Pith. sign in

REVIEW 3 major objections 6 minor 37 references

SecureSpeech: Prompt-based Speaker and Content Protection

T0 review · 3 major / 6 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 This paper claims that a prompt-based ASR-NER-LLM-TTS pipeline can anonymize both who is speaking and what is said, replacing the speaker with a text-described pseudo-voice and swapping recognized named entities for generic alternatives…

desk verdict A competent, honest assembly of off-the-shelf speech-privacy tools; the speaker half is near-tautological and the content half is under-measured, but the paper is clearly written and worth a serious referee. read the letter →

arxiv 2507.07799 v1 pith:SRDEZH6W submitted 2025-07-10 cs.SD eess.AS

Belinda Soh Hui Hui , Xiaoxiao Miao , Xin Wang This is my paper
classification cs.SDeess.AS
keywords speechanonymizationspeakerprivacycontentnamedentityrecognitionprompt-basedtext-to-speechlargelanguagemodelvoiceASR-NER-TTSpipeline
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

The paper argues that speech privacy should cover both who is speaking and what is said, and that both can be handled in one prompt-based pipeline. It transcribes the input, uses named-entity recognition plus a language model to replace sensitive entities with generic alternatives, and synthesizes a new voice from a text description instead of from the original speaker's voice. On the SLUE-VoxPopuli benchmark, the pipeline reports a false acceptance rate of 0% under an automatic speaker-verification attacker, a word error rate of 15.50%, and a predicted mean opinion score of 4.01. The point of the work is that the anonymized voice is unlinkable to the source speaker while the sanitized text stays intelligible, and that users can control the pseudo-speaker through natural-language descriptions.

What carries the argument

The load-bearing component is the pairing of a text-based speaker description with a prompt-conditioned TTS model (Parler-TTS): because the generated voice is conditioned on words rather than on the source utterance, no encoder-derived speaker representation of the original speaker enters the synthesis path, so the output voiceprint is not tied to the input. The content side is carried by the NER-to-LLM replacement step, where the NER model (DeBERTa-L) tags entities and the LLM (Llama-3.2) rewrites them into generic alternatives under a structured prompt. Together they turn a raw utterance into (description, sanitized transcript) pairs that feed the TTS decoder.

What would settle it

Transcribe anonymized outputs and have an independent language-model-based attacker try to infer the source speaker from the remaining semantics, including details that are not named entities (such as a unique biographical fact or an untagged number); if the attacker succeeds while WER stays low, the content-anonymization claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a speech-to-speech pipeline built from off-the-shelf ASR, NER, an open-weight language model, and a prompt-conditioned text-to-speech model can achieve dual anonymization: the generated utterance sounds like a person who is not the source speaker, and the spoken content no longer contains the recognized sensitive entities. Speaker identity is concealed by ignoring the original voice entirely and synthesizing from a text prompt that describes a pseudo-speaker; content is concealed by detecting named entities in the ASR transcript and prompting the language model to replace them with generic alternatives. The paper reports that the approach obtains a 0% false acceptance rate against an ECAPA-TDNN speaker-verification attacker, 15.50% WER and 4.01 PMOS, and a named-entity replacement accuracy of 99.95% given correct detection, with NER F1 of 71.80% on the predicted transcription. It also claims to be the first to use prompt-based speaker-attribute editing for both speaker and content anonymization, and presents an analysis of how accent, gender, pitch, speaking rate, and channel-condition descriptions affect utility.

Load-bearing premise

The system assumes that removing the named entities that the NER model detects is enough to remove sensitive content; anything sensitive that falls outside those entities, or is missed (the reported F1 is 71.80%), stays in the output and can re-identify the speaker.

