Pith. sign in

REVIEW 5 major objections 5 minor 34 references

MDD: a Mask Diffusion Detector to Protect Speaker Verification Systems from Adversarial Perturbations

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper proposes MDD, a text-conditioned masked diffusion model trained only on clean speech, and shows it detects and purifies adversarial perturbations for speaker verification, cutting PGD attack EER from 73.2% to 18.0% while keeping c

desk verdict A genuine clean-only masked-diffusion defense with strong reported numbers, but the missing vocoder-only control undercuts the mechanistic claim and the abstract overstates the purification result. read the letter →

arxiv 2508.19180 v1 pith:GHDF7PGN submitted 2025-08-26 eess.AS cs.SD

classification eess.AScs.SD
keywords speakerverificationadversarialperturbationsdiffusionmodelspectralmaskingpurificationattackdetectionPGDequalerrorrate
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

MDD is a defence against adversarial audio for speaker verification. The paper's central claim is that a diffusion model trained solely on clean speech, with partial masking of Mel-spectrograms and a text transcription as conditioning, can both detect and remove adversarial perturbations, without ever seeing an attack during training. Detection works by comparing the ASV score on the original input with the score after diffusion purification: clean inputs yield similar scores, attacked inputs yield large differences. The 10% masking variant detects 98% of PGD attacks at a 0.1 false-positive rate and, when used as a purifier, cuts the adversarial equal error rate from 73.2% to 18.0% at a cost of raising the clean EER from 1.4% to 4.0%. The result matters because it suggests an attacker-independent, plug-in defence is possible for deployed verifiers.

What carries the argument

The load-bearing mechanism is the composite noise target Ntarget = xm + σ·ε formed from a partially masked spectrogram, combined with the clean input by xt = √ᾱt·x0 + √(1−ᾱt)·Ntarget. This steers the diffusion process toward a masked, noised state rather than pure Gaussian noise, so the reverse model must learn to fill in missing spectral patches while denoising. The text condition, an embedding of the ASR transcript, anchors the reconstruction to the utterance's linguistic content, and the backend detector converts reconstruction stability into a detection statistic: the difference between original and purified ASV scores. The masking ratio is the controlled variable; 10% masking is the o

What would settle it

Run MDD against adversarial examples generated with a different gradient attack (FGSM or BIM) or an adaptive attack that back-propagates through the full purification and detector: if Detection Rate at FPR=0.1 drops well below 98%, or an adversary can push clean score differences across the threshold, the attacker-independent claim is refuted. A second concrete check: feed attacked audio whose ASR transcript is wrong; if detection or purification collapses when the conditioning text is corrupt, the method's dependence on transcription is exposed.

Watch

Extended reading notes

Core claim

The paper introduces the Mask Diffusion Detector (MDD), a text-conditioned masked diffusion model. Its forward process mixes a randomly masked Mel-spectrogram with Gaussian noise according to a DDPM-style schedule, and the reverse process reconstructs the clean spectrogram conditioned on an ASR-derived transcription; a neural vocoder turns the reconstruction back into a waveform. A frozen ASV system scores both original and purified audio, and their absolute score difference is thresholded to label the input adversarial. Trained only on the clean LibriSpeech train-clean-100 subset and tested on PGD attacks against a VoxCeleb1 ECAPA-TDNN verifier, the 10% masking configuration achieves 98.0%

Load-bearing premise

The whole scheme rests on the reverse diffusion process, trained only on clean speech, actually removing adversarial perturbations from never-seen attacked inputs — and on the ASR-derived transcription staying reliable on attacked audio; if either fails, the purification and the score-difference detector both degrade.

