-O and Vox1-E are likely within evaluation noise; report confidence intervals or significance tests.
  4. [Section 4.2] The 'scaled linear noise schedule' and the alpha_t values are not fully specified; provide schedule parameters or an exact reference for reproducibility.
  5. [Author list and abstract] Minor typos: 'P ark' and 'Y u' in the author list, and 'inferencing' should be 'inference'.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: SEED's claimed gains are measured on held-out external benchmarks, and the diffusion machinery is not defined in terms of the recognition improvement it reports.

full rationale

The central derivation chain is self-contained. Training objective Eq. (9) is a paired L2 sample-prediction loss between corrupted embeddings (clean x_t and noisy y_t^k) and the clean target x_0; inference is a single-step DDIM-style application f_theta(y_50, 50) of the trained network. There is no step in which the reported EER/minDCF numbers are fed back into the model definition, and no parameter is fitted to the evaluation protocols (VoxSRC23, VC-Mix, VoxCeleb1-O/E/H). The authors' prior DRL work [2] appears only as a comparison baseline needing labels, not as justification of SEED; citations [8], [37], and [38] are background or dataset sources. The 'scaled noise' term in Eqs. (7)-(9) follows from the same-epsilon reparameterization (xt - yt = sqrt(alpha_t)(x0 - y0)) and is an algebraic identity, not a circular definition. The main legitimate weaknesses are empirical: no non-diffusion regression baseline is run, so the contribution of the diffusion formalism over a plain paired regressor is not isolated, and the WavLM-ECAPA feature ensemble is an extra inference modification; these affect evidence strength, not circularity. The conclusion's admission of possible training instability is a stated limitation, not a circular step. Overall circularity burden is low.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The ledger is light: this is an empirical ML paper with no new physical entities. The load-bearing assumptions are about the tractability of the clean/noisy gap and the transferability of synthetic augmentation. The two hand-picked inference choices (t=50 and the WavLM feature ensemble) are the main free parameters that affect the reported results.

free parameters (2)
  • Inference timestep t = 50
    Set for single-step sampling in Section 4.2; the paper notes 'empirically we observed negligible performance differences compared to multi-step sampling', suggesting selection was based on evaluation results, and this parameter controls the noise level the model expects at test time.
  • WavLM feature ensemble weight = 1.0 (equal sum)
    Section 4.2: 'For WavLM-ECAPA, we apply a feature ensemble technique where the original speaker embedding and SEED's output embedding are summed'; applied only to one backbone and appears tuned to improve results on evaluation sets.
assumptions (4)
  • standard math DDPM/DDIM forward and reverse processes (Eqs. 1-6) are valid when applied directly to speaker embeddings
    The paper adopts Ho et al. [9] and Song et al. [10] without modification; embeddings are treated as continuous vectors in R^D and Gaussian corruption is applied in embedding space.
  • domain assumption x_0 and y_0 from the same utterance are close enough for the scaled noise term in Eq. (8) to remain manageable
    Section 3.3.3: 'under our assumption that x_0 and y_0 come from the same utterance of the same speaker, we expect the scaled noise to be moderate.' If real noisy embeddings are far from clean, training destabilizes, as the conclusion concedes.
  • domain assumption Synthetic augmentation (MUSAN noise and simulated RIR) transfers to real environment mismatch
    Training uses only clean speech with added noise/reverberation (Section 4.2); the model is evaluated on VoxSRC23 and VC-Mix which are meant to reflect realistic mismatch. No evidence is provided that the synthetic embedding gap distribution matches real mismatches.
  • ad hoc to paper Single-step inference at t=50 is equivalent to the full reversal for test inputs
    Section 4.2: 'We set the timestep t to 50 for single-step sampling.' This bypasses the diffusion process and is an empirical choice with no derivation that t=50 is optimal.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SEED: Speaker Embedding Enhancement Diffusion Model." pith.science (2026). https://pith.science/paper/JEET5DUQ

@misc{pith2026250516798,
  author       = {Pith},
  title        = {Pith review of: SEED: Speaker Embedding Enhancement Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEET5DUQ}},
  note         = {Machine review of arXiv:2505.16798}
}
read the original abstract

A primary challenge when deploying speaker recognition systems in real-world applications is performance degradation caused by environmental mismatch. We propose a diffusion-based method that takes speaker embeddings extracted from a pre-trained speaker recognition model and generates refined embeddings. For training, our approach progressively adds Gaussian noise to both clean and noisy speaker embeddings extracted from clean and noisy speech, respectively, via forward process of a diffusion model, and then reconstructs them to clean embeddings in the reverse process. While inferencing, all embeddings are regenerated via diffusion process. Our method needs neither speaker label nor any modification to the existing speaker recognition pipeline. Experiments on evaluation sets simulating environment mismatch scenarios show that our method can improve recognition accuracy by up to 19.6% over baseline models while retaining performance on conventional scenarios. We publish our code here https://github.com/kaistmm/seed-pytorch

Figures

Figures reproduced from arXiv: 2505.16798 by the authors.