Editorial extensions

If this is right

  • An attacker using an ECAPA-TDNN speaker-verification model cannot link the anonymized utterance to the enrollment speech of the source speaker, because the synthesized voice is drawn from a text-described pseudo-speaker rather than from the source utterance.
  • When the NER model tags an entity, the language model replaces it with a generic alternative with 99.95% accuracy, so the removed words no longer appear in the transcript that is read aloud.
  • The anonymized speech keeps acceptable intelligibility and quality on the SLUE-VoxPopuli set, with WER of 15.50% and PMOS of 4.01, comparable to zero-shot TTS baselines that condition on the source speaker's audio.
  • Users can steer the anonymized voice by editing natural-language attributes; the reported utility varies with the description, e.g., the Slovak and Australian accents give the lowest WER while the Italian accent gives the highest (23.76%).
  • Because the pipeline never uses the source speaker's voice in synthesis, the original voiceprint cannot be stolen from the output for voice cloning.

Reading between the lines

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

  • A stricter content-privacy test would use semantic re-identification attacks rather than WER, since the paper itself notes WER does not capture context or reasoning that can still reveal identity.
  • The same speaker description can sound like different people across generations, so for applications needing a fixed anonymized persona one would add an explicit speaker specification or a deterministic voice-locking mechanism.
  • The prompt-based setup makes demographic bias in TTS measurable: one can systematically vary one attribute at a time (as the paper does) and flag combinations that hurt intelligibility, such as very fast speaking rate or certain accents.
  • A reversible variant could keep a private mapping between original and anonymized speaker/content for forensic use, at the cost of weakening the privacy guarantee if that mapping leaks.
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 / 6 minor

Summary. The paper proposes SecureSpeech, a pipeline for dual anonymization of speech: it transcribes audio with an ASR model, detects and replaces named entities with an NER model and Llama-3.2, synthesizes speech with a prompt-based TTS model (Parler-TTS) conditioned on a user-supplied speaker description, and evaluates speaker privacy (FAR), content privacy (WER, NER F1, replacement accuracy), and utility (WER, PMOS) on SLUE-VoxPopuli. The authors report FAR=0%, WER 15.50%, and PMOS 4.01 for their pipeline, and compare against VALL-E and XTTS-v2. They also analyze how varying speaker-description attributes (accent, gender, pitch, speaking rate, channel condition) affect WER and PMOS. The central claim is that the pipeline, by not conditioning on the source voice and by replacing recognized named entities, protects both speaker identity and spoken content while maintaining intelligibility and quality.

Significance. If the dual anonymization claim were fully established, the paper would be a useful contribution to speech privacy: it demonstrates an off-the-shelf, prompt-controllable pipeline, ships code and audio samples, and provides guidance on how speaker descriptions affect downstream utility. The attribute-ablation analysis (Tables 2 and 3) is a practical addition. However, the load-bearing evidence for content anonymization is missing, and the speaker-privacy metric is near-trivially satisfied by construction. The paper's own Section 4 concedes several of these gaps, but the conclusions in the abstract and Section 5 are not yet supported by the experiments.

