Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Children's Voice Privacy: First Steps And Emerging Challenges

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Adult-trained voice anonymization raises speaker verification error for children's speech to adult-like levels, but it increases word error rate by about 17 percentage points on average and over 45 in the worst case.

desk verdict A solid first baseline for child voice anonymization; the WV-MOS bias finding is the real gem, but the abstract oversells the privacy claim and the missing error bars matter. read the letter →

arxiv 2506.00100 v2 pith:RWGGDQGB submitted 2025-05-30 cs.CY cs.AIcs.CLcs.CR

Ajinkya Kulkarni , Francisco Teixeira , Enno Hermann , Thomas Rolland , Isabel Trancoso , Mathew Magimai Doss This is my paper
classification cs.CYcs.AIcs.CLcs.CR
keywords children'sspeechvoiceprivacyspeakeranonymizationautomaticverificationutilityconversionrecognitionsubjectivequality
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

This paper asks whether voice anonymization systems built for adult speech can protect children's voices while keeping them usable. Across three children's datasets and six anonymization methods, the authors find that the systems do raise speaker verification error rates for children's speech to levels comparable to their effect on adult speech, suggesting real privacy gains. The cost is much higher utility loss: on MyST, word error rates rise by about 17 percentage points on average and by over 45 points in the worst case, far exceeding the at-most-4-point degradation seen on adult LibriSpeech. The paper also shows that automatic quality scores such as WV-MOS can be misleading for children's speech, since converting a child's voice to an adult-like voice can raise the score even though human listeners rate the original as more natural. The authors present these results as a first baseline for children's voice anonymization, with the caveat that the privacy measure relies on an adult-trained speaker verifier.

What carries the argument

The evaluation machinery is a transfer test: adult-trained anonymization pipelines (McAdams coefficient, x-vector bottleneck features, phonetic-transcription GAN, neural audio codec, and KNN voice conversion with and without rhythm change) are applied unchanged to children's speech, and the outputs are scored by an ECAPA-TDNN x-vector speaker verifier trained on adult VoxCeleb2 for privacy and by Whisper-based ASR for intelligibility. The load-bearing comparison is the gap between EER and WER on children versus adult speech under the same systems. A second mechanism is the contrast between objective quality scores (WV-MOS) and a subjective listening test, which exposes that the automatic metric is biased toward adult-like voices.

What would settle it

Recompute EER on the same anonymized children's utterances using a speaker verifier that has been trained or fine-tuned on children's speech, and compare with the adult-trained scores. If EER drops substantially toward original-speech levels, the privacy protection is mostly an artifact of adult-child domain mismatch; if EER remains high, the privacy claim survives a stronger test. A human re-identification listening test on anonymized child voices would settle the practical question directly.

Watch

Extended reading notes

Core claim

The paper's central claim is that existing adult-oriented anonymization systems, when applied without adaptation to children's voices, preserve speaker privacy at levels comparable to adult anonymization while degrading linguistic utility far more severely. Privacy is measured as equal error rate of an ECAPA-TDNN speaker verifier: original MyST speech has EER 5.0 percent, anonymized versions range roughly 23 to 48 percent, compared with LibriSpeech adult original EER 0.41 percent and anonymized EERs around 25 to 49 percent. Utility is measured as WER: anonymized adult speech degrades by at most 4 absolute points, while anonymized children's speech on MyST degrades by about 17 points on average and over 45 points for the STTTS system. The authors conclude that current systems are not ready for children without adaptation, identify the neural audio codec system NAC as offering the best privacy-utility trade-off among the tested baselines, and present the results as establishing a baseline for the under-explored area of children's voice anonymization.

Load-bearing premise

The privacy result rests on the assumption that equal error rates from an adult-trained speaker verifier correctly measure how hard it is to re-identify a child from anonymized speech; if the verifier fails on children's voices for reasons unrelated to anonymization, higher EER could reflect domain mismatch rather than real protection.

