Child-Centric Voice Anonymization in Single and Multi-Speaker Speech via Domain-Adapted SSL Models

Pranav Tushar; Rong Tong; Xiao Xiao Miao

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · grok-4.3

Adapting a self-supervised voice anonymization system to child speech improves intelligibility and perceptual quality without compromising privacy.

2026-06-30 05:24 UTC pith:I4PCFEPR

load-bearing objection This adapts existing SSL anonymization to child speech on MyST with a multi-speaker extension, but the abstract supplies no metrics or cross-corpus checks so the claimed gains are hard to evaluate. the 1 major comments →

arxiv 2606.29897 v1 pith:I4PCFEPR submitted 2026-06-29 cs.SD cs.AIeess.SP

Child-Centric Voice Anonymization in Single and Multi-Speaker Speech via Domain-Adapted SSL Models

Pranav Tushar , Xiao Xiao Miao , Rong Tong This is my paper

classification cs.SD cs.AIeess.SP

keywords child speechvoice anonymizationdomain adaptationself-supervised learningprivacy protectionmulti-speaker speechintelligibility

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Most voice anonymization tools are trained on adult speech, causing issues when used with children. This work adapts an SSL-based anonymization pipeline specifically to child speech using data from the MyST corpus. Tests in single-speaker and two-speaker settings show gains in how understandable and natural the anonymized speech sounds. Privacy against speaker identification stays strong, and the method extends to mixtures by adding speaker extraction to keep conversation flow.

Core claim

Child-centric voice anonymization is achieved by domain-adapting an SSL model on the MyST corpus, leading to better intelligibility and quality with maintained privacy in both single-speaker and multi-speaker scenarios, where target speaker extraction helps preserve conversational structure.

What carries the argument

Domain-adapted self-supervised learning (SSL) models for voice anonymization in the child speech domain.

Load-bearing premise

Gains observed from adaptation on the MyST corpus will apply generally to other child speech evaluation settings and datasets.

What would settle it

A test showing that the child-adapted model performs no better than the adult model on intelligibility, quality, or privacy metrics when applied to new child speech samples.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Improved intelligibility and perceptual quality for anonymized child speech.
Strong privacy protection is preserved post-adaptation.
Multi-speaker anonymization maintains conversational structure when combined with target speaker extraction.
The adaptation makes anonymization systems more suitable for child speech applications.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Adaptation techniques like this may be needed for other specialized speech domains to achieve effective anonymization.
Child-specific systems could enable safer voice data use in educational technology or pediatric research.
Generalization tests across different child age groups or recording conditions would strengthen the findings.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

This adapts existing SSL anonymization to child speech on MyST with a multi-speaker extension, but the abstract supplies no metrics or cross-corpus checks so the claimed gains are hard to evaluate.

read the letter

The main takeaway is that the authors take a standard SSL-based voice anonymization pipeline, adapt it on the MyST child corpus, and report that it improves intelligibility and quality in both single-speaker and two-speaker settings while holding privacy. They also combine it with speaker extraction for the multi-speaker case.

What stands out is the explicit focus on the child domain, which adult systems handle poorly. That gap is real for any child-facing application, and the multi-speaker extension is a logical next step rather than a trivial add-on.

The soft spots are the missing details. The abstract states improvements but gives no numbers, baselines, statistical tests, or error bars, so it is impossible to judge effect size or whether the adaptation actually delivers. There is also no mention of held-out testing on other child corpora, which leaves open the possibility that any gains are just reduced mismatch to MyST rather than a general child-centric method. If test utterances were not strictly separated from adaptation data, the results would be overstated.

The work is straightforward experimental adaptation with no equations or circular claims. It is aimed at researchers working on speech privacy or child speech systems who need domain-specific tweaks. A reader already familiar with SSL anonymization pipelines will see the incremental nature but might still pick up the child-specific framing.

