Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Mitigating Language Mismatch in SSL-Based Speaker Anonymization

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A multilingual HuBERT content encoder fine-tuned on Japanese makes anonymized speech more intelligible in both Japanese and Mandarin while preserving speaker privacy.

desk verdict Solid recipe with a confounded headline claim: the CER gains are real and reproducible, but the 'multilingual' credit isn't isolated without an mHU-EN control. read the letter →

arxiv 2507.00458 v1 pith:YNSJUNC2 submitted 2025-07-01 eess.AS cs.SD

Zhe Zhang , Wen-Chin Huang , Xin Wang , Xiaoxiao Miao , Junichi Yamagishi This is my paper
classification eess.AScs.SD
keywords speakeranonymizationself-supervisedlearninglanguagemismatchmultilingualspeechHuBERTmcharactererrorrateprivacy
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

Most speaker anonymization systems are built and tested on English only, and when they process Japanese or Mandarin, the anonymized speech becomes unintelligible—character error rates can exceed 100%. This paper argues that the root cause is language mismatch in the self-supervised (SSL) content encoder, and it tests a remedy: fine-tuning the encoder on Japanese, first keeping the English-only HuBERT foundation and then replacing it with the multilingual mHuBERT. Fine-tuning the English-only model on Japanese already cuts character error rates for Japanese, and swapping in mHuBERT produces further gains in both Japanese and unseen Mandarin. Equal error rates in speaker verification stay high (for example, 39.44% EER for mHU-JA with the OHNN anonymizer on Japanese), so privacy is not sacrificed. The authors conclude that language adaptation and multilingual SSL pre-training are key to making speaker anonymization work outside English.

What carries the argument

The central object is the soft content encoder: a HuBERT foundation model fine-tuned with an extra linear layer so that it outputs continuous speech-content representations rather than discrete units. The paper compares three instantiations—HU-EN (English-only HuBERT fine-tuned on 100 hours of English read speech), HU-JA (the same HuBERT fine-tuned on 100 hours of Japanese spontaneous speech), and mHU-JA (mHuBERT-147, a multilingual HuBERT, fine-tuned on the same Japanese corpus)—while keeping the speaker encoder, pitch tracker, and vocoder fixed across systems. Because only the foundation model and its fine-tuning data are changed, downstream differences in character error rate and equal error rate are attributed to language adaptation and multilingual pre-training.

What would settle it

Fine-tune mHuBERT on the same 100 hours of English (mHU-EN) under the same recipe and compare its Japanese and Mandarin CERs against mHU-JA; if mHU-EN matches or beats mHU-JA, then the gain is from the multilingual foundation's capacity or pre-training, not from adaptation to Japanese.

Watch

Extended reading notes

Core claim

The paper establishes that, within the SSL-based soft content encoder framework, replacing the English-only HuBERT foundation with the multilingual mHuBERT and fine-tuning on Japanese improves utility for both the adapted language (Japanese) and an expanded language (Mandarin), while maintaining privacy. In the language-adapted Japanese evaluation, both HU-JA and mHU-JA lower character error rates relative to HU-EN at both Katakana and Kanji transcription levels, with mHU-JA best (e.g., selection-based anonymizer: 4.04% vs 5.57% for HU-EN in Katakana CER). In the language-expanded Mandarin evaluation, mHU-JA yields the largest drop in CER—from 25.97% (HU-EN) to 12.67% with the selection-based anonymizer, and from 25.74% to 14.13% with OHNN—while EERs remain in a protective range. A per-syllable analysis of Japanese shows that units the English-only model fails to render (e.g., ヌ and ヒ) become recognizable with mHU-JA, supporting the claim that multilingual representations capture cross-language phonetic structure. The authors present this as evidence that multilingual SSL pre-training, combined with target-language fine-tuning, mitigates language mismatch and extends speaker anonymization to a broader linguistic scope.

Load-bearing premise

The paper's central attribution—that multilingual pre-training, not just Japanese fine-tuning, causes the CER improvement—assumes that mHuBERT and HuBERT are otherwise comparable, so that the only meaningful difference is the languages they were pre-trained on.

