Pith. sign in

REVIEW 4 major objections 4 minor 33 references

Content Anonymization for Privacy in Long-form Audio

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Voice anonymization alone fails to protect speakers in long-form audio because linguistic style leaks identity; adding a paraphrasing step closes the gap.

desk verdict A real content-side vulnerability in long-form voice anonymization, and a plausible LLM paraphrase defense—but the defense is tested only against one uninformed, off-domain attacker, so the 'ensure anonymity' recommendation is stronger than the evidence. read the letter →

arxiv 2510.12780 v3 pith:WOAXKZYS submitted 2025-10-14 cs.SD cs.CL

classification cs.SDcs.CL
keywords voiceanonymizationspeechprivacycontent-basedattacksauthorshipattributionlong-formaudioparaphrasingASR-TTSpipelinespeakerre-identification
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 current voice anonymization, which hides acoustic identity, is insufficient for long-form audio such as phone calls, interviews, and meetings: once many utterances from one speaker are available, an attacker can re-identify the speaker from vocabulary, syntax, and phrasing alone. To close this 'content gap', the authors insert a paraphrasing step into an automatic speech recognition-to-text-to-speech pipeline, rewriting transcripts to remove speaker-specific style while preserving meaning. They show that a content-based attacker succeeds against voice-only anonymization and improves as more utterances are aggregated, and that adding paraphrase-based content anonymization flattens the attack to chance-level accuracy. The paper recommends that stakeholders adopt contextual paraphrasing as a standard step in long-form audio anonymization.

What carries the argument

The load-bearing component is a contextual paraphrasing module inserted between automatic speech recognition and text-to-speech. Instead of rewriting utterance-by-utterance, it processes a sliding window of several utterances jointly, using the preceding conversational context, and is prompted to alter utterance length and structure as well as wording. This rewriting is what removes speaker-specific style; the rest of the pipeline handles the voice. The evaluation pairs this with a content-based attribution attack on transcripts under a topic-controlled 'hard' setting, measuring equal error rate as the number of utterances grows.

What would settle it

Train a content-based attribution model on paraphrased (and re-transcribed) transcripts of known speakers, or fine-tune it on anonymized trials, then measure its equal error rate on the hard topic-controlled setting; if the attack accuracy rises well above chance as utterances are aggregated, the paraphrase defense is not robust to an informed attacker.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that linguistic content is a biometric side-channel that survives voice anonymization: a content-only attacker applied to transcripts of voice-anonymized speech re-identifies speakers with growing accuracy as the number of available utterances increases. When the transcripts are first rewritten by a contextual paraphraser that operates on a window of multiple utterances and is prompted to condense, rephrase, and remove personal identifiers, the content attack's accuracy drops to roughly chance and stays flat even at full-conversation length. The same paraphrase-plus-synthesis pipeline also defeats voice-based verification when the synthesize

Load-bearing premise

The recommendation rests on the assumption that the attacker does not adapt to the anonymization method; if an attacker trains on paraphrased or synthesized transcripts, the paraphrase defense may no longer hold.

Editorial extensions

If this is right

  • Long-form audio anonymization needs a content-rewriting step; voice conversion alone leaves speakers re-identifiable from style.
  • The more utterances an attacker can aggregate, the stronger the content-based attack, so privacy protection should be evaluated across utterance counts, not on isolated clips.
  • Small, locally run language models can come close to large API-based models for this paraphrasing task, making on-device private anonymization feasible.
  • Synthesizing and re-transcribing paraphrased content masks some machine-generated style, reducing detectability of the anonymized speech.