I would send this to peer review so the full experiments and any cross-corpus results can be checked; the topic is practical enough to warrant referee time even if the current write-up is light on evidence.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes adapting an SSL-based voice anonymization pipeline to the child speech domain using the MyST corpus. It evaluates the adapted model under single-speaker and two-speaker mixture conditions, claiming gains in intelligibility and perceptual quality with preserved privacy. The work further extends the method to multi-speaker scenarios by combining target speaker extraction with the child-adapted anonymizer to maintain conversational structure.

Significance. If the reported gains prove robust, the work fills a clear gap: most voice anonymization systems are trained on adult data and degrade on child speech. Domain-adapted SSL models could enable privacy-preserving applications (educational tools, voice interfaces) for children while retaining usability. The multi-speaker extension adds practical relevance for conversational settings.

major comments (1)

[Experimental Setup / Evaluation] The central claim that MyST adaptation yields general child-centric improvements in intelligibility/quality with maintained privacy rests on evaluation within the MyST distribution. No cross-corpus held-out testing (e.g., CMU Kids or other child corpora) or explicit confirmation that test utterances were excluded from adaptation is described, so measured gains may reflect reduced domain mismatch rather than a robust method.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback on the manuscript. We respond to the major comment below, providing clarification on the experimental protocol and indicating revisions to address the concern.

read point-by-point responses

Referee: The central claim that MyST adaptation yields general child-centric improvements in intelligibility/quality with maintained privacy rests on evaluation within the MyST distribution. No cross-corpus held-out testing (e.g., CMU Kids or other child corpora) or explicit confirmation that test utterances were excluded from adaptation is described, so measured gains may reflect reduced domain mismatch rather than a robust method.

Authors: We acknowledge the validity of this observation. The current evaluation is performed within the MyST corpus using a standard train/test partition, with adaptation conducted exclusively on the training subset. We will revise the manuscript to explicitly document this disjoint partitioning and confirm that test utterances were withheld from adaptation. We agree that the absence of cross-corpus testing (e.g., on CMU Kids) limits claims of broad generalizability across child speech domains, and the observed gains could partly stem from reduced domain mismatch within MyST. In the revision we will add a dedicated limitations paragraph discussing this scope and the value of future multi-corpus validation, while retaining the within-corpus results as evidence of domain adaptation efficacy for the target setting. revision: partial

Circularity Check

0 steps flagged

No circularity: experimental adaptation pipeline with no derivations or self-referential fits

full rationale

The paper describes an empirical adaptation of an SSL-based voice anonymization system to child speech from the MyST corpus, followed by evaluation on single-speaker and two-speaker conditions. No equations, first-principles derivations, or predictions appear in the abstract or described pipeline. Claims rest on measured intelligibility, quality, and privacy metrics from adaptation and testing steps, without any fitted parameters renamed as predictions or self-definitional loops. The work is self-contained as an experimental evaluation; no load-bearing self-citations or ansatzes reduce the central results to inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, mathematical axioms, or newly invented entities; the described work relies on standard domain-adaptation practices whose concrete implementation details are not stated.

pith-pipeline@v0.9.1-grok · 5669 in / 1110 out tokens · 37532 ms · 2026-06-30T05:24:17.977930+00:00 · methodology

0 comments

read the original abstract

Voice anonymization aims to protect speaker identity while preserving linguistic content and speech usability. However, most anonymization systems are developed on adult speech, leading to degraded performance when applied to child speech. This paper investigates child-centric anonymization by adapting a self-supervised learning (SSL) based anonymization pipeline to the child speech domain. The system is adapted using child speech from the MyST corpus and evaluated under both single-speaker and two-speaker mixture conditions. Experimental results show that child-domain adaptation improves intelligibility and perceptual quality while maintaining strong privacy protection. Extending the approach to multi-speaker further demonstrates that combining target speaker extraction with child-adapted anonymization provides privacy protection while preserving conversational structure. These findings highlight the importance of child-specific adaptation for practical speech anonymization systems.

Figures

Figures reproduced from arXiv: 2606.29897 by Pranav Tushar, Rong Tong, Xiao Xiao Miao.