Figure 1
Figure 1. Illustration of Speaker Embedding Enhancement Diffusion (SEED) model. (a) explains the concept of our diffusion mechanism. (b) shows the whole training process of SEED. (c) illustrates the architecture of SEED. across various domains [11, 12, 13]. These models operate through a forward process that gradually corrupts data with Gaussian noise, followed by a reverse process that reconstructs the original data distribu… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 41 canonical work pages

  1. [27]

    Segan: Speech enhancement generative adversarial network,

    S. Pascual, A. Bonafonte, and J. Serr`a, “Segan: Speech enhancement generative adversarial network,” in Proc. Interspeech, 2017, pp. 3642–3646

    work page 2017

  2. [1]

    Introduction Speaker recognition systems are widely used in various fields such as user authentication, security, and voice interfaces. However, differences in recording equipment, background noise, and other environmental factors in real-world conditions can introduce substan- tial acoustic discrepancies between utterances from the same speaker, degradin...

    work page

  3. [2]

    Related works 2.1. Diffusion Probabilistic Model Diffusion probabilistic models (DPMs) [9, 10] have emerged as a powerful generative framework for high-fidelity data generation arXiv:2505.16798v1 [eess.AS] 22 May 2025 Figure 1:Illustration of Speaker Embedding Enhancement Diffusion (SEED) model. (a) explains the concept of our diffusion mechanism. (b) sho...

    work page Pith review arXiv 2025

  4. [3]

    The concept, process, and model structure of our methodology are shown in Figure 1

    Proposed method This section describes the configuration of our proposed SEED, including the batch setup with audio augmentation and the speaker embedding extraction process. The concept, process, and model structure of our methodology are shown in Figure 1. 3.1. Batch configuration with multi-pair audio augmentation To simulate environment mismatch scena...

    work page

  5. [4]

    Model Configuration 4.1.1

    Experiments 4.1. Model Configuration 4.1.1. Speaker embedding network We employ three pre-trained speaker embedding networks. For the spectrogram-based speaker network, we selected ‘ H / ASP’ version ofResNet34[28] and ECAP A-TDNN[29]. For the raw-waveform-based speaker network, we chose ECAPA-TDNN network combined with ‘WavLM Base+’ version of WavLM [30]...

    work page

  6. [5]

    Results In this section, we summarise and analyse our experimental re- sults. To assess both environment-robustness and generalisation per- formance across diverse speaker recognition systems, we compare three baselines, an adversarial DRL-based method [2] designed to remove environmental factors from the speaker representation, and our proposed method (S...

    work page

  7. [6]

    The SNR is randomly selected within 0–15 dB for background noise and 5–15 dB for music noise

    and music and background noises from MUSAN [5]. The SNR is randomly selected within 0–15 dB for background noise and 5–15 dB for music noise. Consequently, each utterance forms a set of four paired samples (one clean and three noisy variants). T raining & Inference. All of our experiments utilise the PyTorch framework [39] together with the open-source vo...

    work page

  8. [7]

    It seamlessly integrates with pre-trained speaker recognition models, enabling immediate deployment in speaker recognition systems for real-world applications

    Conclusion SEED leverages diffusion models for robust and generalised speaker embedding enhancement without requiring speaker labels or complex training. It seamlessly integrates with pre-trained speaker recognition models, enabling immediate deployment in speaker recognition systems for real-world applications. However, since SEED learns the gap between ...

    work page

Show all 49 references
  1. [8]

    Acknowledgements This work was supported by IITP grant funded by the Korea government (MSIT) (RS-2024- 00457882, National AI Research Lab Project)

    2024

  2. [9]

    Speaker recognition: A tutorial,

    J. P . Campbell, “Speaker recognition: A tutorial,”Proceedings of the IEEE, vol. 85, no. 9, pp. 1437–1462, 1997

    1997

  3. [10]

    Disentangled representation learning for environment-agnostic speaker recognition,

    K. Nam, H.-S. Heo, J. weon Jung, and J. Chung, “Disentangled representation learning for environment-agnostic speaker recognition,” in Proc. Interspeech, 2024, pp. 2130–2134

    2024

  4. [11]

    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

  5. [12]

