Pith. sign in

REVIEW 3 major objections 4 minor 39 references

NouveauVoice: Generating Novel Pseudo Speakers for Voice Anonymization

T0 review · 3 major / 4 minor · reviewed 2026-07-11 · grok-4.5

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

Pith's one-line read A hierarchical variational autoencoder generates diverse pseudo-speakers that anonymize voices while keeping speech usable.

desk verdict Solid plug-in that turns FACodec/CosyVoice2 into anonymizers with a real diversity gain and a usable privacy dial; the headline EER rests on an uninformed attacker only. read the letter →

arxiv 2607.03985 v1 pith:S73JAMUB submitted 2026-07-04 eess.AS cs.AI

Meiying Melissa Chen , Anastasia Kuznetsova , Zhenyu Wang , Zhiyao Duan This is my paper
classification eess.AScs.AI
keywords voiceanonymizationpseudo-speakergenerationhierarchicalVAENVAEspeakerembeddingsconversionprivacy-utilitytrade-offequalerrorrate
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

Neural voice cloning makes it easy to steal a person's identity from a short recording. This paper claims that a hierarchical deep variational autoencoder can learn the distribution of real speaker embeddings and then sample new, never-before-heard identities that still sound natural. The resulting generator is a drop-in module for existing voice-conversion systems; once the original speaker embedding is replaced by a sampled one, the converted speech is hard for an automatic verifier to link back to the source speaker. At the same time the method keeps word error rate and emotion-recognition scores close to those of ordinary reconstruction, and it produces more varied voices than a simple Gaussian-mixture baseline. The hierarchy also lets a user dial the strength of anonymization by deciding how many latent layers to randomize.

What carries the argument

The Nouveau VAE (NVAE) hierarchy: speaker embeddings are encoded into ordered latent groups that are sampled coarse-to-fine under the ELBO objective, then decoded into a new embedding that replaces the original identity cue inside a voice-conversion pipeline.

What would settle it

Retrain or fine-tune the same ECAPA-TDNN verifier on a large set of NouveauVoice-anonymized utterances and re-measure EER; a large drop below 38 percent would falsify the claim of strong identity concealment.

Watch

Extended reading notes

Core claim

NouveauVoice, a hierarchical NVAE trained on speaker embeddings, produces pseudo-speaker vectors that, when plugged into FACodec or CosyVoice2, raise the equal-error rate of an automatic speaker verifier above 38 percent while preserving intelligibility and emotional expressiveness better than a GMM baseline and while increasing speaker diversity.

Load-bearing premise

Privacy is measured only against a fixed, pre-trained speaker verifier that never sees anonymized speech, so the reported protection may not hold once an attacker adapts to the anonymization method.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes NouveauVoice, a hierarchical deep variational autoencoder (NVAE) for generating novel pseudo-speaker embeddings. Trained independently via the ELBO on real speaker embeddings (with free-bits and KL warmup), it serves as a plug-in module that replaces the speaker embedding input of existing VC systems (FACodec and CosyVoice2). Hierarchical sampling from the prior produces diverse embeddings; a progressive latent-group replacement scheme further allows control of anonymization strength. Under a VoicePrivacy-like protocol the authors report EER >38% against a fixed ECAPA-TDNN ASV attacker, together with WER, UAR, top-5 cosine similarity and MMD, claiming a favorable privacy–utility–diversity trade-off relative to a GMM baseline.

Significance. If the privacy numbers survive stronger attackers, NouveauVoice would be a useful modular contribution: a drop-in generator that improves speaker diversity over simple GMM or prompt-pooling methods while offering explicit, layer-wise control of anonymization strength. Concrete strengths include dual-backend validation, the progressive-replacement ablation (Fig. 2), and the joint reporting of diversity (Table II) alongside privacy and utility. These elements are reproducible in principle and address a recognized gap in existing SAS pipelines. The significance of the headline EER claim is currently limited by the threat model, but the architectural idea and experimental design remain of clear interest to the voice-privacy community.