**Figure 1.** Figure 1: Child-centric anonymization. tion, particularly for child–child mixtures. 2. Child-centric anonymization 2.1. SSL-based single-speaker child anonymization Given an input waveform x, we extract three disentangled representations: (i) soft content features c via a HuBERT-based encoder [11], (ii) pitch contour f0, and (iii) a speaker embedding s via ECAPA-TDNN [12]. The content and prosody representations … view at source ↗

**Figure 2.** Figure 2: Subjective evaluation across datasets: listener ratings for naturalness, fluency, speaker similarity & perceived childness. configuration for all subsequent evaluations, denoted as SSLFT. Cross-domain generalization [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Multi-speaker evaluation: tWER, DER, and EER (%) by age-group pairing (AA/CA/CC) and overlap ratio. OO denotes original–original verification trials and OA denotes original–anonymized verification trials. Utility degrades with separation difficulty, especially for child– child mixtures. The second cluster of [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 4 canonical work pages · 1 internal anchor

[1]

Introduction V oice anonymization aims to suppress speaker identity while preserving linguistic content and downstream usability. The V oicePrivacy Challenge [1, 2, 3] has established common eval- uation protocols and reference systems, accelerating progress in both signal-processing and neural approaches [4, 5, 6, 7]. How- ever, most systems in this line...
[2]

Child-Centric Voice Anonymization in Single and Multi-Speaker Speech via Domain-Adapted SSL Models

or clinical sessions, frequently involves conversations be- tween children and adults. In such multi-speaker settings, it is desirable to selectively anonymize the target child while option- ally preserving or separately anonymizing the adult speaker. These limitations motivate child-to-child anonymization, where speaker identity is obfuscated while prese...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[3]

Child-centric anonymization 2.1. SSL-based single-speaker child anonymization Given an input waveformx, we extract three disentangled rep- resentations: (i) soft content featurescvia a HuBERT-based encoder [11], (ii) pitch contourf 0, and (iii) a speaker embed- dingsvia ECAPA-TDNN [12]. The content and prosody rep- resentations are kept intact to preserve...
[4]

Datasets and evaluation metrics Single-Speaker Datasets:Table 1 summarizes the datasets used for adaptation and evaluation

Experiments 3.1. Datasets and evaluation metrics Single-Speaker Datasets:Table 1 summarizes the datasets used for adaptation and evaluation. MyST [15] serves as the in- domain child speech corpus; we use the training partition for child-domain adaptation and reserve the test partition for evalu- ation. Zero-shot generalization is assessed on cross-accent ...

work page arXiv 2015
[5]

Challenges and limitations Children’s speech differs substantially from adult speech, in- troducing challenges for both modeling and evaluation. Child recordings often contain higher background noise, classroom reverberation, microphone variability, and spontaneous vocal behaviors such as disfluencies and non-lexical vocalizations. These factors can creat...
[6]

Adapting the HuBERT content encoder and HiFi-GAN vocoder to child speech improves intelligibility while maintain- ing strong privacy protection across datasets

Conclusions and future work This work presented a child-centric study of voice anonymiza- tion through domain adaptation of self-supervised speech mod- els. Adapting the HuBERT content encoder and HiFi-GAN vocoder to child speech improves intelligibility while maintain- ing strong privacy protection across datasets. Human evalu- ations further indicate th...
[7]

Acknowledgments This work is supported by the Singapore Ministry of Education (MOE) Academic Research Fund (AcRF) Tier 1 grant R-R13- A405-0005
[8]

Generative AI Use Disclosure Generative AI tools were used in a limited manner for (i) lan- guage editing during manuscript preparation and (ii) transcrip- tion support for WER-based evaluation using OpenAI ASR models. In addition, AI-generated text-to-speech (TTS) voices (Typecast and SpeechGen) were used to construct the child-like reference speaker poo...
[9]

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, 2020, pp. 1693–1697

2020
[10]

The voiceprivacy 2022 challenge: Progress and perspec- tives in voice anonymisation,

M. Panariello, N. Tomashenko, X. Wang, X. Miao, P. Champion, H. Nourtel, M. Todisco, N. Evans, E. Vincent, and J. Yamag- ishi, “The voiceprivacy 2022 challenge: Progress and perspec- tives in voice anonymisation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 3477–3491, 2024

