Pith. sign in

REVIEW 3 major objections 5 minor 47 references

SegReConcat: A Data Augmentation Method for Voice Anonymization Attack

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read An attacker can expose anonymized voices by shuffling words and retraining a speaker verifier on the shuffled-plus-original audio, lowering equal error rate on five of seven anonymization systems.

desk verdict New and plausibly useful attack augmentation; the headline '5 of 7' claim is weaker than the evidence as reported. read the letter →

arxiv 2508.18907 v1 pith:2RPTONI2 submitted 2025-08-26 cs.SD cs.AI

Ridwan Arefeen , Xiaoxiao Miao , Rong Tong , Aik Beng Ng , Simon See This is my paper
classification cs.SDcs.AI
keywords voiceanonymizationattackdataaugmentationspeakerverificationword-levelsegmentationutterancerearrangementequalerrorrateresidualcuesECAPA-TDNN
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

Voice anonymization tries to hide who is speaking, but residual cues remain. This paper proposes SegReConcat, an attacker-side data augmentation: chop each anonymized utterance into word segments, rearrange the word order, concatenate the rearranged version with the original, and train a standard speaker-verification network on the combined audio. Tested across seven anonymization systems from the 2024 attacker challenge, the method lowers equal error rate on five of the seven, with the largest improvement roughly an 11-percentage-point drop. The authors argue that wrecking the natural word order forces the model to learn short-term word-level residual cues instead of long-range content structure.

What carries the argument

The central object is the concatenated training sequence w_aug = concat(w_o, w_r), where w_o is the original word segmentation of an anonymized utterance and w_r is a permutation of those word segments. A segmentation pass with Whisper-medium splits the utterance; a rearrangement function disrupts the natural content flow; and concatenation reunites both orders into one training example. The load-bearing mechanism is that permuting words destroys long-term contextual and temporal dependencies, while presenting the same words in two different orders forces the speaker encoder to learn word-level residual speaker traits that are invariant to order.

What would settle it

Retrain the SegReConcat attacker on B5 and T12-5 using forced-alignment word boundaries derived from the original transcripts instead of automatically detected boundaries on anonymized audio; if EER still fails to drop, the vector-quantization explanation holds, and if it drops, the original result depended on segmentation errors in anonymized speech.

Watch

Extended reading notes

Core claim

SegReConcat improves de-anonymization by training the ECAPA-TDNN attacker on concatenations of the original anonymized utterance and a word-level permuted version. Segmentation is done with Whisper-medium; rearrangement is either a random shuffle or a greedy ordering by acoustic (DTW-MFCC) or semantic (Whisper encoder cosine) similarity; the final training input is the original word sequence followed by the rearranged sequence. On the T8-5 system, rearrangement alone raises EER, but adding concatenation drops EER from 37.39% to 26.51%, and combining random rearrangement, concatenation, and SpecAugment reaches 25.94%. Across all seven systems, five show lower EER than the baseline attacker, w

Load-bearing premise

The method assumes word boundaries found by its transcription model on anonymized speech are just as reliable as on ordinary speech, but its accuracy evidence comes from clean un-anonymized recordings.

Editorial extensions

If this is right

  • Word-order randomization is a practical attack augmentation: residual speaker identity survives anonymization well enough to be learned from permuted word segments.
  • Concatenation, not rearrangement alone, drives the gain: shuffling alone hurts the attacker, while shuffling plus the original utterance lowers EER substantially.
  • SegReConcat and SpecAugment are partially complementary: combining them helps on some anonymization systems but not consistently.
  • Anonymization systems built on quantized self-supervised features are less exposed by this attack, suggesting the attack targets continuous temporal dynamics.
  • Future voice anonymization evaluation should include order-perturbation augmentations and should suppress word-level prosodic and articulatory cues, not only speaker-embedding statistics.