Editorial extensions

If this is right

  • Fine-tuning an English-only SSL content encoder on the target language is enough to substantially reduce character error rates for that language, without retraining the speaker encoder or vocoder.
  • A multilingual SSL foundation fine-tuned on a single language carries over to other, unseen languages: mHU-JA improves Mandarin intelligibility even though Mandarin never appeared in fine-tuning or vocoder training.
  • The improvement is consistent across two different anonymizers (selection-based and OHNN-based), indicating that the fix operates at the content representation level rather than at the speaker-anonymization step.
  • Per-syllable analysis shows that specific Japanese units that were unintelligible with the English-only encoder become recognizable with mHU-JA, suggesting the gains are grounded in phonetically meaningful representation quality, not just overall statistics.
  • The privacy-utility tradeoff is favorable: EERs remain high (e.g., 39.44% for mHU-JA with OHNN on Japanese) even as CERs drop, so the method protects identity while restoring content.

Reading between the lines

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

  • The paper compares mHU-JA against HU-JA but not against an mHU-EN (multilingual HuBERT fine-tuned on English); without that arm, part of mHU-JA's advantage could come from mHuBERT's architecture, parameter count, or pre-training data rather than from its multilingual nature. Running mHU-EN would isolate the language-adaptation effect.
  • The English fine-tuning corpus is read audiobook speech while the Japanese corpus is spontaneous conversational speech, so language and speaking style are confounded in the Japanese-vs-English comparison; a matched-style read Japanese corpus would tighten the causal story.
  • If the language-expanded gains are driven by shared phonetic structure, then fine-tuning on a language unrelated to Mandarin (say, French) should transfer less to Mandarin than fine-tuning on Japanese does; this is a direct, testable prediction of the paper's mechanism.
  • The recipe—100 hours of target-language speech to fine-tune a multilingual SSL encoder, with existing speaker and vocoder modules left untouched—is a low-cost practical path to extend English-centric anonymizers to new languages in deployments.
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 studies language mismatch in SSL-based speaker anonymization. The authors fine-tune a HuBERT content encoder on English (HU-EN) and on Japanese (HU-JA), and fine-tune the multilingual mHuBERT-147 model on Japanese (mHU-JA). They compare these encoders within a speaker anonymization framework using selection-based and OHNN-based anonymizers, evaluating utility via CER on Japanese (JVS) and Mandarin (AISHELL-3) test sets and privacy via EER on JTubeSpeech and AISHELL-3. The main reported result is that mHU-JA substantially lowers Mandarin CER relative to HU-EN and HU-JA, while keeping EER at levels the authors consider protective. The paper also presents a per-Katakana phonetic analysis and releases code and pretrained weights.

Significance. If the central claim is supported, the finding that fine-tuning a multilingual SSL model on a single language improves anonymization utility for another unseen language would be practically valuable for broadening SASs beyond English. The paper has clear strengths: it uses standard VPC evaluation protocols, reports both utility and privacy metrics, provides bootstrap confidence intervals for the CER results, includes open code/weights, and performs a fine-grained phonetic analysis. However, the main attributive claim—that multilingual pre-training, rather than the choice of the mHuBERT foundation model, drives the improvement—is not isolated by the current experimental design. The missing mHU-EN control is load-bearing for the central claim and should be addressed before the paper can be accepted.