    V oxCeleb2: Deep Speaker Recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxCeleb2: Deep Speaker Recognition,” inProc. Interspeech, 2018, pp. 1086–1090

    2018

  6. [13]

    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

  7. [14]

    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,” inProc. ICASSP. IEEE, 2017, pp. 5220–5224

    2017

  8. [15]

    Disentangled representation learning,

    X. Wang, H. Chen, S. Tang, Z. Wu, and W. Zhu, “Disentangled representation learning,”arXiv preprint arXiv:2211.11695, 2022

    2022 arXiv

  9. [16]

    Disentangled representation learning for multilingual speaker recognition,

    K. Nam, Y . Kim, J. Huh, H.-S. Heo, J. weon Jung, and J. S. Chung, “Disentangled representation learning for multilingual speaker recognition,” inProc. Interspeech, 2023, pp. 5316–5320

    2023

  10. [17]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  11. [18]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in International Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum?id=St1giarCHLP

    2021

  12. [19]

    Hierarchi- cal text-conditional image generation with clip latents,

    A. Ramesh, P . Dhariwal, A. Nichol, C. Chu, and M. Chen, “Hierarchi- cal text-conditional image generation with clip latents,”arXiv preprint arXiv:2204.06125, vol. 1, no. 2, p. 3, 2022

    2022 arXiv

  13. [20]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P . Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proc. CVPR, 2022, pp. 10 684–10 695

    2022

  14. [21]

    Audioldm: text-to-audio generation with latent diffusion models,

    H. Liu, Z. Chen, Y . Y uan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. D. Plumbley, “Audioldm: text-to-audio generation with latent diffusion models,” inProc. ICML, ser. ICML’23. JMLR.org, 2023

    2023

  15. [22]

    Suppression of acoustic noise in speech using spectral subtraction,

    S. Boll, “Suppression of acoustic noise in speech using spectral subtraction,”IEEE Transactions on acoustics, speech, and signal processing, vol. 27, no. 2, pp. 113–120, 1979

    1979

  16. [23]

    Enhancement and bandwidth compression of noisy speech,

    J. S. Lim and A. V . Oppenheim, “Enhancement and bandwidth compression of noisy speech,” Proceedings of the IEEE, vol. 67, no. 12, pp. 1586–1604, 1979

    1979

  17. [24]

    Speech enhancement using a minimum- mean square error short-time spectral amplitude estimator,

    Y . Ephraim and D. Malah, “Speech enhancement using a minimum- mean square error short-time spectral amplitude estimator,”IEEE Transactions on acoustics, speech, and signal processing, vol. 32, no. 6, pp. 1109–1121, 1984

    1984

  18. [25]

    Speech enhancement based on deep denoising autoencoder

    X. Lu, Y . Tsao, S. Matsuda, and C. Hori, “Speech enhancement based on deep denoising autoencoder.” inProc. Interspeech, vol. 2013, 2013, pp. 436–440

    2013

  19. [26]

    A new framework for cnn-based speech enhancement in the time domain,

    A. Pandey and D. Wang, “A new framework for cnn-based speech enhancement in the time domain,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 27, no. 7, pp. 1179–1188, 2019

    2019

  20. [28]

    A variance modeling framework based on variational autoencoders for speech enhancement,

    S. Leglaive, L. Girin, and R. Horaud, “A variance modeling framework based on variational autoencoders for speech enhancement,” in2018 IEEE 28th international workshop on machine learning for signal processing (MLSP). IEEE, 2018, pp. 1–6

    2018

  21. [29]

    Speech enhancement with score-based generative models in the complex stft domain,

    S. Welker, J. Richter, and T. Gerkmann, “Speech enhancement with score-based generative models in the complex stft domain,” inProc. Interspeech, 2022, pp. 2928–2932

    2022

  22. [30]

    How bad are artifacts?: Analyzing the impact of speech enhancement errors on asr,

    K. Iwamoto, T. Ochiai, M. Delcroix, R. Ikeshita, H. Sato, S. Araki, and S. Katagiri, “How bad are artifacts?: Analyzing the impact of speech enhancement errors on asr,” inProc. Interspeech, 2022, pp. 5418–5422

    2022

  23. [31]

    Audio enhancing with dnn autoencoder for speaker recognition,

    O. Plchot, L. Burget, H. Aronowitz, and P . Matejka, “Audio enhancing with dnn autoencoder for speaker recognition,” in Proc. ICASSP. IEEE, 2016, pp. 5090–5094

    2016

  24. [32]

    Front-end speech enhancement for commercial speaker verification systems,

    S. E. Eskimez, P . Soufleris, Z. Duan, and W. Heinzelman, “Front-end speech enhancement for commercial speaker verification systems,” Speech Communication, vol. 99, pp. 101–113, 2018