Reading between the lines

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

  • The two failures, B5 and T12-5, could be caused by Whisper segmenting quantized anonymized audio poorly rather than by the rearrangement idea failing; a forced-alignment control would separate word-boundary error from the paper's vector-quantization explanation.
  • Because rearrangement alone hurts while rearranged-plus-original helps, SegReConcat may work mainly as a viewpoint-invariance regularizer, suggesting that other structure-preserving corruptions of the same utterance might give similar or better gains.
  • For defenders, the practical lesson is that anonymization must remove cues that survive arbitrary word order, pushing residual-speaker removal down to sub-word timescales.
  • The method is backend-agnostic in principle, so the same augmentation could be applied to other speaker-verification architectures and other anonymization pipelines, but the paper only demonstrates it with ECAPA-TDNN.
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

3 major / 5 minor

Summary. The paper proposes SegReConcat, an attacker-side data augmentation method for ASV-based de-anonymization of voice anonymization systems. It uses Whisper-medium to segment anonymized utterances at word level, rearranges the word sequence with random, acoustic-DTW, or semantic-cosine strategies, and concatenates the rearranged utterance with the original before training an ECAPA-TDNN speaker verification attacker. Experiments follow the VPAC 2024 protocol on seven anonymization systems. Table I gives an ablation on T8-5; Table II reports per-system EER for a baseline and the better of SegReConcat and SegReConcat+SpecAug. The paper claims improvements on five of seven systems, with the strongest effect on T8-5 (average EER 37.39% to 25.94%).

Significance. If confirmed, SegReConcat is a useful and cheap addition to the voice-anonymization attacker toolbox. The idea of disrupting long-range temporal context while retaining word-level acoustics is distinct from SpecAug and spectrogram resizing, and the method is model-agnostic. The evaluation follows a fixed public challenge protocol, the code is released, and no fitted parameters enter the EER comparisons, so the result is not forced by the method's construction. The main limitation is statistical: all EER numbers come from single runs, the per-system configuration is selected after seeing results, and most of the reported gains are small. The T8-5 improvement is large enough to be credible, but the abstract's 'five out of seven' claim is not yet established.

major comments (3)
  1. [IV-B.2, Table II] The central claim of improvement on five of seven systems rests on a post-hoc selection. Section IV-B.2 states that both SegReConcat alone and SegReConcat+SpecAug were evaluated for every system and 'report the best results achieved'; Table II therefore shows the better of two variants per system. Under a null hypothesis of no true effect, taking the minimum of two noisy EER estimates makes apparent improvements more frequent than 50%. All values come from single training runs; no standard deviations, confidence intervals, or significance tests are reported. Several gains are small (B3: -0.94 pp; T10-2: -3.49 pp; T25-1: -2.46 pp; B4: -3.80 pp) and not consistent across dev/test subsets (e.g., B3 female dev worsens 27.56 to 27.58 while male test improves 27.84 to 24.28). Please add multi-seed repeats with variability, show both configurations for all systems, and/or apply a correction for
  2. [III-A] The segmentation stage assumes Whisper-medium detects word boundaries reliably on anonymized speech, but the cited 2.9% WER was measured on original LibriSpeech test-clean. Anonymization by TTS, codecs, or VQ bottlenecks can introduce phonetic distortion, timing changes, and insertions/deletions, making Whisper's boundaries potentially unreliable on exactly the inputs used. Since the proposed method and the AR/SR similarity measurements operate on 'word' segments, the validity of the word-level interpretability depends on this assumption. Please validate segmentation on anonymized data (e.g., forced alignment or manual checks on a sample from each of the seven systems) or soften the claim: if segmentation is inaccurate, SegReConcat still acts as chunk-level shuffling, which may explain the results without invoking word-level speaker cues.
  3. [IV-B.2, last paragraph] The post hoc explanation for failures on B5 and T12-5—that VQ layers 'remove continuity'—does not account for T25-1, which also uses VQ-BN features and improves. This should be labeled as a hypothesis rather than a mechanism, and it does not substitute for an empirical characterization of for which systems the method works. The paper would be strengthened by reporting per-system segmentation quality or a failure analysis tying the negative results to a measurable property of the anonymization pipeline.
minor comments (5)
  1. [III-B] 'SegReConcat specifically denotes the use of RR as the similarity metric during the rearrangement stage' is misleading: RR is a random strategy, not a similarity metric. Rephrase as 'rearrangement strategy'.
  2. [III-A] There is a typo: 'd′ is is the dimension' should be 'd′ is the dimension'.
  3. [IV-B.2] The sentence 'The absolute average EER reductions range from 1% to 3%' should say 'percentage points' and, excluding T8-5, the observed reductions are 0.94–3.80 pp.
  4. [Table II caption] The caption should state explicitly that the SegReConcat row for each system shows the better of SegReConcat alone and SegReConcat+SpecAug, rather than a single fixed configuration.
  5. [References] References [35] and [46] are the same SpecAugment paper and should be merged or cross-referenced.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the EER results come from a fixed external benchmark and no fitted parameter is repackaged as a prediction.