major comments (3)
  1. [3.2 and 5.3] The comparison between mHU-JA and HU-JA does not isolate the effect of multilingual pre-training. As stated in Section 3.2, the two conditions differ not only in the fine-tuning language but also in the foundation model: HuBERT is trained on English-only data, while mHuBERT-147 is trained on a 147-language corpus and has a different architecture/recipe. The sentence in Section 1 claiming that the choice of HuBERT-based models 'eliminate[s] the impact of different SSL architectures' is not accurate because HuBERT and mHuBERT are not the same architecture. Without an mHU-EN condition (mHuBERT fine-tuned on the same English data and protocol), the large Mandarin CER gains (e.g., OHNN 22.76 to 14.13 in Table 1) cannot be attributed specifically to multilinguality; they could come from the different pre-training data scale, model capacity, or training procedure. Please add this control, or revise the Section 5.3 claim to state that the improvement is due to the mHuBERT-based content encoder rather than multilingual pre-training per se.
  2. [4 and 5.2] The HU-EN versus HU-JA comparison is confounded by domain. HU-EN is fine-tuned on LibriSpeech-train-clean-100, which contains read audiobook speech, whereas HU-JA and mHU-JA are fine-tuned on 100 hours sampled from the Corpus of Spontaneous Japanese, which is spontaneous speech. The CER improvements attributed to 'language adaptation' in Section 5.2 could therefore reflect adaptation to spontaneous speech rather than to Japanese. The authors should either fine-tune an English model on a matched-domain spontaneous English corpus, or clearly acknowledge this confound and provide an additional controlled comparison.
  3. [5.5 and Table 1] The statement that 'the EER of ASV results slightly decrease when using the HU-JA and mHU-JA models' is not supported by Table 1. Several EER differences are sizable: for Mandarin with the OHNN anonymizer, EER drops from 42.55 (HU-EN) to 31.20 (mHU-JA); for Mandarin with the selection-based anonymizer, from 44.33 to 35.21; and for Japanese with the selection-based anonymizer, from 47.87 to 39.91. These reductions are not 'slight' and need to be discussed quantitatively, especially because the paper concludes that privacy is 'maintained.' Moreover, the bootstrapping significance test described in Section 5.1 is applied only to the ASR/CER results, so it is unclear whether the ASV/EER differences are statistically reliable. Please add confidence intervals or significance testing for the ASV results, or temper the privacy-maintenance claim.
minor comments (5)
  1. [1, 2.1, 8] The name 'VoicePrivacy' appears as 'V oicePrivacy' in several places (e.g., Section 1 and references [1, 2, 3]); this formatting artifact should be cleaned.
  2. [3.1] The pitch tracker is written as 'YAAPT algorithm' but appears as 'Y AAPT' in the text; please fix the spacing.
  3. [5.1] The description of the bootstrapping significance test does not state the number of resamples or the confidence level used to produce the confidence intervals in Figure 3; please include these details for reproducibility.
  4. [4] Please clarify whether the randomly sampled 100 hours of Japanese speech from CSJ is exactly matched in duration distribution and speaker count to LibriSpeech-clean-100, or whether the two corpora differ in these respects; this information is relevant to the domain-confounding discussion in Major Comment 2.
  5. [5.2] The Japanese CER differences between HU-JA and mHU-JA are small (e.g., OHNN CERkanji 8.95 vs 8.90, and selection CERkata 4.88 vs 4.18); it would strengthen the paper to report which of these small differences are actually statistically significant, since the claim in Section 5.2 that 'the multilingual models can further improve the intelligibility' relies on these numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are direct empirical comparisons on held-out test sets, with no fitted parameter renamed as a prediction and no load-bearing self-citation chain.

full rationale

The paper's central claims are empirical comparisons among three content encoders (HU-EN, HU-JA, mHU-JA) evaluated with held-out Japanese and Mandarin corpora. No parameter is fitted to the test data and then reported as a prediction; the CER and EER numbers come from fixed evaluation pipelines (Whisper ASR, ECAPA-TDNN ASV) on unseen utterances. The main weakness that a skeptic might raise is the missing mHU-EN control, which means the mHU-JA versus HU-JA contrast does not isolate multilingual pre-training from other differences between HuBERT Base and mHuBERT-147. That is a validity or confound concern, not a circularity concern: the comparison is still an independent empirical measurement, not a reduction of the conclusion to the construction of the models. Similarly, the HU-EN versus HU-JA comparison confounds language with corpus domain (LibriSpeech read speech versus CSJ spontaneous speech), but again this is an experimental-design limitation rather than circular reasoning. The paper's self-citations (e.g., the SSL-based framework from prior work by the authors) are used as building blocks and baselines whose behavior is measured externally; they are not invoked to define away the target result or to forbid alternative explanations via an unverified uniqueness claim. There is no equation in the paper that is equivalent to the paper's own input, no fitted parameter renamed as a prediction, and no first-principles derivation that turns out to rest on its own conclusion. Accordingly, the honest circularity score is 0.