Editorial extensions

If this is right

  • Adult-trained anonymization systems can serve as a starting point for children's voice privacy, but only with utility-oriented adaptation, since WER degradation is several times larger for children than for adults.
  • Systems that rely on latent features, such as NAC, degrade utility less than systems with hard phonetic decisions, such as STTTS, pointing to architecture choice as a key lever for child anonymization.
  • Older children's anonymized speech is recognized more accurately than younger children's, while younger children's speech shows higher EER, suggesting age-dependent behavior in both privacy and utility.
  • Automatic speech-quality metrics trained on adult voices should not be trusted for children's anonymization without subjective validation, because WV-MOS can rise after anonymization while human listeners prefer the original.
  • The results motivate child-to-child anonymization as a future direction, while flagging ethical concerns about identity falsification and misrepresentation.

Reading between the lines

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

  • A direct extension would be to train or adapt the ASR, speaker-embedding, and vocoder components inside anonymization pipelines on children's speech; the paper's diagnosis of recognition errors in the linguistic branch suggests this should narrow the utility gap.
  • If privacy is measured with a child-adapted speaker verifier and still holds strongly, the case for converting child voices into adult voices weakens, and child-to-child conversion might deliver similar privacy with better intelligibility and more natural prosody.
  • The WV-MOS bias result implies that future benchmarks of children's speech quality should include subjective listening as ground truth, since objective-only leaderboards could reward systems that merely make children sound adult.
  • The tested languages cover English, non-native English, and Icelandic; testing on more languages and a fuller range of ages, including children with speech disorders, would show whether the utility collapse is worsened by language-dependent modules.
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 / 5 minor

Summary. The paper evaluates six voice anonymization systems (four VPC24 baselines plus KNN-VC and KNN-VC+R) applied to children's speech across three datasets (MyST, SpeechOcean762, Samrómur) and compares results with adult LibriSpeech. Privacy is measured by EER from an ECAPA-TDNN speaker verifier trained on adult VoxCeleb2; utility is measured by WER from Whisper-based ASR and by WV-MOS; a small subjective listening test supplements the objective metrics. The main reported findings are that adult-trained anonymization systems raise EER for children's speech to levels comparable to adult anonymization, but cause much larger WER degradation on children's speech, and that WV-MOS appears biased toward adult-like anonymized outputs.

Significance. If the privacy result holds, the paper provides a useful first baseline for child voice privacy, with a commendably broad experimental matrix: three children's datasets in two languages, six anonymization systems, objective and subjective metrics, and an adult comparison condition. The authors also make speech samples publicly available and explicitly discuss several limitations. However, the central privacy claim depends on an adult-trained speaker verifier whose behavior on original child speech already deviates strongly from its behavior on adult speech, and the paper's own Section 4 cautions that real-world re-identification risk is uncertain. The utility and subjective-testing conclusions are likewise based on point estimates and very small samples. The contribution is a reasonable exploratory baseline, but the abstract overstates what is established.