major comments (3)
  1. [§3.2.1, §3.4.1] The WER metric cannot support the content-privacy claim. In §3.4.1 the authors state that when calculating the WER of anonymized speech, they use the transcripts with sensitive content removed as the ground truth. This means WER measures how faithfully the TTS reproduces the sanitized prompt, not whether sensitive content was removed from the original audio. A low WER can be achieved even if many named entities were missed by NER and remain unremoved; it is a utility metric, not a privacy metric. Section 4(iii) explicitly acknowledges that "the conventional WER metric is insufficient to evaluate content privacy protection," which should be reflected in the main evaluation. The paper needs a direct leakage measurement, such as running NER on the synthesized speech and comparing detected entities against the original gold entities, or a content-based re-identification attack.
  2. [§3.4.1, §4(i)-(iii)] The cumulative errors of the ASR and NER stages undermine the content-anonymization claim. The ASR WER of 19.00% and NER F1 of 71.80% on predicted transcriptions mean that roughly 28% of gold named entities are not recognized and therefore are not replaced (or the transcript itself is wrong). Because the final synthetic speech is generated from the sanitized text, any unrecognized entity survives into the output. The paper acknowledges this limitation in Section 4 items (i) and (ii), but the abstract and conclusion claim that sensitive content is removed to a degree that provides privacy. To support the dual-anonymization claim, the authors should quantify the fraction of gold sensitive entities that survive into the final synthetic speech. Without such a measurement, the content half of the central claim is not established.
  3. [§3.4.1, Table 1] The speaker-privacy evaluation is not a meaningful test of the proposed method. As the paper itself notes, the TTS pipeline is trained on data from speakers disjoint from those in SLUE-VoxPopuli, and no source speaker information is used, so FAR=0% is expected for any TTS system in this configuration. The comparison with VALL-E and XTTS-v2 is also confounded: the paper does not specify what reference audio is used for these zero-shot systems, nor why FAR=0% would be non-trivial for them. A meaningful speaker-privacy evaluation should at least include a scenario in which the source speaker is a plausible match (e.g., a TTS model with overlapping speakers, or an attack that uses prosody or content cues) and should report additional metrics such as EER or attack success under different thresholds. As written, Table 1 does not support the claim that the prompt-based design contributes to speaker privacy over simpler baselines.
minor comments (6)
  1. [References] Reference [24] is cited for Llama-3.2, but the reference listed is Petroni et al., "Language models as knowledge bases?". The authors should cite the actual Llama-3.2 model paper or its official documentation.
  2. [Author affiliation] There is a typo in the first author's affiliation: "Sinagpore Institute of Technology" should be "Singapore Institute of Technology".
  3. [Throughout] The model name "VALL-E" is inconsistently formatted as "V ALL-E" and "V ALLE-X" across the text, tables, and figure captions. Please standardize.
  4. [§3.2.1, §3.2.2] WER appears as a content-privacy metric in §3.2.1 and as a utility metric in §3.2.2 with the same definition. The distinction is confusing and should be clarified: the content-privacy role of WER is not defensible, and renaming or removing it from the privacy block would help.
  5. [§3.4.1] The "named entity replacement accuracy" of 99.95% is not defined. It is unclear whether this is measured on recognized entities only or on gold entities, and how errors are counted. A precise definition is needed.
  6. [§4] In the last paragraph of Section 4, "disscussed" should be "discussed".

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pipeline is evaluated with external component metrics, and the expected FAR=0 result is explicitly acknowledged as a construction artifact rather than a derived prediction.

full rationale

This is an empirical system paper with no fitted parameters or analytic derivation chain that could reduce to its own inputs. The content-anonymization claim rests on separately evaluated off-the-shelf components: ASR WER 19.00%, NER F1 71.80%, and named-entity replacement accuracy 99.95% (Section 3.4.1). None of these numbers is defined in terms of the final privacy claim; they are measured against external benchmarks and model outputs. The paper explicitly concedes the limitations of its content-privacy evaluation in Section 4, noting that WER 'does not account for context, long-term relationships, or reasoning that may still reveal the speaker's identity' and that the NER assumption is 'a narrow definition.' These are honest limitations, not circular reasoning. The speaker-privacy result FAR=0 is also not presented as a surprise or as a consequence of a fitted model: the paper states, 'This is not surprising because the TTS pipeline trained on data of speakers who are disjoint from the target speakers in the evaluation set. The anonymized voice produced by TTS models is expected to be different from the target speakers.' Thus the 0% FAR is an acknowledged expected consequence of using a TTS with disjoint training speakers, not a hidden tautological prediction. The only self-citations (e.g., VoicePrivacy challenge references for evaluation conventions) are used for standard FAR/EER definitions and prior anonymization baselines, which are externally established benchmarks rather than load-bearing unverified assertions. No derivation, equation, or stated claim reduces by construction to its own input. The paper is self-contained as an empirical evaluation and its central claims have independent content.