Reading between the lines

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

  • If paraphrasing becomes a standard defense, attackers will likely adapt by training content-attribution models on paraphrased or synthesized speech; the paper's own limitations note that semi-informed settings may require different strategies.
  • The same content-attack risk should apply to other long-form audio settings (meetings, podcasts, medical or legal recordings), so the recommended paraphrase step may generalize beyond telephone conversations.
  • A testable extension would be to run the pipeline on non-English conversational speech or on spontaneous multi-speaker meetings, where context windows and transcription errors differ.
  • The privacy-utility metrics introduced here (attacker error rate plus semantic similarity, naturalness, and detectability) could become a template for benchmarking content anonymization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper argues that existing voice anonymization, which is typically evaluated at the utterance level, is insufficient for long-form audio because linguistic content acts as a biometric side-channel. To address this, the authors propose content anonymization via LLM-based paraphrasing (utterance-level with GPT-4o-mini, or segment-level with Gemma-3-4B and GPT-5) inserted into an ASR-TTS pipeline. Experiments on the Fisher corpus show that a content-only attacker (LUAR) achieves lower EER as more utterances are aggregated when only the voice is anonymized, while adding content paraphrasing flattens the content-attack curve to near 50% EER. The paper recommends that stakeholders adopt paraphrasing to ensure anonymity in long-form audio.

Significance. If the central claim holds, this is a valuable and timely result: it highlights a vulnerability overlooked by utterance-level privacy benchmarks and offers a practical defense. The use of public datasets and standard metrics (EER, UTMOS, Binoculars) is a strength, and the paper is transparent about the ASR-TTS pipeline and the specific attack models used. However, the generalizability of the privacy claim rests on a single, domain-mismatched, uninformed attacker model, and the over-strong recommendation in the abstract ('ensure anonymity') goes beyond the evidence. The paper also ships no actual code or prompts, though it promises to release them.

major comments (4)
  1. [§6.1, §5.2] The central recommendation ('ensure anonymity') is not supported by the experimental scope. The content attack is evaluated only with LUAR, trained on Reddit comments and applied to Fisher telephone transcripts (§5.2), and only in the uninformed attacker setting (§3). The paper itself concedes in §6.1 that 'content-based attacks may prove even more effective in semi-informed attacker settings... and may in turn require different anonymization strategies.' This is a load-bearing caveat: if a stronger attacker (e.g., one fine-tuned on paraphrased transcripts or using a domain-matched stylometric model) defeats the defense, the observed flattening to ~0.5 EER would not hold. The authors should either temper the recommendation to the specific evaluated setting or add experiments with additional attack models, including semi-informed variants.
  2. [§5.1, Fig. 1] The key empirical claims—that content-attack EER decreases with more utterances and that paraphrasing flattens this curve—are made from point estimates without confidence intervals or statistical tests. Given the relatively small number of trials (1944 total, from [4]) and the fact that the 'full' condition shows EER exceeding 50% (which itself remains unexplained), the authors should report bootstrap CIs, per-trial variability, or significance tests to support the claimed flattening. Without this, it is difficult to distinguish a real effect from noise.
  3. [§4, §5.1] The anonymization method depends critically on the prompt instructions ('The formulation of these prompts... is a critical factor'), yet the actual prompts are omitted and only promised for future release. This makes the method irreproducible and prevents readers from assessing what specific textual transformations drive the privacy improvement. Please include the full prompt templates in an appendix or supplementary material.
  4. [§5.4] The claim that voice+content anonymization yields an EER 'exceeding 50%' at the full-conversation condition is anomalous. EER above 50% implies the attacker is systematically inverting its scores, which is not equivalent to chance and suggests a possible artifact in the trial construction or in the LUAR scoring after paraphrase. This should be investigated and explained, or the result should be capped at 50% with an appropriate discussion.
minor comments (4)
  1. [§5] The text says 'The attacker’s performance improves (higher EER)' — this is backwards. For an error rate, attacker improvement corresponds to lower EER. The later §5.4 correctly states 'lowers EER'. Please correct.
  2. [§2] Minor grammar: 'they does not address' should be 'it does not address'.
  3. [§5.1] The experimental setup relies on the 'hard' setting from [4], but the paper does not specify the number of speakers, the utterance segmentation procedure, or how 'full' is defined. Please provide these details to make the results reproducible.
  4. [§5.5] Reference [26] is about diarization and does not appear directly relevant to ASR error propagation in the pipeline. Please check or provide a more appropriate citation.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation; the content-attack threat is re-run, not assumed, and the defense is measured against a fixed public attacker model.