Editorial extensions

If this is right

  • A single MDD trained on clean speech can be inserted in front of a frozen ASV system, providing detection and purification without retraining the verifier.
  • Detection thresholds can be set to a desired false-positive rate using only bona fide data, with no attack examples needed for calibration.
  • Moderate spectral masking acts as an effective regularizer: 10% masking beats both no masking and heavy masking for detecting attacks.
  • Purification improves PGD attack EER roughly fourfold, from 73.2% to 18.0%, while clean EER rises only from 1.4% to 4.0%.
  • Under matched limited training, MDD outperforms both an earlier diffusion-based purifier and neural codec-based detectors.

Reading between the lines

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

  • Not tested in the paper: an attack that targets the conditioning ASR itself. If a perturbation changes the transcription, the reverse process may faithfully reconstruct the wrong utterance, which would sever the speaker-identity link and degrade purification.
  • The fixed 10% mask is chosen on one attack setup; an adaptive masking ratio, or a mask roughly aligned with the perturbed regions, could generalize to stronger or white-box attacks.
  • The detector's threshold is calibrated from clean score differences on one ASV system; deploying with a different verifier, microphone, or domain will likely require recalibration and may change the clean/attack separation.
  • The 4.0% clean EER cost marks a practical trade-off; deployments needing near-original clean accuracy would need a purification objective that preserves more speaker-discriminative detail.
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

5 major / 5 minor

Summary. The paper proposes MDD, a text-conditioned masked diffusion model trained on clean LibriSpeech audio, intended to detect and purify adversarial perturbations in speaker verification. During inference, the input spectrogram is masked/noised, reconstructed by the diffusion model conditioned on a whisper-small transcription, and resynthesized by a HiFi-GAN vocoder. Detection is performed by thresholding the absolute difference between ASV scores computed on the original and purified waveforms, with the threshold calibrated on clean data to a target FPR. Experiments on 1,000 PGD-attacked VoxCeleb1 utterances report DR of 98.0% at FPR=0.1 for the 10% masking configuration, and purification reduces the PGD EER from 73.2% to 18.0%, at the cost of raising the clean EER from 1.4% to 4.0%. The paper compares against DAP and three neural codec baselines and claims state-of-the-art detection and purification performance without adversarial training or large-scale pretraining.

Significance. If the proposed mechanism were rigorously established, MDD would be an interesting contribution: it trains only on clean data, uses a clean-only threshold calibration, and reports competitive detection rates against several neural codec baselines. The clean-training property and the simple score-difference detector are attractive for practical deployment. However, the current evidence does not yet support the central claim that the masked diffusion model is the component responsible for removing adversarial perturbations, because the purification pipeline always ends with a lossy HiFi-GAN vocoder and no vocoder-only control is reported. The abstract's 'close to clean' claim is contradicted by the paper's own Table III. In addition, the hyper-parameter selection and the single-attack evaluation limit the generality of the reported gains. These issues are load-bearing and require additional experiments before the main claims can be accepted.