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

The central claims are empirical, so there are no fitted constants in a derivation. The ledger instead lists the domain assumptions and the implicit attribution assumptions that the experimental design does not isolate. No new entities are postulated.

assumptions (4)
  • domain assumption Fine-tuning an SSL model with an extra linear layer produces content representations that suppress speaker identity.
    Invoked in Section 3.2 following van Niekerk et al. [26].
  • domain assumption ASR (whisper-large-v3) and ASV (ECAPA-TDNN) metrics faithfully measure intelligibility and privacy.
    Section 5.1 assumes these tools are valid across languages.
  • ad hoc to paper The difference between HU-JA and mHU-JA is caused by multilingual pre-training, not by architecture or training data.
    No mHU-EN control; Table 1 comparison confounds these factors.
  • ad hoc to paper The only relevant difference between HU-EN and HU-JA is the language of fine-tuning data.
    Section 4 datasets differ in domain (LibriSpeech read vs CSJ spontaneous).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Language Mismatch in SSL-Based Speaker Anonymization." pith.science (2026). https://pith.science/paper/YNSJUNC2

@misc{pith2026250700458,
  author       = {Pith},
  title        = {Pith review of: Mitigating Language Mismatch in SSL-Based Speaker Anonymization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YNSJUNC2}},
  note         = {Machine review of arXiv:2507.00458}
}
read the original abstract

Speaker anonymization aims to protect speaker identity while preserving content information and the intelligibility of speech. However, most speaker anonymization systems (SASs) are developed and evaluated using only English, resulting in degraded utility for other languages. This paper investigates language mismatch in SASs for Japanese and Mandarin speech. First, we fine-tune a self-supervised learning (SSL)-based content encoder with Japanese speech to verify effective language adaptation. Then, we propose fine-tuning a multilingual SSL model with Japanese speech and evaluating the SAS in Japanese and Mandarin. Downstream experiments show that fine-tuning an English-only SSL model with the target language enhances intelligibility while maintaining privacy and that multilingual SSL further extends SASs' utility across different languages. These findings highlight the importance of language adaptation and multilingual pre-training of SSLs for robust multilingual speaker anonymization.

Figures

Figures reproduced from arXiv: 2507.00458 by the authors.

Figure 1
Figure 1. Utility evaluation via ASR task on anonymized Japanese and Mandarin speech using VPC baselines (B2 - B6). low equal error rate (EER) of 4.52% in the automatic speaker verify (ASV) experiments [3]. Improving the utility of SASs while maintaining privacy is an important topic, especially for extending the usage of English-only models to different lan￾guages. Among the various SAS methods, the self-supervised learning … view at source ↗
Figure 2
Figure 2. Framework of the SSL-based multilingual SAS. learning, showing remarkable performance for speech synthe￾sis [23, 24]. XLS-R [10] is used as a teacher model in distil￾lation in [5]. A language-independent speaker anonymization approach [6] adopts an SSL-based soft content encoder, which demonstrates better performance on unseen language. A fol￾lowing analysis [7] discussed the language mismatch of SSL￾based SASs in d… view at source ↗
Figure 3
Figure 3. CER scores and confidence intervals. 5. Evaluation 5.1. Experiment Setup To assess the anonymizer’s ability to protect speaker identity, we adopted the privacy metric and evaluation methods from VPC. Specifically, we tested the ASV performance in terms of EER by using an ECAPA-TDNN [28] model trained on mul￾tilingual speech data. ASV experiments are conducted under the ignorant [25] setting, where attackers are unaw… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Per-Katakana CER of synthesized Japanese Speech. 5.5. The Privacy-Utility Tradeoff Despite successfully confirming the effectiveness of multilin￾gual SSL models, we notice that the EER of ASV results slightly decrease when using the HU-JA and mHU-JA mod￾els. This is be…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages

  1. [1]

    Introduction Speaker anonymization has gained increasing attention as pri- vacy concerns grow in speech applications. With the rapid ad- vancement of AI-driven speech technologies, it is crucial to ensure that sensitive speaker characteristics remain concealed while preserving speech’s linguistic content and naturalness. Recent advances, as demonstrated i...

    work page

  2. [2]

    Fine-tuning a multilingual HuBERT model, mHuBERT [8], with a specific language, i.e., Japanese, to investigate whether the language-adapted SSL model is helpful for adapting the SAS to a particular language

    work page

  3. [3]

    Mitigating Language Mismatch in SSL-Based Speaker Anonymization

    Use the above multilingual HuBERT model fine-tuned to the specific language and examine whether it also contains mean- ingful representations for other languages, such as Mandarin. This is an essential step towards properly understanding the root of the language mismatch problem in multilingual speaker anonymization systems and proposing a solution. Note ...

    work page Pith review arXiv 2025

  4. [4]

    English-Based SASs Speaker anonymization is a voice privacy solution to conceal speaker identity without degrading intelligibility and natural- ness [1]

    Related Work 2.1. English-Based SASs Speaker anonymization is a voice privacy solution to conceal speaker identity without degrading intelligibility and natural- ness [1]. Aiming at standardizing and advancing the develop- ment of voice privacy preservation techniques, the V oicePrivacy Challenge (VPC) was initiated [1] and held in 2022 [2] and 2024 [3], ...

    work page 2022

  5. [5]

    Then, we introduce the details of the multilingual soft content encoder, including the foundation models and fine-tuning procedures

    Methods In this section, we describe framework of the proposed SSL- based multilingual SAS, which is illustrated in Figure 2. Then, we introduce the details of the multilingual soft content encoder, including the foundation models and fine-tuning procedures. 3.1. SSL-based Speaker Anonymization System Our SAS framework consists of a HuBERT-based soft cont...

    work page

  6. [6]

    Dataset In the training stage, LibriSpeech-train-clean-100 dataset is used [31] to fine-tune the HuBERT foundation model in HU-EN. For a fair comparison, we randomly sampled 100 hours of Japanese speech from Corpus of Spontaneous Japanese dataset [32] and segmented the speech audio to utterance-level with random durations between 2 and 20 seconds to fine-...

    work page

  7. [7]

    This study was carried out using the TSUB- AME4.0 supercomputer at Institute of Science Tokyo

    Acknowledgments This study is partially supported by JST AIP Acceleration Research (JPMJCR24U3) and by MEXT KAKENHI Grants (24K21324). This study was carried out using the TSUB- AME4.0 supercomputer at Institute of Science Tokyo

    work page

  8. [8]

    Experiment Setup To assess the anonymizer’s ability to protect speaker identity, we adopted the privacy metric and evaluation methods from VPC

    Evaluation 5.1. Experiment Setup To assess the anonymizer’s ability to protect speaker identity, we adopted the privacy metric and evaluation methods from VPC. Specifically, we tested the ASV performance in terms of EER by using an ECAPA-TDNN [28] model trained on mul- tilingual speech data. ASV experiments are conducted under the ignorant [25] setting, w...

    work page

Show all 46 references
  1. [9]

    Our study revealed that English-only SASs lead to significant utility degradation when processing Japanese and Mandarin speech

    Conclusion In this paper, we presented an extended investigation of lan- guage mismatch in SSL-based speaker anonymization. Our study revealed that English-only SASs lead to significant utility degradation when processing Japanese and Mandarin speech. By fine-tuning content en...