full rationale

SegReConcat is presented as an empirical augmentation technique, not as a first-principles derivation. The central claim—'SegReConcat improves de-anonymization on five out of seven systems'—is an observed EER comparison on the fixed VPAC 2024 benchmark, not a quantity computed from the method's own equations or normalizations. No parameter is fitted to the reported EERs and then renamed as a prediction: the augmentation variants (RR/AR/SR, with/without concatenation and SpecAug) are fixed procedures, and the ECAPA-TDNN attacker is trained from scratch on anonymized data. Whisper segmentation is an external tool justified by its public 2.9% WER on LibriSpeech test-clean; whether that accuracy transfers to anonymized speech is a robustness/correctness concern, not circularity. The authors do co-author several cited anonymization and challenge papers ([12], [13], [19], [20], [24], [25]), but those citations provide evaluation protocol and background—not a uniqueness theorem or a forbidden alternative—and the VPAC 2024 benchmark is externally specified. The main methodological weakness is disclosed in Section IV-B.2: 'We evaluated both SegReConcat alone and in combination with SpecAug for all anonymization systems, and report the best results achieved.' This is post-hoc selection without significance testing, which weakens the strength of the '5 of 7' claim, but it does not reduce the result to a definitional identity, since the reported EERs remain independent measurements under the challenge protocol. No circular step is present.

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

The central claim rests on three domain assumptions: reliable word segmentation on anonymized audio, the existence of residual speaker cues in prosody and temporal dynamics, and the VPAC EER protocol as a valid privacy measure. There are no free parameters fitted to produce the EER numbers; the method reports direct comparisons on a public benchmark. No invented entities are introduced.

assumptions (4)
  • domain assumption Whisper-medium word segmentation on anonymized speech is reliable
    Section III-A assumes word boundaries from anonymized utterances are correct; the quoted 2.9% WER is on original LibriSpeech test-clean, not on anonymized speech.
  • domain assumption Residual speaker cues persist in prosody, phonetics, and temporal dynamics after anonymization
    Introduction cites [37], [20], [15], [16], [38], [39] for this premise, which motivates the entire augmentation design.
  • domain assumption Lower EER on the VPAC dev/test protocol indicates a stronger de-anonymization attack
    Section II-A adopts the VPAC 2024 convention that EER is the privacy metric and lower EER means more identity leakage.
  • domain assumption ECAPA-TDNN trained from scratch on anonymized data is a representative attacker model
    The paper follows the VPAC baseline architecture but does not compare against top challenge attackers or other backends, so generalizability is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SegReConcat: A Data Augmentation Method for Voice Anonymization Attack." pith.science (2026). https://pith.science/paper/2RPTONI2

@misc{pith2026250818907,
  author       = {Pith},
  title        = {Pith review of: SegReConcat: A Data Augmentation Method for Voice Anonymization Attack},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2RPTONI2}},
  note         = {Machine review of arXiv:2508.18907}
}
read the original abstract

Anonymization of voice seeks to conceal the identity of the speaker while maintaining the utility of speech data. However, residual speaker cues often persist, which pose privacy risks. We propose SegReConcat, a data augmentation method for attacker-side enhancement of automatic speaker verification systems. SegReConcat segments anonymized speech at the word level, rearranges segments using random or similarity-based strategies to disrupt long-term contextual cues, and concatenates them with the original utterance, allowing an attacker to learn source speaker traits from multiple perspectives. The proposed method has been evaluated in the VoicePrivacy Attacker Challenge 2024 framework across seven anonymization systems, SegReConcat improves de-anonymization on five out of seven systems.

Figures

Figures reproduced from arXiv: 2508.18907 by the authors.

Figure 1
Figure 1. (1) SegReConcat Flow Diagram • T8-5 [15]: a hybrid scheme randomly choosing between (1) ASR(Whisper)+TTS(VITS) and (2) k-NN voice con￾version on WavLM features. • T10-2 [43]: a neural codec (using VQ) that explicitly disentangles linguistic content, speaker identity, and emo￾tion. • T12-5 [44]: an extension of B5 with additional pitch smoothing. • T25-1 [45]: VQ-BN content features with global style tokens and emoti… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 41 canonical work pages

  1. [1]

    General data protection regulation (GDPR),

    “General data protection regulation (GDPR),” https://gdpr.eu/ what-is-gdpr

    work page

  2. [2]