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

No constants are fitted in this paper; all model weights come from public checkpoints. The central claims depend on domain assumptions about what counts as sensitive content, on the attacker model for FAR, and on the disjointness of Parler-TTS training speakers from the evaluation set. No new physical or conceptual entities are introduced.

assumptions (4)
  • domain assumption Sensitive content is assumed to be limited to named entities recognized by the NER model.
    Stated in footnote 3 and discussed in Section 4; the pipeline replaces only NER-detected entities, so any sensitive content outside this class is not removed.
  • domain assumption Privacy can be measured by FAR against an ECAPA-TDNN attacker with an EER threshold, using target-speaker enrollment.
    Section 3.2.1 defines this evaluation; it is a standard simplification but does not model attackers who know the pipeline or link multiple anonymized utterances.
  • domain assumption The TTS model's training speakers are disjoint from SLUE-VoxPopuli speakers, so generated voices are not expected to match enrolled targets.
    Acknowledged in Section 3.4.1; the FAR 0% result depends on this external model property rather than a pipeline-specific mechanism.
  • domain assumption The pre-trained ASR, NER, LLM, and TTS checkpoints behave as expected for the test data.
    The paper relies on off-the-shelf public models (Sections 2.1 and 2.2) without independent verification beyond the reported metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SecureSpeech: Prompt-based Speaker and Content Protection." pith.science (2026). https://pith.science/paper/SRDEZH6W

@misc{pith2026250707799,
  author       = {Pith},
  title        = {Pith review of: SecureSpeech: Prompt-based Speaker and Content Protection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SRDEZH6W}},
  note         = {Machine review of arXiv:2507.07799}
}
read the original abstract

Given the increasing privacy concerns from identity theft and the re-identification of speakers through content in the speech field, this paper proposes a prompt-based speech generation pipeline that ensures dual anonymization of both speaker identity and spoken content. This is addressed through 1) generating a speaker identity unlinkable to the source speaker, controlled by descriptors, and 2) replacing sensitive content within the original text using a name entity recognition model and a large language model. The pipeline utilizes the anonymized speaker identity and text to generate high-fidelity, privacy-friendly speech via a text-to-speech synthesis model. Experimental results demonstrate an achievement of significant privacy protection while maintaining a decent level of content retention and audio quality. This paper also investigates the impact of varying speaker descriptions on the utility and privacy of generated speech to determine potential biases.

Figures

Figures reproduced from arXiv: 2507.07799 by the authors.

Figure 1
Figure 1. Speech data flow comparison: a typical method (left) and the proposed prompt-based speaker and content anonymization method [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of User Prompt Used in Proposed Pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Sorted WER (top) and sorted PMOS (bottom) for differ [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 31 canonical work pages

  1. [1]

    Alharthi, R

    D. Alharthi, R. S. Sharma, H. Dhamyal, S. Maiti, B. Raj, and R. Singh. Evaluating speech synthesis by training recogniz- ers on synthetic speech. In Synthetic Data’s Transformative Role in Foundational Speech Models, pages 66–70, 2024

    work page 2024

  2. [2]

    S. Arik, J. Chen, K. Peng, W. Ping, and Y . Zhou. Neural voice cloning with a few samples. Advances in neural infor- mation processing systems, 31, 2018

    work page 2018

  3. [3]

    Baevski, H

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli. wav2vec 2.0: a framework for self-supervised learning of speech rep- resentations. 2020

    work page 2020

  4. [4]

    People are poorly equipped to detect AI-powered voice clones

    S. Barrington and H. Farid. People are poorly equipped to detect ai-powered voice clones. arXiv preprint arXiv:2410.03791, 2024

    work page Pith review arXiv 2024

  5. [5]

    S. M. Bsharat, A. Myrzakhan, and Z. Shen. Principled in- structions are all you need for questioning llama-1/2, gpt- 3.5/4. arXiv preprint arXiv:2312.16171, 2023

    arXiv 2023

  6. [6]