major comments (4)
  1. [§2.3 and Table 1] The central privacy claim is not established because EER is computed only with an ECAPA-TDNN trained on adult VoxCeleb2. On original (non-anonymized) children's speech this verifier already yields much higher EER than on adult LibriSpeech (MyST 5.0% vs 0.41%; Samrómur AgeGrp1 11.5%), and the anonymization systems all convert child voices into adult-like voices. The large EER increases after anonymization (e.g., MyST original 5.0% to 47.8% for ASR-BN) could therefore reflect the verifier's failure to produce stable embeddings for distorted or domain-shifted child speech rather than removal of speaker identity. The authors themselves state in Section 4 that "the actual risk of re-identification in real-world scenarios remains uncertain." To support the abstract's claim that existing systems "are still able to protect children's voice privacy," the paper needs either a child-competent ASV (trained or adapted on child speech) or a direct analysis showing that EER increases track identity removal and not merely signal distortion (e.g., by comparing ASV score distributions on original versus anonymized child speech with a matched child verifier).
  2. [§3 and Table 1] All EER and WER results are reported as point estimates without confidence intervals or significance tests. Several datasets have small speaker samples (e.g., SpeechOcean uses 10 child speakers; Samrómur about 13 speakers per age group), so differences such as MyST NAC (44.40%) versus ASR-BN (47.80%) may be within sampling error. Without bootstrap intervals or a paired test across speakers, the claims that NAC achieves the best privacy-utility trade-off and that age-group effects are consistent are not supported by the reported evidence.
  3. [§2.3 and Table 2] The subjective listening test uses only 5 utterances and 12 listeners, with no confidence intervals, inter-rater reliability, or statistical comparison. This small sample is used to conclude that WV-MOS is biased toward adult speech and that listeners perceive 73%–95% of outputs from neural anonymization systems as adult-like. These conclusions are load-bearing for the paper's secondary claim about "the challenges of automatic evaluation methods for speech quality in children's speech," but the sample size is too small to support them. The authors should treat the listening test as a pilot and temper the conclusion, or expand the test and report inferential statistics.
  4. [§3 and §4] The utility-degradation claim compares WER increases on children's speech with those on adult speech, but the ASR baseline on original child speech is already much worse (MyST original WER 13.42% vs LibriSpeech 2.56%). The absolute WER increase after anonymization may partly reflect the adult-trained ASR's pre-existing difficulty with child speech, not purely the anonymization's destruction of linguistic content. The authors acknowledge in Section 4 that WER degradation may overestimate utility loss; this caveat should be incorporated into the abstract and conclusions, or supported by a human intelligibility test.
minor comments (5)
  1. [§2.3 / Table 2] The acronym "ND-MOS" in Table 2 is not defined in the text; please define it or use the full phrase "naturalness MOS" consistently.
  2. [Throughout] There are inconsistent spellings and spacing, e.g., "V oice" instead of "Voice" in several places ("V oice Privacy Challenges", "V oice Conversion"), and "Samromur" versus "Samrómur" used inconsistently.
  3. [§2.1] The description of SpeechOcean speaker selection is unclear: the text says "two speakers from each age" and "a total of 10 speakers," but the age groups span five years each; please clarify the selection procedure and the number of speakers per age.
  4. [Table 1] Please include the number of speakers and utterances per dataset in the caption or table, since the trial counts alone do not convey the small sample sizes that affect the interpretation of the results.
  5. [§3] The sentence "privacy results for children in age group one are consistently higher" is ambiguous because higher EER means better privacy, yet the following sentence suggests the effect may be an artifact of poorer audio quality; please rephrase to make the direction of the effect and the interpretation clear.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports empirical measurements against external, pre-trained systems, and the metric-validity caveats it raises are acknowledged limitations, not constructional circularity.

full rationale

This paper contains no derivation chain that reduces to its own inputs. Its central claims are empirical measurements: six off-the-shelf anonymization systems (VPC24 baselines plus KNN-VC variants) are applied to three children's datasets and evaluated with external, pre-trained models (ECAPA-TDNN ASV trained on VoxCeleb2 for EER, Whisper large-v3 for WER, WV-MOS for quality) and a fresh subjective listening study. No parameter is fitted to any subset of the data and then renamed as a prediction; the anonymization systems, ASV, ASR, and MOS models are all fixed and used without fine-tuning on children's speech. The abstract's privacy claim ('existing systems for adults are still able to protect children's voice privacy') is an inference from EER measurements on held-out trial pairs, not a consequence of how EER is defined. The main validity threats — that the adult-trained ECAPA-TDNN may yield inflated EER on children's speech due to domain mismatch, and that WV-MOS is biased toward adult-sounding speech — are explicitly acknowledged by the authors (Section 3: 'the poorer quality of the generated anonymized utterances may have a negative impact on ASV performance'; Section 4: 'the actual risk of re-identification in real-world scenarios remains uncertain'; Table 2 discussion hypothesizes WV-MOS's adult bias), and the subjective ND-MOS test is used as an independent check that partially corrects the automated metric. Self-citations ([7] for ASR background, [28] for the KNN-VC+R rhythm-modification variant) are method or background references and are not load-bearing for the paper's core conclusion, which stands independently of them. Since nothing is fitted and every metric is an external benchmark evaluated without modification, the correct finding is no significant circularity (score 0).