major comments (5)
  1. [III-B, V, Table III] The purification pipeline always terminates with a pretrained HiFi-GAN vocoder (Section III-B). The paper provides no control condition in which the same vocoder is applied to the input spectrogram without the diffusion model. Since HiFi-GAN is a lossy neural resynthesizer, it may attenuate small adversarial perturbations independently of the learned denoiser. The clean EER degradation from 1.4% to 4.0% in Table III shows the pipeline modifies clean inputs materially. Without a vocoder-only baseline (and, if feasible, a diffusion-only reconstruction), the attribution of the PGD EER reduction from 73.2% to 18.0% to the masked diffusion mechanism is not established.
  2. [Abstract, V, Table III] The abstract claims purification 'restoring speaker verification performance to levels close to those observed under clean conditions'. Table III contradicts this: with 10% MDD, the clean EER is 4.0% vs 1.4% without purification, and the PGD EER is 18.0% vs 1.4% clean. A fourfold-to-thirteenfold degradation is not 'close'. Please temper the claim and report the precise trade-off.
  3. [IV-A, IV-B, V] The key hyper-parameter (mask ratio 10%) and the noise control factor sigma=0.1 are selected after observing test performance on the same 1,000-utterance VoxCeleb1 subset. There is no held-out validation or error bar; the detection and purification results may overfit to this particular subset and attack. Please provide confidence intervals or multiple VoxCeleb1 subsets, and select the mask ratio on a development set disjoint from the test trials.
  4. [IV-A, IV-C] All adversarial results are obtained with a single PGD attack (50 iterations, l2 norm). The paper claims 'robustness and generalisation across attack types' and 'attacker-independent' properties, but no other attack (e.g., FGSM, BIM, or an adaptive attack) is evaluated. This is a load-bearing gap for the generalization claim. At minimum, evaluate two additional attack types and report the same DR/EER metrics.
  5. [III-A.2, IV-A] The reverse process is conditioned on transcriptions from whisper-small, but the reliability of whisper-small on attacked audio is not verified. If ASR fails or changes the transcription, the conditioning could inject wrong content and destroy speaker identity. Please measure ASR accuracy (e.g., WER) on the PGD inputs, or show that purification is robust to conditioning errors.
minor comments (5)
  1. [II-C / Table II] The DAP method is referenced as [15] in the text but as [14] in Table II. Check the cross-references.
  2. [Abstract, I] The paper claims MDD 'does not require large-scale pretraining', but the pipeline uses whisper-small, Stella, and an AudioLDM HiFi-GAN vocoder, all of which are large pretrained models. Clarify that only the diffusion model itself is trained from scratch.
  3. [III-C, Eq. (3)] The minimization in Eq. (3) is unusual as written. State explicitly that tau_det is the smallest threshold satisfying the empirical FPR bound, and describe how the minimization is performed.
  4. [V, Table III] The trial types 'target vs. adversarial non-target' are not defined precisely. Clarify whether the adversarial examples are added to impostor utterances and whether the ASV score is computed against the claimed target enrollment.
  5. [IV-A] No code or model release is mentioned. For a defense paper with several modular components, providing reproducible code or at least exact configuration details would strengthen the contribution.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: MDD's clean-only training and out-of-sample PGD evaluation make the detection/purification result empirical; self-citations [14-16] are context, not load-bearing.

full rationale

The core detection claim is not circular. The diffusion model is trained only on clean LibriSpeech speech, while detection rate is measured on PGD-adversarial VoxCeleb1 samples; the detection threshold (Eq. 3) is calibrated on clean score differences to a target FPR, so the reported DR figures are out-of-sample operating points rather than fitted predictions. Similarly, the purification results in Table III compare unprotected versus MDD-processed PGD trials in a white-box setting, with no adversarial examples used in training, so the EER reduction is an empirical result rather than a construction. Equations (1)-(2) define the forward corruption and the reverse model learns to invert that corruption, which is a standard denoising objective and is not equivalent to the downstream detection metric. The paper's self-citations [14-16] are used only as prior work and baseline context; no uniqueness theorem, fitted ansatz, or load-bearing premise is imported from them. The remaining concerns - tuning sigma and mask ratio on the reported test outcomes, and the absence of a HiFi-GAN-only control - are experimental-design or selection-bias issues, not derivational circularity. Score 2 reflects the presence of minor, non-load-bearing self-citations; no circular step was identified.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central result rests on transfer from clean-only training to adversarial inputs, and on the reliability of ASR text conditioning and the score-difference detector. The numeric degrees of freedom are the mask ratio, sigma, and threshold; no new entities are postulated.