full rationale

The paper's central claim—paraphrasing in an ASR-TTS pipeline mitigates content-based re-identification in long-form audio—is supported by a direct experiment, not by a definition or a fitted parameter. The threat is first demonstrated in Fig. 1 by applying the LUAR content-attack model to Fisher transcripts, so the paper does not rely solely on citations [4,5] to establish that linguistic content is a biometric channel. The attacker model LUAR ([14,20]) is a fixed, publicly released artifact (HuggingFace LUAR-MUD) and is not retrained on the anonymized trials; the uninformed-attacker setting (§3) explicitly excludes access to anonymized data. The proposed content anonymization is evaluated by EER against this fixed model and by utility metrics; no equation equates the output to the input. The self-citations to the authors' prior work are real external publications and the empirical setup is re-executed in this paper, so they are not load-bearing in a circular sense. The §6.1 limitation ('content-based attacks may prove even more effective in semi-informed attacker settings where anonymized data is available to improve the attack model') is a scope restriction, not a circular step. The only mild self-referential aspect is that the same group built both the attack model and the defense, which is a normal evaluation choice and not a reduction to inputs. Overall: no significant circularity; score 1 reflects the minor in-house origin of the attacker model rather than any definitional circularity.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The paper contributes an empirical pipeline evaluation; its load-bearing premises are almost all domain assumptions inherited from prior work rather than derived here. Most consequential: (i) the uninformed-attacker threat model, (ii) the transfer validity of a Reddit-trained authorship model to telephone transcripts, (iii) the assumption that ASR-TTS exactly removes acoustic identity, and (iv) that SBERT-alignment similarity validly measures meaning preservation. Hand-chosen control settings (window size, segment size, preservation ratio, embedding composition, prompt instructions) shape every reported number and are not fitted but are also not swept or ablated. No invented entities: everything used is an existing system.

free parameters (5)
  • Context window size N = 8 previous utterances
    Set by hand in §4 to contextualize paraphrases; affects both privacy and fluency of rewrites.
  • Segment size for segment-based paraphrasing = 16 utterances or ≈300 tokens
    Hand-chosen in §4; defines the unit of joint rewriting that the authors argue is essential vs utterance-by-utterance.
  • Gemma3-4Bc utterance preservation ratio = 50% (half of original utterances preserved)
    Conservative variant defined in §5.4; changes privacy vs detectability tradeoff.
  • Pseudo target-speaker embedding composition = randomly weighted combination of 5–6 VoxCeleb2 speakers' longest utterances
    Hand-chosen in §5.1 to synthesize a consistent pseudo identity; affects TTS output voice.
  • Prompt instruction set
    Instructions to condense, change utterance length, replace PII with gender-aligned fictions (§5.4); the paper calls prompt formulation 'a critical factor' (§4) but does not reproduce the prompts.
assumptions (7)
  • domain assumption Speaker identity is recoverable from transcript content alone, even with ASR errors
    Invoked via refs [4],[5] in §2; the premise of the entire attack side.
  • domain assumption The uninformed attacker (no access to anonymized trials) is the relevant threat model
    §3 'we focus on the uninformed attacker setting'; paper's own §6.1 limits generalizability.
  • domain assumption LUAR, trained on Reddit comments, is a strong content-attribution model on telephone transcripts
    §5.2 selects LUAR as 'a top performing authorship attribution model'; cross-domain transfer is assumed, not measured.
  • domain assumption Whisper→XTTS pipeline completely removes the source speaker's acoustic identity
    §5/Fig 1a use high voice-attack EER as evidence; residual voiceprint leakage is not quantified.
  • standard math EER of 50% equals perfect anonymization
    §3, standard definition from VoicePrivacy [1].
  • domain assumption SBERT-embedding greedy/DTW alignment validly measures semantic preservation
    §5.3; used to claim utility retention at GAS 0.65–0.70.
  • domain assumption Binoculars and SSL-AASIST provide valid zero-shot machine-content detection
    §5.3, §5.5; used to argue anonymized speech is not machine-identifiable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Content Anonymization for Privacy in Long-form Audio." pith.science (2026). https://pith.science/paper/WOAXKZYS

@misc{pith2026251012780,
  author       = {Pith},
  title        = {Pith review of: Content Anonymization for Privacy in Long-form Audio},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WOAXKZYS}},
  note         = {Machine review of arXiv:2510.12780}
}
read the original abstract