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

No parameters are fitted in this paper; system hyperparameters are inherited from the public VPC24 configurations. The load-bearing assumptions are the validity of the adult-trained ASV as a privacy oracle for children, the chosen lazy-informed attacker model, and the use of WV-MOS as a naturalness proxy, which the paper itself later shows to be biased for children's speech.

assumptions (3)
  • domain assumption The ECAPA-TDNN ASV trained on VoxCeleb2 adult speech is a valid privacy oracle for children's anonymized speech.
    EER from this model is the primary privacy metric (Section 2.3); Section 3 acknowledges that poor output quality could artificially inflate EER for younger children.
  • domain assumption The lazy-informed attacker, who has anonymized data but does not fine-tune the ASV, is an appropriate threat model for the privacy claim.
    Adopted in Section 2.3; Section 4 concedes stronger attackers may reduce privacy guarantees.
  • domain assumption WV-MOS predicts subjective naturalness for anonymized children's speech.
    Used as a secondary utility metric (Section 2.3) and then shown to be inconsistent with human ND-MOS results in Table 2 and Section 3, so this assumption is questionable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Children's Voice Privacy: First Steps And Emerging Challenges." pith.science (2026). https://pith.science/paper/RWGGDQGB

@misc{pith2026250600100,
  author       = {Pith},
  title        = {Pith review of: Children's Voice Privacy: First Steps And Emerging Challenges},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RWGGDQGB}},
  note         = {Machine review of arXiv:2506.00100}
}
read the original abstract

Children are one of the most under-represented groups in speech technologies, as well as one of the most vulnerable in terms of privacy. Despite this, anonymization techniques targeting this population have received little attention. In this study, we seek to bridge this gap, and establish a baseline for the use of voice anonymization techniques designed for adult speech when applied to children's voices. Such an evaluation is essential, as children's speech presents a distinct set of challenges when compared to that of adults. This study comprises three children's datasets, six anonymization methods, and objective and subjective utility metrics for evaluation. Our results show that existing systems for adults are still able to protect children's voice privacy, but suffer from much higher utility degradation. In addition, our subjective study displays the challenges of automatic evaluation methods for speech quality in children's speech, highlighting the need for further research.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 36 canonical work pages

  1. [1]

    Introduction The rapid evolution of technology over recent decades has significantly transformed children’s activities, shifting many from physical spaces to digital environments. The increas- ing adoption of speech technologies underscores their impor- tance in children’s education, offering advanced learning tools such as interactive reading tutors and ...

    work page 2020

  2. [2]

    Methodology To explore the questions raised in the previous section, we examined different datasets combined with several voice anonymization systems, focusing on, as a first step, anonymiz- arXiv:2506.00100v2 [cs.CY] 4 Jun 2025 ing children’s speech to adult speech. We analyzed the im- pact of anonymization across three children’s speech datasets, each w...

    arXiv 2025

  3. [3]

    Results and Discussion We start by discussing the results obtained for the MyST and LibriSpeech data, comparing the performance of the anonymization systems applied to children and adult speech 4. These results can be found in the first two lines of Table 1, and show that, while existing anonymization systems are able to protect the privacy of children’s ...

    work page arXiv

  4. [4]

    One of the major challenges is identifying the exact sources of degradation in utility metrics