free parameters (3)
  • Mask ratio = 10%
    Selected post hoc from Table I as best on the PGD test set; no held-out validation.
  • Noise control factor sigma = 0.1
    Hand-tuned in Section IV-A for a good balance between attack mitigation and fidelity; no grid or validation protocol.
  • Detection threshold tau_det = Empirically optimized per Eq. (3)
    Set on a clean-only set T to meet target FPR; standard calibration, but the clean set is not precisely specified.
assumptions (4)
  • domain assumption A diffusion model trained on clean speech projects adversarial inputs onto the clean-speech manifold.
    Section III-A.2: 'By learning the distribution of clean data, MDD is able to remove adversarial perturbations from masked inputs.' Unproven generalization claim that is the basis of purification.
  • domain assumption Text-conditioning from an ASR transcription preserves speaker-discriminative information during reconstruction.
    Section III-A uses text condition c in both forward and reverse processes; assumes whisper-small transcriptions of possibly attacked audio are accurate enough to guide reconstruction without harming speaker identity.
  • domain assumption The ASV score difference d=|s-s'| separates clean from adversarial inputs.
    Section III-C defines the detector; assumes clean inputs yield small d and adversarial inputs yield large d, which is only demonstrated for one PGD attack.
  • standard math Standard DDPM forward and reverse process equations hold with the modified noise target Ntarget.
    Equations (1)-(2) and the 1000-step DDPM schedule are taken from Ho et al. [17] without re-derivation; this is standard, but the modified noise target changes the usual Gaussian noise assumptions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MDD: a Mask Diffusion Detector to Protect Speaker Verification Systems from Adversarial Perturbations." pith.science (2026). https://pith.science/paper/GHDF7PGN

@misc{pith2026250819180,
  author       = {Pith},
  title        = {Pith review of: MDD: a Mask Diffusion Detector to Protect Speaker Verification Systems from Adversarial Perturbations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GHDF7PGN}},
  note         = {Machine review of arXiv:2508.19180}
}
read the original abstract

Speaker verification systems are increasingly deployed in security-sensitive applications but remain highly vulnerable to adversarial perturbations. In this work, we propose the Mask Diffusion Detector (MDD), a novel adversarial detection and purification framework based on a \textit{text-conditioned masked diffusion model}. During training, MDD applies partial masking to Mel-spectrograms and progressively adds noise through a forward diffusion process, simulating the degradation of clean speech features. A reverse process then reconstructs the clean representation conditioned on the input transcription. Unlike prior approaches, MDD does not require adversarial examples or large-scale pretraining. Experimental results show that MDD achieves strong adversarial detection performance and outperforms prior state-of-the-art methods, including both diffusion-based and neural codec-based approaches. Furthermore, MDD effectively purifies adversarially-manipulated speech, restoring speaker verification performance to levels close to those observed under clean conditions. These findings demonstrate the potential of diffusion-based masking strategies for secure and reliable speaker verification systems.

Figures

Figures reproduced from arXiv: 2508.19180 by the authors.

Figure 1
Figure 1. An illustration of the workflow of the proposed MDD method. Given an input Mel-spectrogram, MDD first generates a masked version of the feature [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The distributions of score differences when applying 10% MDD and [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

34 extracted references · 31 canonical work pages

  1. [1]

    Malacopula: adversarial automatic speaker verification attacks using a neural-based gener- alised hammerstein model,

    M. Todisco, M. Panariello, X. Wang, H. Delgado, K.-A. Lee, and N. Evans, “Malacopula: adversarial automatic speaker verification attacks using a neural-based gener- alised hammerstein model,” in Proc. ASVspoof Workshop 2024, 2024

  2. [2]

    Intriguing properties of neural networks,

    C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus, “Intriguing properties of neural networks,” arXiv preprint arXiv:1312.6199, 2013

  3. [3]

    x-vectors meet adversarial attacks: Benchmarking adversarial robustness in speaker verification,

    J. Villalba, Y . Zhang, and N. Dehak, “x-vectors meet adversarial attacks: Benchmarking adversarial robustness in speaker verification,” in Interspeech, 2020, pp. 4233– 4237

  4. [4]

    Asvspoof 2019: Future horizons in spoofed and fake audio detection,

    M. Todisco, X. Wang, V . Vestman, M. Sahidullah, H. Del- gado, A. Nautsch, J. Yamagishi, N. Evans, T. H. Kin- nunen, and K. A. Lee, “Asvspoof 2019: Future horizons in spoofed and fake audio detection,” Interspeech 2019, 2019