2022
[11]

The third voiceprivacy challenge: Preserving emotional expres- siveness and linguistic content in voice anonymization,

N. Tomashenko, X. Miao, P. Champion, S. Meyer, M. Panariello, X. Wang, N. Evans, E. Vincent, J. Yamagishi, and M. Todisco, “The third voiceprivacy challenge: Preserving emotional expres- siveness and linguistic content in voice anonymization,”Com- puter Speech & Language, vol. 100, p. 101988, 2026

2026
[12]

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, 2021, pp. 1099–1103

2021
[13]

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 ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 4725– 4729

2024
[14]

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,” in10th ISCA Workshop on Speech Syn- thesis (SSW 10). ISCA, 2019, pp. 155–160

2019
[15]

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

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

2022
[16]

Children’s voice privacy: First steps and emerging challenges,

A. Kulkarni, F. Teixeira, E. Hermann, T. Rolland, I. Trancoso, and M. M. Doss, “Children’s voice privacy: First steps and emerging challenges,” inProc. Interspeech 2025, 2025, pp. 2810–2814

2025
[17]

Speaker anonymization for children’s oral reading assessment,

S. Dhar, S. R. Chetupalli, and P. Rao, “Speaker anonymization for children’s oral reading assessment,” 2026

2026
[18]

Personalized ai-directed tutoring for oral pro- ficiency enhancement in language education,

P. Tushar, B. Zhang, I. Atmosukarto, D. Soh, R. Tong, and I. McLoughlin, “Personalized ai-directed tutoring for oral pro- ficiency enhancement in language education,”Applied Sciences, vol. 16, no. 5, p. 2379, 2026

2026
[19]

A comparison of discrete and soft speech units for improved voice 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 voice conversion,” inICASSP 2022-2022 IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2022, pp. 6562–6566

2022
[20]

Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,

B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” inProc. Interspeech 2020, 2020, pp. 3830–3834

2020
[21]

Privacy and utility of x-vector based speaker anonymiza- tion,

B. M. L. Srivastava, M. Maouche, M. Sahidullah, E. Vincent, A. Bellet, M. Tommasi, N. Tomashenko, X. Wang, and J. Yam- agishi, “Privacy and utility of x-vector based speaker anonymiza- tion,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 2383–2395, 2022

2022
[22]

Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,

J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,”Advances in neural information processing systems, vol. 33, pp. 17 022– 17 033, 2020

2020
[23]

My science tutor (myst)–a large corpus of children’s conversational speech,

S. Pradhan, R. Cole, and W. Ward, “My science tutor (myst)–a large corpus of children’s conversational speech,” inProceedings of the 2024 Joint International Conference on Computational Lin- guistics, Language Resources and Evaluation (LREC-COLING 2024), 2024, pp. 12 040–12 045

2024
[24]

Benchmarking chil- dren’s asr with supervised and self-supervised speech foundation models,

R. Fan, N. Balaji Shankar, and A. Alwan, “Benchmarking chil- dren’s asr with supervised and self-supervised speech foundation models,” inProc. Interspeech 2024, 2024, pp. 5173–5177

2024
[25]

Target speaker extrac- tion with curriculum learning,

Y . Liu, X. Liu, X. Miao, and J. Yamagishi, “Target speaker extrac- tion with curriculum learning,” inProc. Interspeech 2024, 2024, pp. 4348–4352

2024
[26]

Libri2vox dataset: Target speaker extraction with di- verse speaker conditions and synthetic data,

——, “Libri2vox dataset: Target speaker extraction with di- verse speaker conditions and synthetic data,”arXiv preprint arXiv:2412.12512, 2024

work page arXiv 2024
[27]

A dataset and two-pass system for reading miscue detection,

R. Gothi, R. Kumar, M. Pereira, N. Nayak, and P. Rao, “A dataset and two-pass system for reading miscue detection,” inProc. In- terspeech 2024, 2024, pp. 4014–4018