    Challenges and Limitations Despite the promising results in privacy protection, several chal- lenges arise in applying voice anonymization techniques to chil- dren’s speech. One of the major challenges is identifying the exact sources of degradation in utility metrics. The decrease in ASR performance and speech quality may be attributed to mul- tiple fact...

    work page

  5. [5]

    Through experiments on diverse children’s speech datasets, the results indicate that anonymization to adult speech with existing systems can enhance privacy to certain degree

    Conclusions and Future work This study establishes a preliminary understanding of chil- dren’s voice privacy by evaluating the performance of voice anonymization techniques originally developed for adult speech. Through experiments on diverse children’s speech datasets, the results indicate that anonymization to adult speech with existing systems can enha...

    work page

  6. [6]

    219726) and by the Inno- suisse through the flagship project IICT: Inclusive Informa- tion and Communication Technologies (grant agreement no

    Acknowledgments This work was partially supported by the Swiss National Sci- ence Foundation through the project PASS: Pathological Speech Synthesis (grant agreement no. 219726) and by the Inno- suisse through the flagship project IICT: Inclusive Informa- tion and Communication Technologies (grant agreement no. PFFS-21-47), and by the Portuguese national ...

    work page doi:10.54499/uidb/50021/2020

  7. [7]

    Un- veiling biases while embracing sustainability: Assessing the dual challenges of automatic speech recognition systems,

    A. Kulkarni, A. Kulkarni, I. Trancoso, and M. Couceiro, “Un- veiling biases while embracing sustainability: Assessing the dual challenges of automatic speech recognition systems,” inProc. In- terspeech, 2024

    work page 2024

  8. [8]

    Automated speech scoring for non- native middle school students with multiple task types,

    K. Evanini and X. Wang, “Automated speech scoring for non- native middle school students with multiple task types,” inProc. Interspeech, 2013, pp. 2435–2439

    work page 2013

Show all 44 references
  1. [9]

    Teens, social media and technology 2022,

    E. A. V ogels, R. Gelles-Watnick, and N. Massarat, “Teens, social media and technology 2022,”Pew Research Center, vol. 10,

    2022

  2. [10]

    Exploring discrete speech units for privacy-preserving and efficient speech recognition for school-aged and preschool children,

    S. Dutta, D. Irvin, and J. H. Hansen, “Exploring discrete speech units for privacy-preserving and efficient speech recognition for school-aged and preschool children,”International Journal of Human-Computer Studies, 2025

    2025

  3. [11]

    R. Kats. (2018) Are Kids and Teens Using Smart Speakers? Ac- cessed: 2024-02-20. [Online]. Available: https://www.emarketer. com/content/the-smart-speaker-series-kids-teens-infographic

    2018

  4. [12]

    Youth, privacy and online me- dia: Framing the right to privacy in public policy-making,

    G. H. Lapenta and R. F. Jørgensen, “Youth, privacy and online me- dia: Framing the right to privacy in public policy-making,”First Monday, vol. 20, no. 3, 2015

    2015

  5. [13]

    Acoustics of children’s speech: Developmental changes of temporal and spectral parame- ters,

    S. Lee, A. Potamianos, and S. Narayanan, “Acoustics of children’s speech: Developmental changes of temporal and spectral parame- ters,”The Journal of the Acoustical Society of America, vol. 105, no. 3, pp. 1455–1468, 1999

    1999

  6. [14]

    Speech production variability in fricatives of children and adults: Results of func- tional data analysis,

    L. L. Koenig, J. C. Lucero, and E. Perlman, “Speech production variability in fricatives of children and adults: Results of func- tional data analysis,”The Journal of the Acoustical Society of America, vol. 124, no. 5, pp. 3158–3170, 2008

    2008

  7. [15]

    The VoicePrivacy 2024 Challenge Evaluation Plan,

    N. Tomashenko, X. Miao, P. Championet al., “The VoicePrivacy 2024 Challenge Evaluation Plan,” inProc. Symposium on Security and Privacy in Speech Communication, 2024