    2018

  25. [33]

    Joint feature en- hancement and speaker recognition with multi-objective task-oriented network

    Y . Wu, L. Wang, K. A. Lee, M. Liu, and J. Dang, “Joint feature en- hancement and speaker recognition with multi-objective task-oriented network.” inProc. Interspeech, 2021, pp. 1089–1093

    2021

  26. [34]

    Analysis of i-vector length normalization in speaker recognition systems

    D. Garcia-Romero and C. Y . Espy-Wilson, “Analysis of i-vector length normalization in speaker recognition systems.” inProc. Interspeech, vol. 2011, 2011, pp. 249–252

    2011

  27. [35]

    Selective deep speaker embedding enhancement for speaker verification,

    J.-W . Jung, J.-H. Kim, H.-J. Shim, S. bin Kim, and H.-J. Y u, “Selective deep speaker embedding enhancement for speaker verification,” in The Speaker and Language Recognition W orkshop (Odyssey 2020), 2020, pp. 171–178

    2020

  28. [36]

    The ins and outs of speaker recognition: lessons from V oxSRC 2020,

    Y . Kwon, H.-S. Heo, B.-J. Lee, and J. S. Chung, “The ins and outs of speaker recognition: lessons from V oxSRC 2020,” inProc. ICASSP. IEEE, 2021, pp. 5809–5813

    2020

  29. [37]

    ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA-TDNN: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” inProc. Interspeech, 2020

    2020

  30. [38]

    Wavlm: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen, C. Wang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T. Y oshioka, X. Xiaoet al., “Wavlm: Large-scale self-supervised pre-training for full stack speech processing,”IEEE Journal of Selected T opics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    2022

  31. [39]

    Real time speech enhancement in the waveform domain,

    A. D´efossez, G. Synnaeve, and Y . Adi, “Real time speech enhancement in the waveform domain,” inProc. Interspeech, 2020, pp. 3291–3295

    2020

  32. [40]

    Return of unconditional generation: A self-supervised representation generation method,

    T. Li, D. Katabi, and K. He, “Return of unconditional generation: A self-supervised representation generation method,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  33. [41]

    Layer normalization,

    J. Lei Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,”ArXiv e-prints, pp. arXiv–1607, 2016

    2016

  34. [42]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,

    S. Elfwing, E. Uchibe, and K. Doya, “Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,” Neural networks, vol. 107, pp. 3–11, 2018

    2018

  35. [43]

    Libritts-r: A restored multi-speaker text-to-speech corpus,

    Y . Koizumi, H. Zen, S. Karita, Y . Ding, K. Y atabe, N. Morioka, M. Bacchiani, Y . Zhang, W . Han, and A. Bapna, “Libritts-r: A restored multi-speaker text-to-speech corpus,” inProc. Interspeech, 2023, pp. 5496–5500

    2023

  36. [44]

    Libri-light: A benchmark for asr with limited or no supervision,

    J. Kahn, M. Riviere, W. Zheng, E. Kharitonov, Q. Xu, P .-E. Mazar´e, J. Karadayi, V . Liptchinsky, R. Collobert, C. Fuegenet al., “Libri-light: A benchmark for asr with limited or no supervision,” inProc. ICASSP. IEEE, 2020, pp. 7669–7673

    2020

  37. [45]

    The vox celeb speaker recognition challenge: A retrospective,

    J. Huh, J. S. Chung, A. Nagrani, A. Brown, J.-w. Jung, D. Garcia- Romero, and A. Zisserman, “The vox celeb speaker recognition challenge: A retrospective,”IEEE/ACM Trans. on Audio, Speech, and Language Processing, 2024

    2024

  38. [46]

    Rethinking session variability: Leveraging session embeddings for session robustness in speaker verification,

    H.-S. Heo, K. Nam, B.-J. Lee, Y . Kwon, M. Lee, Y . J. Kim, and J. S. Chung, “Rethinking session variability: Leveraging session embeddings for session robustness in speaker verification,” inProc. ICASSP, 2023

    2023

  39. [47]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga et al., “Pytorch: An imperative style, high-performance deep learning library,” inProc. NeurIPS, vol. 32, 2019

    2019

  40. [48]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in Proc. ICLR, 2019

    2019

  41. [49]

    The 2018 nist speaker recognition evaluation,

    O. Sadjadi, C. Greenberg, E. Singer, D. Reynolds, L. Mason, and J. Hernandez-Cordero, “The 2018 nist speaker recognition evaluation,” in Proc. Interspeech, 2019

    2018

Pith tools

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