    Casanova, K

    E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea, L. Hart, A. Aljafari, J. Meyer, R. Morais, S. Olayemi, and J. Weber. Xtts: a massively multilingual zero-shot text-to- speech model. In Interspeech 2024, pages 4978–4982, 2024

    work page 2024

  7. [7]

    Champion

    P. Champion. Anonymizing Speech: Evaluating and Design- ing Speaker Anonymization Techniques. PhD thesis, 2023

    work page 2023

  8. [8]

    C.-F. R. Chen, Q. Fan, and R. Panda. CrossViT: Cross- Attention Multi-Scale Vision Transformer for Image Clas- sification. In International Conference on Computer Vision (ICCV), 2021

    work page 2021

Show all 37 references
  1. [9]

    Desplanques, J

    B. Desplanques, J. Thienpondt, and K. Demuynck. ECAPA- TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification. In Proc. Interspeech, pages 3830–3834, 2020

    2020

  2. [10]

    F. Fang, X. Wang, J. Yamagishi, I. Echizen, M. Todisco, N. Evans, and J.-F. Bonastre. Speaker anonymization us- ing x-vector and neural waveform models. Proc. 10th ISCA Speech Synthesis Workshop, pages 155–160, 9 2019

    2019

  3. [11]

    A. Hannun. Sequence modeling with ctc. Distill, 2017. https://distill.pub/2017/ctc

    2017

  4. [12]

    Y . Hu, R. Li, S. Wang, F. Tao, and Z. Sun. Speechhide: A hybrid privacy-preserving mechanism for speech content and voiceprint in speech data sharing. In 2022 7th IEEE Inter- national Conference on Data Science in Cyberspace (DSC), pages 345–352. IEEE, 2022

    2022

  5. [13]

    Justin, V

    T. Justin, V . ˇStruc, S. Dobri ˇsek, B. Vesnicer, I. Ip ˇsi´c, and F. Miheliˇc. Speaker de-identification using diphone recogni- tion and speech synthesis. In 2015 11th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition (FG), volume 4, pages...

    2015

  6. [14]

    Kumar, K

    A. Kumar, K. Tan, Z. Ni, P. Manocha, X. Zhang, E. Hender- son, and B. Xu. Torchaudio-squim: Reference-less speech quality and intelligibility measures in torchaudio. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 1–...

    2023

  7. [15]

    Kumar, P

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Ku- mar. High-fidelity audio compression with improved rvqgan, 2023

    2023

  8. [16]

    Lacombe, V

    Y . Lacombe, V . Srivastav, and S. Gandhi. Parler-tts.https: //github.com/huggingface/parler-tts, 2024

    2024

  9. [17]

    Lyth and S

    D. Lyth and S. King. Natural language guidance of high- fidelity text-to-speech with synthetic annotations. arXiv preprint arXiv:2402.01912, 2024

    2024 arXiv

  10. [18]

    Meyer, P

    S. Meyer, P. Tilli, P. Denisov, F. Lux, J. Koch, and N. T. Vu. Anonymizing speech with generative adversarial networks to preserve speaker privacy. In Proc. Spoken Language Tech- nology Workshop (SLT), pages 912–919, 2022

    2022

  11. [19]

    X. Miao, R. Tao, C. Zeng, and X. Wang. A benchmark for multi-speaker anonymization. IEEE Transactions on Infor- mation Forensics and Security, 2025

    2025

  12. [20]

    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) , pages 279–286, 2022

    2022

  13. [21]

    X. Miao, X. Wang, E. Cooper, J. Yamagishi, and N. Tomashenko. Speaker anonymization using orthogo- nal householder neural network. IEEE/ACM Trans. Audio, Speech, and Language Processing, 31:3681–3695, 2023

    2023

  14. [22]

    Panariello, F

    M. Panariello, F. Nespoli, M. Todisco, and N. Evans. Speaker anonymization using neural audio codec language models. In Proc. ICASSP, pages 4725–4729, 2024

    2024

  15. [23]

    Panariello, N

    M. Panariello, N. Tomashenko, X. Wang, X. Miao, P. Cham- pion, H. Nourtel, M. Todisco, N. Evans, E. Vincent, and J. Yamagishi. The voiceprivacy 2022 challenge: Progress and perspectives in voice anonymisation. IEEE/ACM Transactions on Audio, Speech, and Language Processing , 3...