    The voiceprivacy 2020 challenge: Results and findings,

    N. Tomashenko, X. Wang, E. Vincent, J. Patino, B. M. L. Srivastava, P.-G. No´e, A. Nautsch, N. Evans, J. Yamagishi, B. O’Brien, A. Chanclu, J.-F. Bonastre, M. Todisco, and M. Maouche, “The voiceprivacy 2020 challenge: Results and findings,” Computer Speech & Language, vol. 74, p. 101362, Jul. 2022

    work page 2020

  3. [4]

    The V oicePrivacy 2022 challenge: Progress and perspectives in voice anonymisation,

    M. Panariello, N. Tomashenko, X. Wang, X. Miao, P. Champion, H. Nourtel, M. Todisco, N. Evans, E. Vincent, and J. Yamagishi, “The V oicePrivacy 2022 challenge: Progress and perspectives in voice anonymisation,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, pp. 1–14, 2024

    work page 2022

  4. [5]

    The V oicePri- vacy 2024 challenge evaluation plan,

    N. Tomashenko, X. Miao, P. Champion, S. Meyer, X. Wang, E. Vincent, M. Panariello, N. Evans, J. Yamagishi, and M. Todisco, “The V oicePri- vacy 2024 challenge evaluation plan,” arXiv preprint arXiv:2404.02677, 2024

    arXiv 2024

  5. [6]

    Anonymizing speech: Evaluating and designing speaker anonymization techniques,

    P. Champion, S. Ouni, D. Jouvet, and A. Larcher, “Anonymizing speech: Evaluating and designing speaker anonymization techniques,” arXiv preprint, vol. arXiv:2308.04455v4, 2023

    arXiv 2023

  6. [7]

    Speaker anonymization using neural audio codec language models,

    M. Panariello, F. Nespoli, M. Todisco, and N. Evans, “Speaker anonymization using neural audio codec language models,” 2024

    work page 2024

  7. [8]

    Prosody is not identity: A speaker anonymization approach using prosody cloning,

    S. Meyer, F. Lux, J. Koch, P. Denisov, P. Tilli, and N. T. Vu, “Prosody is not identity: A speaker anonymization approach using prosody cloning,” in Proc. IEEE ICASSP . IEEE, 2023, pp. 1–5

    work page 2023

  8. [9]

    Differentially private speaker anonymization,

    A. S. Shamsabadi, B. M. L. Srivastava, A. Bellet, N. Vauquier, E. Vin- cent, M. Maouche, M. Tommasi, and N. Papernot, “Differentially private speaker anonymization,” Proceedings on Privacy Enhancing Technolo- gies, vol. 2023, no. 1, Jan. 2023

    work page 2023

Show all 47 references
  1. [10]

    Speaker anonymization by modifying fundamental frequency and x- vector singular value,

    C. O. Mawalim, K. Galajit, J. Karnjana, S. Kidani, and M. Unoki, “Speaker anonymization by modifying fundamental frequency and x- vector singular value,” Computer Speech & Language , vol. 73, p. 101326, 2022

    2022

  2. [11]

    MUSA: Multi-lingual speaker anonymization via serial disentanglement,

    J. Yao, Q. Wang, P. Guo, Z. Ning, Y . Yang, Y . Pan, and L. Xie, “MUSA: Multi-lingual speaker anonymization via serial disentanglement,” arXiv preprint arXiv:2407.11629, 2024

    2024 arXiv

  3. [12]

    Language-independent speaker anonymization approach using self- supervised pre-trained models,

    X. Miao, X. Wang, E. Cooper, J. Yamagishi, and N. Tomashenko, “Language-independent speaker anonymization approach using self- supervised pre-trained models,” in Proc. The Speaker and Language Recognition Workshop (Odyssey 2022) , 2022, pp. 279–286

    2022

  4. [13]

    Speaker anonymization using orthogonal householder neural network,

    ——, “Speaker anonymization using orthogonal householder neural network,” IEEE/ACM Trans. Audio, Speech, and Language Processing , vol. 31, pp. 3681–3695, 2023

    2023

  5. [14]

    Distinctive and natural speaker anonymization via singular value transformation-assisted 5 matrix,

    J. Yao, Q. Wang, P. Guo, Z. Ning, and L. Xie, “Distinctive and natural speaker anonymization via singular value transformation-assisted 5 matrix,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024

    2024

  6. [15]

    Hltcoe jhu submission to the voice privacy challenge 2024,

    H. L. Xinyuan, Z. Cai, A. Garg, K. Duh, L. P. Garc ´ıa-Perera, S. Khudan- pur, N. Andrews, and M. Wiesner, “Hltcoe jhu submission to the voice privacy challenge 2024,” arXiv preprint arXiv:2409.08913 , 2024