  5. [5]

    Adversarial attacks and defenses in speaker recognition systems: A survey,

    J. Lan, R. Zhang, Z. Yan, J. Wang, Y . Chen, and R. Hou, “Adversarial attacks and defenses in speaker recognition systems: A survey,” Journal of Systems Architecture, vol. 127, p. 102526, 2022

  6. [6]

    Adversarial attack and defense strategies of speaker recognition systems: A survey,

    H. Tan, L. Wang, H. Zhang, J. Zhang, M. Shafiq, and Z. Gu, “Adversarial attack and defense strategies of speaker recognition systems: A survey,” Electronics, vol. 11, no. 14, p. 2183, 2022

  7. [7]

    Review of data processing methods used in predictive maintenance for next generation heavy machinery,

    I. U. Hassan, K. Panduru, and J. Walsh, “Review of data processing methods used in predictive maintenance for next generation heavy machinery,” Data, vol. 9, no. 5, p. 69, 2024

  8. [8]

    Scalable ensemble-based detection method against adversarial at- tacks for speaker verification,

    H. Wu, H.-C. Kuo, Y . Tsao, and H.-y. Lee, “Scalable ensemble-based detection method against adversarial at- tacks for speaker verification,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and 5 Signal Processing (ICASSP) . IEEE, 2024, pp. 4670– 4674

Show all 34 references
  1. [9]

    Lmd: A learnable mask network to detect ad- versarial examples for speaker verification,

    X. Chen, J. Wang, X.-L. Zhang, W.-Q. Zhang, and K. Yang, “Lmd: A learnable mask network to detect ad- versarial examples for speaker verification,” IEEE/ACM Transactions on Audio, Speech, and Language Process- ing, vol. 31, pp. 2476–2490, 2023

  2. [10]

    Improving the adversarial robustness for speaker veri- fication by self-supervised learning,

    H. Wu, X. Li, A. T. Liu, Z. Wu, H. Meng, and H.-Y . Lee, “Improving the adversarial robustness for speaker veri- fication by self-supervised learning,” IEEE/ACM Trans- actions on Audio, Speech, and Language Processing , vol. 30, pp. 202–217, 2021

  3. [11]

    Neural codec-based adversarial sample detection for speaker verification,

    X. Chen, J. Du, H. Wu, J.-S. R. Jang, and H. yi Lee, “Neural codec-based adversarial sample detection for speaker verification,” in Interspeech 2024, 2024, pp. 522– 526

  4. [12]

    Investigating robustness of adversarial samples detection for automatic speaker verification,

    X. Li, N. Li, J. Zhong, X. Wu, X. Liu, D. Su, D. Yu, and H. Meng, “Investigating robustness of adversarial samples detection for automatic speaker verification,” in Interspeech 2020, 2020, pp. 1540–1544

  5. [13]

    Advest: Adversarial perturbation estimation to classify and detect adversarial attacks against speaker identification,

    S. Joshi, S. Kataria, J. Villalba, and N. Dehak, “Advest: Adversarial perturbation estimation to classify and detect adversarial attacks against speaker identification,” in In- terspeech 2022, 2022, pp. 5060–5064

  6. [14]

    Diffusion-based adver- sarial purification for speaker verification,

    Y . Bai, X.-L. Zhang, and X. Li, “Diffusion-based adver- sarial purification for speaker verification,” IEEE Signal Processing Letters, 2024

  7. [15]

    Textual- driven adversarial purification for speaker verification,

    S. Chen, Y . Bai, J. Yao, X.-L. Zhang, and X. Li, “Textual- driven adversarial purification for speaker verification,” in Proc. Interspeech 2024 , 2024, pp. 527–531

  8. [16]