Voice anonymization techniques have been found to successfully obscure a speaker's acoustic identity in short, isolated utterances in benchmarks such as the VoicePrivacy Challenge. In practice, however, utterances seldom occur in isolation: long-form audio is commonplace in domains such as interviews, phone calls, and meetings. In these cases, many utterances from the same speaker are available, which pose a significantly greater privacy risk: given multiple utterances from the same speaker, an attacker could exploit an individual's vocabulary, syntax, and turns of phrase to re-identify them, even when their voice is completely disguised. To address this risk, we propose a new approach that performs a contextual rewriting of the transcripts in an ASR-TTS pipeline to eliminate speaker-specific style while preserving meaning. We present results in a long-form telephone conversation setting demonstrating the effectiveness of a content-based attack on voice-anonymized speech. Then we show how the proposed content-based anonymization methods can mitigate this risk while preserving speech utility. Overall, we find that paraphrasing is an effective defense against content-based attacks and recommend that stakeholders adopt this step to ensure anonymity in long-form audio.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 3 linked inside Pith

  1. [4]

    METHODOLOGY To address this multi-modal threat, our methodology centers onjoint anonymizationof both the acoustic voice and the linguistic content. We use an ASR-TTS pipeline and introduce novel paraphrasing tech- niques for the intermediate content anonymization step: 1.Utterance-by-utterance paraphrasing (GPT-4o-mini): Each utterance is independently pa...

  2. [1]

    Exist- ing approaches, driven by benchmarks like the V oicePrivacy Chal- lenge, have shown success at the utterance-level [1]

    INTRODUCTION V oice anonymization aims to mitigate privacy risks by modifying speech to conceal speaker identity while preserving utility. Exist- ing approaches, driven by benchmarks like the V oicePrivacy Chal- lenge, have shown success at the utterance-level [1]. Simple methods such as kNN-based voice conversion can achieve strong anonymity even in the ...

  3. [2]

    This provides a balance between privacy and utility, unlike encryption, which renders data unusable for downstream tasks, or redaction, which destroys utility

    RELA TED WORK V oice AnonymizationThe task of voice anonymization aims to suppress speaker identity while retaining other useful attributes of speech. This provides a balance between privacy and utility, unlike encryption, which renders data unusable for downstream tasks, or redaction, which destroys utility. Research in this area is largely driven by ben...

  4. [3]

    In this approach, the original audio is first transcribed into text by an ASR system

    PRELIMINARIES ASR-TTS Anonymization PipelineA common and effective method for voice anonymization is a cascaded pipeline that com- bines automatic speech recognition (ASR) and text-to-speech (TTS) synthesis [11]. In this approach, the original audio is first transcribed into text by an ASR system. This intermediate text representation provides a powerful ...

  5. [5]

    flattens

    EXPERIMENTS Our experimental framework is designed to evaluate the vulnerabil- ity of different anonymization strategies to content-based attacks. The results shown in Figure 1 demonstrate the core problem: a stan- dard ASR-TTS pipeline, which perfectly anonymizes the acoustic voice, remains highly vulnerable to a content-only attacker. The attacker’s per...

  6. [6]

    content gap

    CONCLUSION Existing voice anonymization techniques are built on the flawed premise that obscuring a speaker’s voice is sufficient to protect their identity. This assumption breaks down in real-world scenarios involving long-form audio, where the linguistic content can be ex- ploited by an attacker [4]. Our work highlights this “content gap” compared to vo...

  7. [7]

    ACKNOWLEDGMENTS This research is supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via the HIATUS Program contract #D2022-2205150003 and the ARTS Program contract #D2023- 2308110001. The views and conclusions contained herein are those of the authors and should not be...