  2. [10]

    XLS-R: Self-supervised Cross-lingual Speech Rep- resentation Learning at Scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. V on Platen, Y . Saraf, J. Pino, A. Baevski, A. Conneau, and M. Auli, “XLS-R: Self-supervised Cross-lingual Speech Rep- resentation Learning at Scale,” in Proc. Interspeech, 2022, pp. 2278–2282

    2022

  3. [11]

    Introducing the V oicePrivacy Initiative,

    N. Tomashenko, B. M. L. Srivastava, X. Wang, E. Vincent, A. Nautsch, J. Yamagishi, N. Evans, J. Patino, J.-F. Bonastre, P.-G. No´e, and M. Todisco, “Introducing the V oicePrivacy Initiative,” in Proc. Interspeech, 2020, pp. 1693–1697

    2020

  4. [12]

    The V oicePrivacy 2022 Challenge Evaluation Plan,

    N. Tomashenko, X. Wang, X. Miao, H. Nourtel, P. Champion, M. Todisco, E. Vincent, N. Evans, J. Yamagishi, and J.-F. Bonas- tre, “The V oicePrivacy 2022 Challenge Evaluation Plan,” arXiv, no. arXiv:2203.12468, Sep. 2022

    2022 arXiv

  5. [13]

    The V oicePrivacy 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 oicePrivacy 2024 Challenge Evaluation Plan,” arXiv, no. arXiv:2404.02677, Jun. 2024

    2024 arXiv

  6. [14]

    Probing the Feasibility of Multi- lingual Speaker Anonymization,

    S. Meyer, F. Lux, and N. T. Vu, “Probing the Feasibility of Multi- lingual Speaker Anonymization,” in Proc. Interspeech, 2024, pp. 4448–4452

    2024

  7. [15]

    MUSA: Multi-lingual Speaker Anonymization via Serial Disen- tanglement,

    J. Yao, Q. Wang, P. Guo, Z. Ning, Y . Yang, Y . Pan, and L. Xie, “MUSA: Multi-lingual Speaker Anonymization via Serial Disen- tanglement,” arXiv, no. arXiv:2407.11629, 2024

    2024 arXiv

  8. [16]

    Language-Independent Speaker Anonymization Approach Us- ing Self-Supervised Pre-Trained Models,

    X. Miao, X. Wang, E. Cooper, J. Yamagishi, and N. Tomashenko, “Language-Independent Speaker Anonymization Approach Us- ing Self-Supervised Pre-Trained Models,” inProc. Odyssey, 2022, pp. 279–286

    2022

  9. [17]

    Analyzing Language-Independent Speaker Anonymiza- tion Framework under Unseen Conditions,

    ——, “Analyzing Language-Independent Speaker Anonymiza- tion Framework under Unseen Conditions,” arXiv, no. arXiv:2203.14834, Mar. 2022

    2022 arXiv

  10. [18]

    mHuBERT-147: A Compact Multilingual HuBERT Model,

    M. Zanon Boito, V . Iyer, N. Lagos, L. Besacier, and I. Calapode- scu, “mHuBERT-147: A Compact Multilingual HuBERT Model,” in Proc. Interspeech, 2024, pp. 3939–3943

    2024

  11. [19]

    Wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “Wav2vec 2.0: A framework for self-supervised learning of speech representa- tions,” in Proc. NeurIPS, ser. NIPS ’20, Dec. 2020, pp. 12 449– 12 460

    2020

  12. [20]

    Generating Identi- ties with Mixture Models for Speaker Anonymization,

    H. Turner, G. Lovisotto, and I. Martinovic, “Generating Identi- ties with Mixture Models for Speaker Anonymization,” Comput. Speech Lang., vol. 72, no. C, Mar. 2022

    2022

  13. [21]

    Speaker Anonymisation Using the McAdams Coefficient,

    J. Patino, N. Tomashenko, M. Todisco, A. Nautsch, and N. Evans, “Speaker Anonymisation Using the McAdams Coefficient,” in Proc. Interspeech, 2021, pp. 1099–1103

    2021

  14. [22]

    Speaker Anonymization Using X-vector and Neural Waveform Models,

    F. Fang, X. Wang, J. Yamagishi, I. Echizen, M. Todisco, N. Evans, and J.-F. Bonastre, “Speaker Anonymization Using X-vector and Neural Waveform Models,” in 10th ISCA Workshop on Speech Synthesis (SSW 10), Sep. 2019, pp. 155–160

    2019

  15. [23]

    The V oicePrivacy 2020 Challenge: Results and findings,

    N. Tomashenko, X. Wang, E. Vincent, J. Patino, B. M. L. Srivas- tava, P.-G. No´e, A. Nautsch, N. Evans, J. Yamagishi, B. O’Brien, A. Chanclu, J.-F. Bonastre, M. Todisco, and M. Maouche, “The V oicePrivacy 2020 Challenge: Results and findings,” Comput. Speech Lang., vol. 74, no...