    Adversarial purification for speaker verification by two-stage diffusion models,

    Y . Bai, X.-L. Zhang, and X. Li, “Adversarial purification for speaker verification by two-stage diffusion models,” in 2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 1158–1164

  9. [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

  10. [18]

    Deep unsupervised learning using nonequi- librium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequi- librium thermodynamics,” in International conference on machine learning. PMLR, 2015, pp. 2256–2265

  11. [19]

    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 2018 IEEE international con- ference on acoustics, speech and signal processing (ICASSP). IEEE, 2018, pp. 5329–5333

  12. [20]

    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,” in INTERSPEECH. ISCA, 2020

  13. [21]

    Large-scale self-supervised speech representation learning for automatic speaker ver- ification,

    Z. Chen, S. Chen, Y . Wu, Y . Qian, C. Wang, S. Liu, Y . Qian, and M. Zeng, “Large-scale self-supervised speech representation learning for automatic speaker ver- ification,” in ICASSP 2022. IEEE, 2022, pp. 6147–6151

  14. [22]

    Adversarial examples in the physical world,

    A. Kurakin, I. J. Goodfellow, and S. Bengio, “Adversarial examples in the physical world,” in Artificial intelligence safety and security . Chapman and Hall/CRC, 2018, pp. 99–112

  15. [23]

    Towards deep learning models resistant to ad- versarial attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards deep learning models resistant to ad- versarial attacks,” in International Conference on Learn- ing Representations, 2018

  16. [24]

    Explaining and harnessing adversarial examples,

    I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,” in International Conference on Learning Representations , 2014

  17. [25]

    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 Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

  18. [26]

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

    H. Liu, Z. Chen, Y . Yuan, X. Mei, X. Liu, D. Mandic, W. Wang, and M. D. Plumbley, “Audioldm: Text-to-audio generation with latent diffusion models,” in International Conference on Machine Learning . PMLR, 2023, pp. 21 450–21 474

  19. [27]

    Diffusion models for image restoration and enhancement–a comprehensive survey,

    X. Li, Y . Ren, X. Jin, C. Lan, X. Wang, W. Zeng, X. Wang, and Z. Chen, “Diffusion models for image restoration and enhancement–a comprehensive survey,” arXiv preprint arXiv:2308.09388 , 2023

  20. [28]

    Conditional diffusion probabilistic model for speech enhancement,

    Y .-J. Lu, Z.-Q. Wang, S. Watanabe, A. Richard, C. Yu, and Y . Tsao, “Conditional diffusion probabilistic model for speech enhancement,” in ICASSP 2022-2022 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7402–7406

  21. [29]

    Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,” Advances in neural information processing systems, vol. 33, pp. 17 022–17 033, 2020

  22. [30]

    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 INTER- SPEECH. ISCA, 2017

  23. [31]

    Wespeaker: A research and production oriented speaker embedding learning toolkit,

    H. Wang, C. Liang, S. Wang, Z. Chen, B. Zhang, X. Xi- ang, Y . Deng, and Y . Qian, “Wespeaker: A research and production oriented speaker embedding learning toolkit,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, ...

  24. [32]

    Hifi-codec: Group-residual vector quantization for high fidelity audio codec,

    D. Yang, S. Liu, R. Huang, J. Tian, C. Weng, and Y . Zou, “Hifi-codec: Group-residual vector quantization for high fidelity audio codec,” arXiv preprint arXiv:2305.02765 , 2023

  25. [33]

    Speech- tokenizer: Unified speech tokenizer for speech language models,

    X. Zhang, D. Zhang, S. Li, Y . Zhou, and X. Qiu, “Speech- tokenizer: Unified speech tokenizer for speech language models,” in The Twelfth International Conference on Learning Representations

  26. [34]

    High-fidelity audio compression with im- proved rvqgan,

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, “High-fidelity audio compression with im- proved rvqgan,”Advances in Neural Information Process- ing Systems, vol. 36, pp. 27 980–27 993, 2023. 6

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.