2024
[28]

speechocean762: An open-source non- native english speech corpus for pronunciation assessment,

J. Zhang, Z. Zhang, Y . Wang, Z. Yan, Q. Song, Y . Huang, K. Li, D. Povey, and Y . Wang, “speechocean762: An open-source non- native english speech corpus for pronunciation assessment,” in Proc. Interspeech 2021, 2021

2021
[29]

V oxCeleb2: Deep Speaker Recognition,

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

2018
[30]

Robust speech recognition via large-scale weak supervision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

2023
[31]

NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,

G. Mittag, B. Naderi, A. Chehadi, and S. M ¨oller, “NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,” inProc. Inter- speech 2021, 2021, pp. 2127–2131

2021
[32]

Fine-tune before structured pruning: Towards compact and accurate self-supervised models for speaker diariza- tion,

J. Han, F. Landini, J. Rohdin, A. Silnova, M. Diez, J. ˇCernock`y, and L. Burget, “Fine-tune before structured pruning: Towards compact and accurate self-supervised models for speaker diariza- tion,” inProc. Interspeech 2025, 2025, pp. 1583–1587

2025
[33]

pyannote.audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,

H. Bredin, “pyannote.audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,” inProc. Interspeech 2023, 2023, pp. 1983–1987

2023
[34]

LibriMix: An Open-Source Dataset for Generalizable Speech Separation

J. Cosentino, M. Pariente, S. Cornell, A. Deleforge, and E. Vin- cent, “Librimix: An open-source dataset for generalizable speech separation,”arXiv preprint arXiv:2005.11262, 2020

work page Pith review arXiv 2005
[35]

Montreal forced aligner: Trainable text-speech align- ment using kaldi

M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Son- deregger, “Montreal forced aligner: Trainable text-speech align- ment using kaldi.” inInterspeech, vol. 2017, 2017, pp. 498–502. A. Additional Experimental Details A.1. Comparison of WER The multi-speaker experiments in the main paper evaluate target-speaker intelligibility using pseudo-reference ...

2017

[1] [1]

Introduction V oice anonymization aims to suppress speaker identity while preserving linguistic content and downstream usability. The V oicePrivacy Challenge [1, 2, 3] has established common eval- uation protocols and reference systems, accelerating progress in both signal-processing and neural approaches [4, 5, 6, 7]. How- ever, most systems in this line...

[2] [2]

Child-Centric Voice Anonymization in Single and Multi-Speaker Speech via Domain-Adapted SSL Models

or clinical sessions, frequently involves conversations be- tween children and adults. In such multi-speaker settings, it is desirable to selectively anonymize the target child while option- ally preserving or separately anonymizing the adult speaker. These limitations motivate child-to-child anonymization, where speaker identity is obfuscated while prese...

work page internal anchor Pith review Pith/arXiv arXiv 2026

[3] [3]

Child-centric anonymization 2.1. SSL-based single-speaker child anonymization Given an input waveformx, we extract three disentangled rep- resentations: (i) soft content featurescvia a HuBERT-based encoder [11], (ii) pitch contourf 0, and (iii) a speaker embed- dingsvia ECAPA-TDNN [12]. The content and prosody rep- resentations are kept intact to preserve...

[4] [4]

Datasets and evaluation metrics Single-Speaker Datasets:Table 1 summarizes the datasets used for adaptation and evaluation

Experiments 3.1. Datasets and evaluation metrics Single-Speaker Datasets:Table 1 summarizes the datasets used for adaptation and evaluation. MyST [15] serves as the in- domain child speech corpus; we use the training partition for child-domain adaptation and reserve the test partition for evalu- ation. Zero-shot generalization is assessed on cross-accent ...

work page arXiv 2015

[5] [5]

Challenges and limitations Children’s speech differs substantially from adult speech, in- troducing challenges for both modeling and evaluation. Child recordings often contain higher background noise, classroom reverberation, microphone variability, and spontaneous vocal behaviors such as disfluencies and non-lexical vocalizations. These factors can creat...