    2024 arXiv

  7. [16]

    Privacy versus emotion preservation trade-offs in emotion-preserving speaker anonymization,

    Z. Cai, H. L. Xinyuan, A. Garg, L. P. Garc ´ıa-Perera, K. Duh, S. Khudan- pur, N. Andrews, and M. Wiesner, “Privacy versus emotion preservation trade-offs in emotion-preserving speaker anonymization,” in 2024 IEEE Spoken Language Technology Workshop (SLT) . IEEE, 2024, pp. 409– 414

    2024

  8. [17]

    Anonymising elderly and pathological speech: V oice conversion using ddsp and query-by-example,

    S. Ghosh, M. Jouaiti, A. Das, Y . Sinha, T. Polzehl, I. Siegert, and S. Sto- ber, “Anonymising elderly and pathological speech: V oice conversion using ddsp and query-by-example,” Interspeech, 2024

    2024

  9. [18]

    V oicePM: A robust privacy measurement on voice anonymity,

    Z. Shaohu, L. Zhouyu, and D. Anupam, “V oicePM: A robust privacy measurement on voice anonymity,” in Proc. 16th ACM Conference on Security and Privacy in Wireless and Mobile Networks (WiSec) , 2023, p. 215–226

    2023

  10. [19]

    A benchmark for multi- speaker anonymization,

    X. Miao, R. Tao, C. Zeng, and X. Wang, “A benchmark for multi- speaker anonymization,” IEEE Transactions on Information Forensics and Security, vol. 20, pp. 3819–3833, 2025

    2025

  11. [20]

    Adapting general disentanglement-based speaker anonymization for enhanced emotion preservation,

    X. Miao, Y . Zhang, X. Wang, N. Tomashenko, D. C. L. Soh, and I. Mcloughlin, “Adapting general disentanglement-based speaker anonymization for enhanced emotion preservation,” Computer Speech & Language, p. 101810, 2025

    2025

  12. [21]

    Easy: Emotion-aware speaker anonymization via factorized distillation,

    J. Yao, H. Liu, E. S. Chng, and L. Xie, “Easy: Emotion-aware speaker anonymization via factorized distillation,” Accepted by Interspeech 2025, 2025

    2025

  13. [22]

    Private knn-vc: Inter- pretable anonymization of converted speech,

    C. Franzreb, A. Das, T. Polzehl, and S. M ¨oller, “Private knn-vc: Inter- pretable anonymization of converted speech,” in Accepted by Interspeech 2025, 2025

    2025

  14. [23]

    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 Interspeech 2020. ISCA, 2020

    2020

  15. [24]

    The first voiceprivacy attacker challenge evaluation plan,

    N. Tomashenko, X. Miao, E. Vincent, and J. Yamagishi, “The first voiceprivacy attacker challenge evaluation plan,” 2024

    2024

  16. [25]

    The first voiceprivacy attacker challenge,

    ——, “The first voiceprivacy attacker challenge,” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, Apr. 2025, p. 1–2

    2025

  17. [26]

    Specwav- attack: Leveraging spectrogram resizing and wav2vec 2.0 for attacking anonymized speech,

    Y . Li, Y . Zheng, Z. Guo, Y . Wang, J. Yin, and H. Fei, “Specwav- attack: Leveraging spectrogram resizing and wav2vec 2.0 for attacking anonymized speech,” in ICASSP 2025-2025 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–2

    2025

  18. [27]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” in Proc. NIPS, vol. 33, 2020, pp. 12 449–12 460

    2020

  19. [28]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  20. [29]

    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

  21. [30]

    Fast adaptation of pretrained speaker verification system for source speaker tracking,

    X. Lyu, Y . Wang, T. Zhao, and H. Liu, “Fast adaptation of pretrained speaker verification system for source speaker tracking,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–2

    2025

  22. [31]

    Fine-tuning titanet-large model for speaker anonymization attacker systems,

    C. O. Mawalim, A. Adila, and M. Unoki, “Fine-tuning titanet-large model for speaker anonymization attacker systems,” in ICASSP 2025- 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2025, pp. 1–2

    2025

  23. [32]

    Attacking voice anonymization systems with augmented feature and speaker identity difference,

    Y . Zhang, Z. Bi, F. Xiao, X. Yang, Q. Zhu, and J. Guan, “Attacking voice anonymization systems with augmented feature and speaker identity difference,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–2