    2024

  8. [16]

    Analysis of disfluency in children’s speech,

    T. Tran, M. Tinkler, G. Yeung, A. Alwan, and M. Ostendorf, “Analysis of disfluency in children’s speech,” inProc. Inter- speech, 2020, pp. 4278–4282

    2020

  9. [17]

    On the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (GDPR),

    European Parliament and Council, “On the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46/EC (GDPR),”Regulation 2016/679, April 2016

    2016

  10. [18]

    Samrómur Children: An Icelandic Speech Corpus,

    C. D. H. Mena, D. E. Mollberg, M. Borský, and J. Guðnason, “Samrómur Children: An Icelandic Speech Corpus,” inProc. LREC, 2022

    2022

  11. [19]

    Ad- dressing challenges in speaker anonymization to maintain utility while ensuring privacy of pathological speech,

    S. Tayebi Arasteh, T. Arias-Vergara, P. A. Pérez-Toroet al., “Ad- dressing challenges in speaker anonymization to maintain utility while ensuring privacy of pathological speech,”Communications Medicine, vol. 4, no. 1, p. 182, 2024

    2024

  12. [20]

    Differential privacy for protecting patient data in speech disorder detection using deep learning,

    S. T. Arasteh, M. Lotfinia, P. A. Perez-Toroet al., “Differential privacy for protecting patient data in speech disorder detection using deep learning,”arXiv preprint arXiv:2409.19078, 2024

    2024

  13. [21]

    Privacy- preserving speech analytics for automatic assessment of student collaboration,

    N. Bassiou, A. Tsiartas, J. Smith, H. Brattet al., “Privacy- preserving speech analytics for automatic assessment of student collaboration,” inProc. Interspeech, 2016, pp. 888–892

    2016

  14. [22]

    The V oicePrivacy 2020 Challenge Evaluation Plan,

    N. Tomashenko, B. M. L. Srivastava, X. Wanget al., “The V oicePrivacy 2020 Challenge Evaluation Plan,” 2020

    2020

  15. [23]

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

    S. Meyer, F. Lux, J. Kochet al., “Prosody is not identity: A speaker anonymization approach using prosody cloning,” inProc. ICASSP, 2023, pp. 1–5

    2023

  16. [24]

    My science tutor: A con- versational multimedia virtual tutor

    W. Ward, R. Cole, D. Bolañoset al., “My science tutor: A con- versational multimedia virtual tutor.”Journal of Educational Psy- chology, vol. 105, no. 4, p. 1115, 2013

    2013

  17. [25]

    speechocean762: An Open- Source Non-Native English Speech Corpus for Pronunciation As- sessment,

    J. Zhang, Z. Zhang, Y . Wanget al., “speechocean762: An Open- Source Non-Native English Speech Corpus for Pronunciation As- sessment,” inProc. Interspeech, 2021, pp. 3710–3714

    2021

  18. [26]

    V oice Conversion With Just Nearest Neighbors,

    M. Baas, B. Van Niekerk, and H. Kamper, “V oice Conversion With Just Nearest Neighbors,” inProc. Interspeech, 2023, pp. 2053–2057

    2023

  19. [27]

    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,” Proc. ICASSP, 2015

    2015

  20. [28]

    Common V oice: A Massively-Multilingual Speech Corpus,

    R. Ardila, M. Branson, K. Daviset al., “Common V oice: A Massively-Multilingual Speech Corpus,” inProc. LREC, 2020, pp. 4218–4222

    2020

  21. [29]

    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

  22. [30]

    Anonymizing Speech: Evaluating and Designing Speaker Anonymization Techniques,

    P. Champion, “Anonymizing Speech: Evaluating and Designing Speaker Anonymization Techniques,” Ph.D. dissertation, Univer- sité de Lorraine, 2023

    2023

  23. [31]

    We obtained ASV scores by computing the cosine similarity between the mean speaker embedding of enrollment/impostor utterances and test utterances

    trained on the V oxCeleb 2 dataset [32]. We obtained ASV scores by computing the cosine similarity between the mean speaker embedding of enrollment/impostor utterances and test utterances. We assume alazy-informedattacker [33], who has access to anonymized enrollment and trial...