[6] [6]

Adapting the HuBERT content encoder and HiFi-GAN vocoder to child speech improves intelligibility while maintain- ing strong privacy protection across datasets

Conclusions and future work This work presented a child-centric study of voice anonymiza- tion through domain adaptation of self-supervised speech mod- els. Adapting the HuBERT content encoder and HiFi-GAN vocoder to child speech improves intelligibility while maintain- ing strong privacy protection across datasets. Human evalu- ations further indicate th...

[7] [7]

Acknowledgments This work is supported by the Singapore Ministry of Education (MOE) Academic Research Fund (AcRF) Tier 1 grant R-R13- A405-0005

[8] [8]

Generative AI Use Disclosure Generative AI tools were used in a limited manner for (i) lan- guage editing during manuscript preparation and (ii) transcrip- tion support for WER-based evaluation using OpenAI ASR models. In addition, AI-generated text-to-speech (TTS) voices (Typecast and SpeechGen) were used to construct the child-like reference speaker poo...

[9] [9]

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, 2020, pp. 1693–1697

2020

[10] [10]

The voiceprivacy 2022 challenge: Progress and perspec- tives in voice anonymisation,

M. Panariello, N. Tomashenko, X. Wang, X. Miao, P. Champion, H. Nourtel, M. Todisco, N. Evans, E. Vincent, and J. Yamag- ishi, “The voiceprivacy 2022 challenge: Progress and perspec- tives in voice anonymisation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 3477–3491, 2024

2022

[11] [11]

The third voiceprivacy challenge: Preserving emotional expres- siveness and linguistic content in voice anonymization,

N. Tomashenko, X. Miao, P. Champion, S. Meyer, M. Panariello, X. Wang, N. Evans, E. Vincent, J. Yamagishi, and M. Todisco, “The third voiceprivacy challenge: Preserving emotional expres- siveness and linguistic content in voice anonymization,”Com- puter Speech & Language, vol. 100, p. 101988, 2026

2026

[12] [12]

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, 2021, pp. 1099–1103

2021

[13] [13]

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 ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 4725– 4729

2024

[14] [14]

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,” in10th ISCA Workshop on Speech Syn- thesis (SSW 10). ISCA, 2019, pp. 155–160

2019

[15] [15]

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

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

2022

[16] [16]

Children’s voice privacy: First steps and emerging challenges,

A. Kulkarni, F. Teixeira, E. Hermann, T. Rolland, I. Trancoso, and M. M. Doss, “Children’s voice privacy: First steps and emerging challenges,” inProc. Interspeech 2025, 2025, pp. 2810–2814

2025

[17] [17]

Speaker anonymization for children’s oral reading assessment,

S. Dhar, S. R. Chetupalli, and P. Rao, “Speaker anonymization for children’s oral reading assessment,” 2026

2026

[18] [18]

Personalized ai-directed tutoring for oral pro- ficiency enhancement in language education,

P. Tushar, B. Zhang, I. Atmosukarto, D. Soh, R. Tong, and I. McLoughlin, “Personalized ai-directed tutoring for oral pro- ficiency enhancement in language education,”Applied Sciences, vol. 16, no. 5, p. 2379, 2026

2026

[19] [19]

A comparison of discrete and soft speech units for improved voice 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 voice conversion,” inICASSP 2022-2022 IEEE In- ternational Conference on Acoustics, Speech and Signal Process- ing (ICASSP). IEEE, 2022, pp. 6562–6566

2022

[20] [20]

Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,

B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,” inProc. Interspeech 2020, 2020, pp. 3830–3834

2020

[21] [21]

Privacy and utility of x-vector based speaker anonymiza- tion,

B. M. L. Srivastava, M. Maouche, M. Sahidullah, E. Vincent, A. Bellet, M. Tommasi, N. Tomashenko, X. Wang, and J. Yam- agishi, “Privacy and utility of x-vector based speaker anonymiza- tion,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 30, pp. 2383–2395, 2022

2022

[22] [22]

Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,

J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial net- works for efficient and high fidelity speech synthesis,”Advances in neural information processing systems, vol. 33, pp. 17 022– 17 033, 2020