    2025

  24. [33]

    Probabilistic linear discriminant analysis for inferences about identity,

    S. J. Prince and J. H. Elder, “Probabilistic linear discriminant analysis for inferences about identity,” in 2007 IEEE 11th international conference on computer vision . IEEE, 2007, pp. 1–8

    2007

  25. [34]

    Probabilistic linear discriminant analysis,

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

    2006

  26. [35]

    Specaugment: A simple data augmentation method for automatic speech recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” in Interspeech 2019. ISCA, Sep. 2019

    2019

  27. [36]

    Hltcoe submission to the voicepri- vacy attacker challenge,

    H. L. Xinyuan, A. Garg, Z. Cai, K. Duh, L. P. Garc ´ıa-Perera, S. Khudan- pur, N. Andrews, and M. Wiesner, “Hltcoe submission to the voicepri- vacy attacker challenge,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE...

    2025

  28. [37]

    Why disentanglement-based speaker anonymization systems fail at preserving emotions?

    U. E. Gaznepoglu and N. Peters, “Why disentanglement-based speaker anonymization systems fail at preserving emotions?” in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

    2025

  29. [38]

    Analysis of speech temporal dynamics in the context of speaker verification and voice anonymization,

    N. Tomashenko, E. Vincent, and M. Tommasi, “Analysis of speech temporal dynamics in the context of speaker verification and voice anonymization,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–5

    2025

  30. [39]

    Exploiting context-dependent duration features for voice anonymization attack systems,

    ——, “Exploiting context-dependent duration features for voice anonymization attack systems,” in Interspeech 2025, 2025

    2025

  31. [40]

    Inference Attacks for X- Vector Speaker Anonymization ,

    W. Bao, M. Jadhav, and V . Bindschaedler, “ Inference Attacks for X- Vector Speaker Anonymization ,” in 2025 IEEE Security and Privacy Workshops (SPW), 2025, pp. 152–159

    2025

  32. [41]

    LibriSpeech: an ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “LibriSpeech: an ASR corpus based on public domain audio books,” in Proc. ICASSP . IEEE, 2015, pp. 5206–5210

    2015

  33. [42]

    Anonymiz- ing speech with generative adversarial networks to preserve speaker privacy,

    S. Meyer, P. Tilli, P. Denisov, F. Lux, J. Koch, and N. T. Vu, “Anonymiz- ing speech with generative adversarial networks to preserve speaker privacy,” in 2022 IEEE Spoken Language Technology Workshop (SLT) , 2023, pp. 912–919

    2022

  34. [43]

    NPU-NTU System for V oice Privacy 2024 Challenge,

    J. Yao, N. Kuzmin, Q. Wang, P. Guo, Z. Ning, D. Guo, K. A. Lee, E.-S. Chng, and L. Xie, “NPU-NTU System for V oice Privacy 2024 Challenge,” arXiv preprint arXiv:2409.04173 , 2024

    2024 arXiv

  35. [44]

    NTU-NPU System for V oice Privacy 2024 Challenge,

    N. Kuzmin, H.-T. Luong, J. Yao, L. Xie, and K. A. Lee, “NTU-NPU System for V oice Privacy 2024 Challenge,” SPSC 2024, 2024. [Online]. Available: https://www.voiceprivacychallenge.org/vp2024/docs/T12 NTU-NPU System for V oicePrivacy 2024 Challenge.pdf

    2024

  36. [45]

    USTC-PolyU system for the V oicePrivacy 2024 Challenge,

    W. Gu, Z. Liu, L. Chen, R. Wang, C. Guo, W. Guo, K. A. Lee, and Z.-H. Ling, “USTC-PolyU system for the V oicePrivacy 2024 Challenge,” SPSC 2024 , 2024. [Online]. Available: https://www.voiceprivacychallenge.org/vp2024/docs/T25 USTC-PolyU system for the V oicePrivacy2024 Challenge.pdf

    2024

  37. [46]

    Specaugment: A simple data augmentation method for automatic speech recognition,

    D. S. Park, W. Chan, Y . Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V . Le, “Specaugment: A simple data augmentation method for automatic speech recognition,” Interspeech 2019, Sep 2019

    2019

  38. [47]

    Robust speech recognition via large-scale weak supervi- sion,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” 2022

    2022

  39. [48]

    Dynamic time warping,

    M. M ¨uller, “Dynamic time warping,” Information retrieval for music and motion, pp. 69–84, 2007. 6

    2007

Pith tools

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