    2020

  16. [24]

    X-Vectors: Robust DNN Embeddings for Speaker Recogni- tion,

    D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudan- pur, “X-Vectors: Robust DNN Embeddings for Speaker Recogni- tion,” in Proc. ICASSP, 2018, pp. 5329–5333

    2018

  17. [25]

    Design Choices for X-Vector Based Speaker Anonymization,

    B. M. L. Srivastava, N. Tomashenko, X. Wang, E. Vincent, J. Ya- magishi, M. Maouche, A. Bellet, and M. Tommasi, “Design Choices for X-Vector Based Speaker Anonymization,” in Proc. Interspeech, 2020, pp. 1713–1717

    2020

  18. [26]

    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,” Comput. Speech Lang. , vol. 73, no. C, May 2022

    2022

  19. [27]

    Are Disentangled Represen- tations All You Need to Build Speaker Anonymization Systems?

    C. Pierre, A. Larcher, and D. Jouvet, “Are Disentangled Represen- tations All You Need to Build Speaker Anonymization Systems?” in Proc. Interspeech, 2022, pp. 2793–2797

    2022

  20. [28]

    Differentially Private Speaker Anonymization,

    A. Shahin Shamsabadi, B. Mohan Lal Srivastava, A. Bellet, N. Vauquier, E. Vincent, M. Maouche, M. Tommasi, and N. Paper- not, “Differentially Private Speaker Anonymization,”Proc. PETS, vol. 2023, no. 1, pp. 98–114, Jan. 2023

    2023

  21. [29]

    X-Vector Anonymization Using Autoen- coders and Adversarial Training for Preserving Speech Privacy,

    J. M. Perero-Codosero, F. M. Espinoza-Cuadros, and L. A. Hern´andez-G´omez, “X-Vector Anonymization Using Autoen- coders and Adversarial Training for Preserving Speech Privacy,” Comput. Speech Lang., vol. 74, no. C, Jul. 2022

    2022

  22. [30]

    Yet Another Algorithm for Pitch Tracking,

    K. Kasi and S. A. Zahorian, “Yet Another Algorithm for Pitch Tracking,” in Proc. ICASSP, 2002, pp. 361–364

    2002

  23. [31]

    Anonymizing Speech with Generative Adversarial Networks to Preserve Speaker Privacy,

    S. Meyer, P. Tilli, P. Denisov, F. Lux, J. Koch, and N. T. Vu, “Anonymizing Speech with Generative Adversarial Networks to Preserve Speaker Privacy,” in2022 IEEE Spoken Language Tech- nology Workshop (SLT), 2022, pp. 912–919

    2022

  24. [32]

    System description for V oice Privacy Challenge 2022,

    X. Chen, G. Li, H. Huang, W. Zhou, S. Li, Y . Cao, and Y . Zhao, “System description for V oice Privacy Challenge 2022,” inProc. SPSC 2022, 2022

    2022

  25. [33]

    Any-to-One Sequence- to-Sequence V oice Conversion Using Self-Supervised Discrete Speech Representations,

    W.-C. Huang, Y .-C. Wu, and T. Hayashi, “Any-to-One Sequence- to-Sequence V oice Conversion Using Self-Supervised Discrete Speech Representations,” inProc. ICASSP, 2021, pp. 5944–5948

    2021

  26. [34]

    Speech Resynthesis from Discrete Disentangled Self-Supervised Representations,

    A. Polyak, Y . Adi, J. Copet, E. Kharitonov, K. Lakhotia, W.-N. Hsu, A. Mohamed, and E. Dupoux, “Speech Resynthesis from Discrete Disentangled Self-Supervised Representations,” in Proc. Interspeech, Aug. 2021, pp. 3615–3619

    2021

  27. [35]

    was used for both the ASR experiments and ASV experi- ments. We split the utterances in the test set into test trial (88 ut- terances) and enrollment (4,179 utterances) subsets, which pro- duced 10,120 enrollment-test pairs for ASV evaluation. 3https://github.com/facebookresea...