    2022

  16. [24]

    Petroni, T

    F. Petroni, T. Rockt¨aschel, P. Lewis, A. Bakhtin, Y . Wu, A. H. Miller, and S. Riedel. Language models as knowledge bases? In Conference on Empirical Methods in Natural Language Processing, 2019

    2019

  17. [25]

    J. Qian, H. Du, J. Hou, L. Chen, T. Jung, and X.-Y . Li. Speech sanitizer: Speech content desensitization and voice anonymization. IEEE Transactions on Dependable and Se- cure Computing, 18(6):2631–2642, 2019

    2019

  18. [26]

    Raffel, N

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu. Exploring the lim- its of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1), Jan. 2020

    2020

  19. [27]

    Ravanelli, T

    M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cor- nell, L. Lugosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong, J.-C. Chou, S.-L. Yeh, S.-W. Fu, C.-F. Liao, E. Ras- torgueva, F. Grondin, W. Aris, H. Na, Y . Gao, R. D. Mori, and Y . Bengio. SpeechBrain: A general-p...

    2021 arXiv

  20. [28]

    A. S. Shamsabadi, B. M. L. Srivastava, A. Bellet, N. Vauquier, E. Vincent, M. Maouche, M. Tommasi, and N. Papernot. Differentially private speaker anonymization. Proceedings on Privacy Enhancing Technologies , 2023(1), Jan. 2023

    2023

  21. [29]

    S. Shon, A. Pasad, F. Wu, P. Brusco, Y . Artzi, K. Livescu, and K. J. Han. Slue: New benchmark tasks for spo- ken language understanding evaluation on natural speech. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages ...

    2022

  22. [30]

    Tomashenko, X

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

    2024 arXiv

  23. [31]

    Tomashenko, X

    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, et al. The V oicePrivacy 2020 challenge: Results and findings. Computer Speech & Language, 2022

    2020

  24. [32]

    C. S. Veerappan, P. Dhingra, D. Z. Wang, and R. Tong. Speedf - a speech de-identification framework. In TENCON 2024 - 2024 IEEE Region 10 Conference (TENCON), pages 31–34, 2024

    2024

  25. [33]

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li, et al. Neural codec lan- guage models are zero-shot text to speech synthesizers.arXiv preprint arXiv:2301.02111, 2023

    2023 arXiv

  26. [34]

    Williams, J

    J. Williams, J. Yamagishi, P.-G. No ´e, C. V . Botinhao, and J.-F. Bonastre. Revisiting speech content privacy. arXiv preprint arXiv:2110.06760, 2021

    2021 arXiv

  27. [35]

    H. L. Xinyuan, Z. Cai, A. Garg, K. Duh, L. P. Garc´ıa-Perera, S. Khudanpur, N. Andrews, and M. Wiesner. Hltcoe jhu sub- mission to the voice privacy challenge 2024. In 4th Sym- posium on Security and Privacy in Speech Communication , pages 61–66, 2024

    2024

  28. [36]

    J. Yao, Q. Wang, P. Guo, Z. Ning, Y . Yang, Y . Pan, and L. Xie. Musa: Multi-lingual speaker anonymization via se- rial disentanglement. IEEE Transactions on Audio, Speech and Language Processing, 33:1664–1674, 2025

    2025

  29. [37]

    Zhang, L

    Z. Zhang, L. Zhou, C. Wang, S. Chen, Y . Wu, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li, et al. Speak foreign lan- guages with your own voice: Cross-lingual neural codec lan- guage modeling. arXiv preprint arXiv:2303.03926, 2023

    2023 arXiv

Pith tools

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