  8. [8]

    The V oicePrivacy 2024 Challenge Evaluation Plan,

    Natalia Tomashenko, Xiaoxiao Miao, Pierre Champion, Sa- rina Meyer, Xin Wang, Emmanuel Vincent, Michele Pa- nariello, Nicholas Evans, Junichi Yamagishi, and Massimiliano Todisco, “The V oicePrivacy 2024 Challenge Evaluation Plan,” arXiv preprint arXiv:2404.02677, 2024

Show all 33 references
  1. [9]

    HLTCOE Submission to the 2024 V oicePrivacy Challenge,

    Henry Li Xinyuan, Zexin Cai, Ashi Garg, Leibny Paola Garcia- Perera, Kevin Duh, Sanjeev Khudanpur, Nicholas Andrews, and Matthew Wiesner, “HLTCOE Submission to the 2024 V oicePrivacy Challenge,” inProceedings of the 4th Symposium on Security and Privacy in Speech Communication, 2024

  2. [10]

    Use Cases for V oice Anonymization,

    Sarina Meyer and Ngoc Thang Vu, “Use Cases for V oice Anonymization,” in5th Symposium on Security and Privacy in Speech Communication, 2025

  3. [11]

    Can Authorship Attribution Models Distinguish Speakers in Speech Transcripts?,

    Cristina Aggazzotti, Nicholas Andrews, and Elizabeth Al- lyn Smith, “Can Authorship Attribution Models Distinguish Speakers in Speech Transcripts?,”Transactions of the Asso- ciation for Computational Linguistics, vol. 12, pp. 875–891, 2024

  4. [12]

    The Impact of Automatic Speech Transcription on Speaker Attribution,

    Cristina Aggazzotti, Matthew Wiesner, Elizabeth Allyn Smith, and Nicholas Andrews, “The Impact of Automatic Speech Transcription on Speaker Attribution,”Transactions of the As- sociation for Computational Linguistics, in press

  5. [13]

    Gemma 3 Technical Report,

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Ta- tiana Matejovicova, Alexandre Ram ´e, Morgane Rivi `ere, et al., “Gemma 3 Technical Report,”arXiv preprint arXiv:2503.19786, 2025

  6. [14]

    Re- visiting speech content privacy,

    Jennifer Williams, Junichi Yamagishi, Paul-Gauthier No ´e, Cassia Valentini Botinhao, and Jean-Franc ¸ois Bonastre, “Re- visiting speech content privacy,” in2021 ISCA Symposium on Security and Privacy in Speech Communication, 2021

  7. [15]

    New challenges for content privacy in speech and audio,

    Jennifer Williams, Karla Pizzi, Shuvayanti Das, and Paul- Gauthier No´e, “New challenges for content privacy in speech and audio,” in2nd Symposium on Security and Privacy in Speech Communication, 2023

  8. [16]

    Exploratory evaluation of speech content masking,

    Jennifer Williams, Karla Pizzi, Paul-Gauthier Noe, and Sneha Das, “Exploratory evaluation of speech content masking,” in Speech Communication; 15th ITG Conference, 2023, pp. 215– 219

  9. [17]

    Adapting language models when training on privacy-transformed data,

    Tugtekin Turan, Dietrich Klakow, Emmanuel Vincent, and Denis Jouvet, “Adapting language models when training on privacy-transformed data,” inProceedings of the 13th Lan- guage Resources and Evaluation Conference. June 2022, pp. 4367–4373, European Language Resources Association

  10. [18]

    Se- cureSpeech: Prompt-based Speaker and Content Protection,

    Belinda Soh Hui Hui, Xiaoxiao Miao, and Xin Wang, “Se- cureSpeech: Prompt-based Speaker and Content Protection,” inIEEE International Joint Conference on Biometrics, 2025

  11. [19]

    Surveying Stylom- etry Techniques and Applications,

    Tempestt Neal, Kalaivani Sundararajan, Aneez Fatima, Yiming Yan, Yingfei Xiang, and Damon Woodard, “Surveying Stylom- etry Techniques and Applications,”ACM Computing Surveys (CSuR), vol. 50, no. 6, pp. 1–36, 2017

  12. [20]

    A Survey of Modern Authorship At- tribution Methods,

    Efstathios Stamatatos, “A Survey of Modern Authorship At- tribution Methods,”Journal of the American Society for Infor- mation Science and Technology, vol. 60, no. 3, pp. 538–556, 2009

  13. [21]

    Learning Universal Authorship Represen- tations,

    Rafael A. Rivera-Soto, Olivia Elizabeth Miano, Juanita Or- donez, Barry Y . Chen, Aleem Khan, Marcus Bishop, and Nicholas Andrews, “Learning Universal Authorship Represen- tations,” inProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021,...