  28. [36]

    Speaker Anonymization Using Orthogonal Householder Neural Network,

    X. Miao, X. Wang, E. Cooper, J. Yamagishi, and N. Tomashenko, “Speaker Anonymization Using Orthogonal Householder Neural Network,” IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, vol. 31, pp. 3681–3695, 2023

    2023

  29. [37]

    A Comparison of Discrete and Soft Speech Units for Improved V oice Conversion,

    B. van Niekerk, M.-A. Carbonneau, J. Za ¨ıdi, M. Baas, H. Seut ´e, and H. Kamper, “A Comparison of Discrete and Soft Speech Units for Improved V oice Conversion,” in Proc. ICASSP, 2022, pp. 6562–6566

    2022

  30. [38]

    HuBERT: Self-Supervised Speech Rep- resentation Learning by Masked Prediction of Hidden Units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “HuBERT: Self-Supervised Speech Rep- resentation Learning by Masked Prediction of Hidden Units,” IEEE/ACM Trans. Audio, Speech and Lang. , vol. 29, pp. 3451– 3460, Oct. 2021

    2021

  31. [39]

    ECAPA- TDNN: Emphasized Channel Attention, Propagation and Ag- gregation in TDNN Based Speaker Verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA- TDNN: Emphasized Channel Attention, Propagation and Ag- gregation in TDNN Based Speaker Verification,” in Proc. Inter- speech, 2020, pp. 3830–3834

    2020

  32. [40]

    HiFi-GAN: Generative Adversar- ial Networks for Efficient and High Fidelity Speech Synthesis,

    J. Kong, J. Kim, and J. Bae, “HiFi-GAN: Generative Adversar- ial Networks for Efficient and High Fidelity Speech Synthesis,” in Advances in Neural Information Processing Systems , vol. 33, 2020, pp. 17 022–17 033

    2020

  33. [41]

    Lib- rispeech: An ASR corpus based on public domain audio books,

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

    2015

  34. [42]

    Corpus of Spontaneous Japanese: Its Design and Evaluation,

    K. Maekawa, “Corpus of Spontaneous Japanese: Its Design and Evaluation,” in Proc. SSPR 2003, 2003, p. paper MMO2

    2003

  35. [43]

    JVS Corpus: Free Japanese Multi-Speaker V oice Corpus,

    S. Takamichi, K. Mitsui, Y . Saito, T. Koriyama, N. Tanji, and H. Saruwatari, “JVS Corpus: Free Japanese Multi-Speaker V oice Corpus,” arXiv, no. arXiv:1908.06248, Aug. 2019

    1908 arXiv

  36. [44]

    JTubeSpeech: Corpus of Japanese speech collected from YouTube for speech recognition and speaker verification,

    S. Takamichi, L. K ¨urzinger, T. Saeki, S. Shiota, and S. Watan- abe, “JTubeSpeech: Corpus of Japanese speech collected from YouTube for speech recognition and speaker verification,” arXiv, no. arXiv:2112.09323, Dec. 2021

    2021 arXiv

  37. [45]

    AISHELL-3: A Multi-Speaker Mandarin TTS Corpus,

    Y . Shi, H. Bu, X. Xu, S. Zhang, and M. Li, “AISHELL-3: A Multi-Speaker Mandarin TTS Corpus,” in Proc. Interspeech, 2021, pp. 2756–2760

    2021

  38. [46]

    V ocoder Drift in X-Vector–Based Speaker Anonymization,

    M. Panariello, M. Todisco, and N. Evans, “V ocoder Drift in X-Vector–Based Speaker Anonymization,” in Proc. Interspeech, 2023, pp. 2863–2867

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.