  24. [32]

    Design Choices for X-vector Based Speaker Anonymization,

    B. M. L. Srivastava, N. Tomashenko, X. Wanget al., “Design Choices for X-vector Based Speaker Anonymization,” inProc. Interspeech, 2020

    2020

  25. [33]

    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,” in Proc. ICASSP, 2024, pp. 4725–4729

    2024

  26. [34]

    SamróMur MilljóN: An ASR corpus of one million verified read prompts in Icelandic,

    C. D. Hernandez Mena, Þ. D. Gunnarsson, and J. Gudnason, “SamróMur MilljóN: An ASR corpus of one million verified read prompts in Icelandic,” inProc. LREC-COLING, 2024, pp. 14 305–14 312

    2024

  27. [35]

    Com- paring Speech Anonymization Efficacy by V oice Conversion Us- ing KNN and Disentangled Speaker Feature Representations,

    A. Das, C. Franzreb, T. Herzig, P. Pirlet, and T. Polzehl, “Com- paring Speech Anonymization Efficacy by V oice Conversion Us- ing KNN and Disentangled Speaker Feature Representations,” in Proc. Symposium on Security and Privacy in Speech Communica- tion, 2024, pp. 121–126

    2024

  28. [36]

    Unsupervised rhythm and voice conversion of dysarthric to healthy speech for ASR,

    K. E. Hajal, E. Hermann, A. Kulkarni, and M. Magimai.-Doss, “Unsupervised rhythm and voice conversion of dysarthric to healthy speech for ASR,” inProc. ICASSP, 2025

    2025

  29. [37]

    Rhythm modeling for voice conversion,

    B. van Niekerk, M.-A. Carbonneau, and H. Kamper, “Rhythm modeling for voice conversion,”IEEE Signal Processing Letters, vol. 30, pp. 1297–1301, 2023

    2023

  30. [38]

    The LJ speech dataset,

    K. Ito and L. Johnson, “The LJ speech dataset,” https://keithito. com/LJ-Speech-Dataset/, 2017

    2017

  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,” inProc. Inter- speech, 2020, pp. 3830–3834

    2020

  32. [40]

    V oxCeleb2: Deep speaker recognition,

    J. S. Chung, A. Nagrani, and A. Zisserman, “V oxCeleb2: Deep speaker recognition,” inProc. Interspeech, 2018, pp. 1086–1090

    2018

  33. [41]

    Privacy and utility of x-vector based speaker anonymization,

    B. M. L. Srivastava, M. Maouche, M. Sahidullahet al., “Privacy and utility of x-vector based speaker anonymization,”IEEE/ACM TASLP, vol. 30, pp. 2383–2395, 2022

    2022

  34. [43]

    HIFI++: A Uni- fied Framework for Bandwidth Extension and Speech Enhance- ment,

    P. Andreev, A. Alanov, O. Ivanov, and D. Vetrov, “HIFI++: A Uni- fied Framework for Bandwidth Extension and Speech Enhance- ment,” inProc. ICASSP, 2023

    2023

  35. [44]

    BeaqleJS: HTML5 and JavaScript based Framework for the Subjective Evaluation of Audio Quality,

    S. Kraft and U. Zolzer, “BeaqleJS: HTML5 and JavaScript based Framework for the Subjective Evaluation of Audio Quality,” in Proc. Linux Audio Conference, 2014

    2014

  36. [2022]

    Available: https://www.pewresearch.org/internet/ 2022/08/10/teens-social-media-and-technology-2022

    [Online]. Available: https://www.pewresearch.org/internet/ 2022/08/10/teens-social-media-and-technology-2022

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.