  14. [22]

    Same Author or Just Same Topic? Towards Content-Independent Style Representations,

    Anna Wegmann, Marijn Schraagen, and Dong Nguyen, “Same Author or Just Same Topic? Towards Content-Independent Style Representations,” inProceedings of the 7th Workshop on Representation Learning for NLP, 2022, pp. 249–268

  15. [23]

    HLTCOE Submission to the V oicePrivacy Attacker Challenge,

    Henry Li Xinyuan, Ashi Garg, Zexin Cai, Kevin Duh, Leibny Paola Garc´ıa-Perera, Sanjeev Khudanpur, Nicholas An- drews, and Matthew Wiesner, “HLTCOE Submission to the V oicePrivacy Attacker Challenge,” inIEEE International Con- ference on Acoustics, Speech and Signal Processing...

  16. [24]

    The Fisher Corpus: A Resource for the Next Generations of Speech-to-Text,

    Christopher Cieri, David Miller, and Karen Walker, “The Fisher Corpus: A Resource for the Next Generations of Speech-to-Text,” inProc. of the 4th International Conference on Language Resources and Evaluation, 2004, pp. 69–71

  17. [25]

    XTTS: a Mas- sively Multilingual Zero-Shot Text-to-Speech Model,

    Edresson Casanova, Kelly Davis, Eren G¨olge, G¨orkem G¨oknar, Iulian Gulea, Logan Hart, Aya Aljafari, Joshua Meyer, Reuben Morais, Samuel Olayemi, and Julian Weber, “XTTS: a Mas- sively Multilingual Zero-Shot Text-to-Speech Model,” inIn- terspeech 2024, 2024, pp. 4978–4982

  18. [26]

    V oxCeleb2: Deep Speaker Recognition,

    Joon Son Chung, Arsha Nagrani, and Andrew Zisserman, “V oxCeleb2: Deep Speaker Recognition,” inInterspeech 2018, 2018, pp. 1086–1090

  19. [27]

    A Deep Metric Learning Approach to Account Linking,

    Aleem Khan, Elizabeth Fleming, Noah Schofield, Marcus Bishop, and Nicholas Andrews, “A Deep Metric Learning Approach to Account Linking,” inProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technol...

  20. [28]

    Spotting LLMs with Binoc- ulars: Zero-shot detection of machine-generated text,

    Abhimanyu Hans, Avi Schwarzschild, Valeriia Cherepanova, Hamid Kazemi, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein, “Spotting LLMs with Binoc- ulars: Zero-shot detection of machine-generated text,” in Proceedings of the 41st International Conference on Mac...

  21. [29]

    UTMOS: UTokyo-SaruLab System for V oiceMOS Challenge 2022,

    Tomoki Saeki, Wataru Kawamura, Shinnosuke Takamichi, and Hiroshi Saruwatari, “UTMOS: UTokyo-SaruLab System for V oiceMOS Challenge 2022,” inProceedings of the V oiceMOS Challenge 2022, 2022, pp. 33–38

  22. [30]

    Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks,

    Nils Reimers and Iryna Gurevych, “Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks,” inProceedings of the 2019 Conference on Empirical Methods in Natural Lan- guage Processing, 2019

  23. [31]

    GPT-5 System Card,

    OpenAI, “GPT-5 System Card,” August 2025

  24. [32]

    Automatic Speaker Verification Spoofing and Deepfake Detection Using Wav2vec 2.0 and Data Augmentation,

    Hemlata Tak, Massimiliano Todisco, Xin Wang, Jee weon Jung, Junichi Yamagishi, and Nicholas Evans, “Automatic Speaker Verification Spoofing and Deepfake Detection Using Wav2vec 2.0 and Data Augmentation,” inSpeaker and Lan- guage Recognition Workshop (Odyssey), 2022, pp. 112–119

  25. [33]

    Encoder-Decoder Based Attractors for End-to-End Neural Diarization,

    Shota Horiguchi, Yusuke Fujita, Shinji Watanabe, Yawen Xue, and Paola Garc ´ıa, “Encoder-Decoder Based Attractors for End-to-End Neural Diarization,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 1493– 1507, 2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.