2020

[23] [23]

My science tutor (myst)–a large corpus of children’s conversational speech,

S. Pradhan, R. Cole, and W. Ward, “My science tutor (myst)–a large corpus of children’s conversational speech,” inProceedings of the 2024 Joint International Conference on Computational Lin- guistics, Language Resources and Evaluation (LREC-COLING 2024), 2024, pp. 12 040–12 045

2024

[24] [24]

Benchmarking chil- dren’s asr with supervised and self-supervised speech foundation models,

R. Fan, N. Balaji Shankar, and A. Alwan, “Benchmarking chil- dren’s asr with supervised and self-supervised speech foundation models,” inProc. Interspeech 2024, 2024, pp. 5173–5177

2024

[25] [25]

Target speaker extrac- tion with curriculum learning,

Y . Liu, X. Liu, X. Miao, and J. Yamagishi, “Target speaker extrac- tion with curriculum learning,” inProc. Interspeech 2024, 2024, pp. 4348–4352

2024

[26] [26]

Libri2vox dataset: Target speaker extraction with di- verse speaker conditions and synthetic data,

——, “Libri2vox dataset: Target speaker extraction with di- verse speaker conditions and synthetic data,”arXiv preprint arXiv:2412.12512, 2024

work page arXiv 2024

[27] [27]

A dataset and two-pass system for reading miscue detection,

R. Gothi, R. Kumar, M. Pereira, N. Nayak, and P. Rao, “A dataset and two-pass system for reading miscue detection,” inProc. In- terspeech 2024, 2024, pp. 4014–4018

2024

[28] [28]

speechocean762: An open-source non- native english speech corpus for pronunciation assessment,

J. Zhang, Z. Zhang, Y . Wang, Z. Yan, Q. Song, Y . Huang, K. Li, D. Povey, and Y . Wang, “speechocean762: An open-source non- native english speech corpus for pronunciation assessment,” in Proc. Interspeech 2021, 2021

2021

[29] [29]

V oxCeleb2: Deep Speaker Recognition,

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

2018

[30] [30]

Robust speech recognition via large-scale weak supervision,

A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervision,” inInternational conference on machine learning. PMLR, 2023, pp. 28 492–28 518

2023

[31] [31]

NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,

G. Mittag, B. Naderi, A. Chehadi, and S. M ¨oller, “NISQA: A Deep CNN-Self-Attention Model for Multidimensional Speech Quality Prediction with Crowdsourced Datasets,” inProc. Inter- speech 2021, 2021, pp. 2127–2131

2021

[32] [32]

Fine-tune before structured pruning: Towards compact and accurate self-supervised models for speaker diariza- tion,

J. Han, F. Landini, J. Rohdin, A. Silnova, M. Diez, J. ˇCernock`y, and L. Burget, “Fine-tune before structured pruning: Towards compact and accurate self-supervised models for speaker diariza- tion,” inProc. Interspeech 2025, 2025, pp. 1583–1587

2025

[33] [33]

pyannote.audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,

H. Bredin, “pyannote.audio 2.1 speaker diarization pipeline: prin- ciple, benchmark, and recipe,” inProc. Interspeech 2023, 2023, pp. 1983–1987

2023

[34] [34]

LibriMix: An Open-Source Dataset for Generalizable Speech Separation

J. Cosentino, M. Pariente, S. Cornell, A. Deleforge, and E. Vin- cent, “Librimix: An open-source dataset for generalizable speech separation,”arXiv preprint arXiv:2005.11262, 2020

work page Pith review arXiv 2005

[35] [35]

Montreal forced aligner: Trainable text-speech align- ment using kaldi

M. McAuliffe, M. Socolof, S. Mihuc, M. Wagner, and M. Son- deregger, “Montreal forced aligner: Trainable text-speech align- ment using kaldi.” inInterspeech, vol. 2017, 2017, pp. 498–502. A. Additional Experimental Details A.1. Comparison of WER The multi-speaker experiments in the main paper evaluate target-speaker intelligibility using pseudo-reference ...

2017