major comments (3)
  1. [§IV-A, Table I, Fig. 2] The central claim of “strong identity concealment (EER exceeding 38 %)” (abstract, §V-A, Table I, Fig. 2) is obtained exclusively against an uninformed attacker—a fixed ECAPA-TDNN pretrained on VoxCeleb that never sees anonymized speech (§IV-A). The manuscript itself cites prior work [16] and the VoicePrivacy Challenge evaluation plans [17,18] that treat semi-informed / informed attackers (retrained or adapted on anonymized data) as the realistic standard; such attackers are known to collapse EER for many SAS methods. No semi-informed or informed evaluation is reported for NouveauVoice or the GMM baseline. This is load-bearing for every privacy number advertised; without it the quantitative guarantee does not hold under the community threat model.
  2. [§V-A, Table I, Fig. 2] Table I presents single-point EER/WER/UAR estimates on 1 000 utterances with no error bars, multiple random seeds, or statistical tests. Consequently the claimed “more favorable overall balance” of FACodec-NV versus FACodec-GMM (38.26 % vs 42.30 % EER, 7.56 % vs 9.90 % WER, 40.36 % vs 38.59 % UAR) cannot be assessed for reliability. The same limitation applies to the progressive-layer curves in Fig. 2. These differences underwrite the practical advantage asserted for the method.
  3. [§I, §III, Table I] The introduction positions NouveauVoice against VPC baselines B3 (GAN) and B4 (NAC) and against StreamVoiceAnon, yet the experimental section reports only an internal GMM baseline. Without side-by-side EER/WER/UAR numbers under a common protocol it is impossible to judge relative progress on the metrics the community already uses.
minor comments (4)
  1. [passim] Throughout the extracted text many compound terms appear with spurious spaces (“NouveauV oice”, “V oicePrivacy”, “CosyV oice2”). These should be corrected for the camera-ready version.
  2. [§II-A] Hyper-parameters of the free-bits threshold, KL-warmup schedule and quantile normalization are mentioned (§II-A) but never given numerical values, hindering exact reproduction.
  3. [Fig. 1, §II-B] Figure 1 caption and the residual-connection description could more clearly indicate which latent groups are sampled versus reconstructed at inference time.
  4. [footnote 1] The demo link is promised “upon acceptance”; a temporary anonymous repository would strengthen the submission.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: NVAE is trained via standard ELBO on real embeddings and evaluated with external ASV/ASR/SER models on held-out data.

full rationale

The paper's central claims rest on an independently trained hierarchical VAE (NVAE) whose ELBO objective (Eq. 2) reconstructs speaker embeddings extracted from LibriTTS via external encoders (FACodec timbre extractor or CAM++). Sampling at inference draws from the learned hierarchical prior p_θ(z_l | z_<l) without reference to any test speaker or evaluation metric. Privacy (EER), utility (WER, UAR), and diversity (Top-5 cosine, MMD) numbers are produced by fixed external models (ECAPA-TDNN, wav2vec2 ASR/SER) on held-out LibriTTS/IEMOCAP utterances; none of these quantities is fitted or defined in terms of the others. The sole author self-citation ([4] ControlVC) appears only as related work and is not load-bearing for any equation, uniqueness claim, or reported number. Progressive latent-group ablation (Experiment 2) is an empirical sensitivity study, not a derivation that reduces by construction. The derivation chain is therefore self-contained against external benchmarks.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on standard VAE theory plus domain assumptions about speaker embeddings and the uninformed-attacker threat model. Free parameters are ordinary architectural and training choices; no new physical constants or ad-hoc scaling factors are introduced to force the privacy numbers. The only invented entity is the NouveauVoice module itself, which is an engineering construct rather than a new scientific object.

free parameters (4)
  • number of latent groups L = 8
    Fixed at 8; controls the hierarchy depth and the granularity of the progressive-replacement experiment. Chosen by authors, not derived.
  • GMM components k and covariance type = k=16, diagonal
    Baseline uses k=16 diagonal Gaussians; these choices affect the diversity and MMD comparison numbers.
  • free-bits threshold and KL warmup schedule
    Standard VAE regularizers introduced to prevent posterior collapse; values are not reported precisely and affect how well the prior matches the real embedding distribution.
  • quantile-based normalization parameters
    Applied to speaker embeddings before NVAE training; the exact quantiles are not stated and influence the scale of the latent space.
assumptions (4)
  • domain assumption ELBO maximization with hierarchical conditional priors yields samples whose distribution is close enough to real speaker embeddings for natural-sounding synthesis.
    Invoked throughout Section II; standard VAE theory but not guaranteed for the particular speaker-embedding manifold.
  • domain assumption An uninformed ECAPA-TDNN attacker pretrained on VoxCeleb is a representative privacy threat.
    Section IV-A; underpins every EER claim. Prior VoicePrivacy work shows informed attackers are stronger.
  • domain assumption Speaker identity is sufficiently captured by the fixed-dimensional embeddings of FACodec's timbre extractor and CosyVoice2's CAM++.
    Section III; if residual identity leaks through content or prosody tokens, the plug-in cannot fully anonymize.
  • standard math Standard hierarchical VAE math (Eqs. 1-2) and residual spectral-normalized cells.
    Taken from Vahdat & Kautz 2020; used without modification except 1-D convolutions.
invented entities (1)
  • NouveauVoice (NVAE-based pseudo-speaker generator plug-in)
    purpose: Standalone module that samples novel speaker embeddings for any VC/TTS backend that accepts an embedding vector.
    Engineering construct; no claim of a new scientific object beyond the trained model. independent_evidence is false because the only evidence is the paper's own experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NouveauVoice: Generating Novel Pseudo Speakers for Voice Anonymization." pith.science (2026). https://pith.science/paper/S73JAMUB

@misc{pith2026260703985,
  author       = {Pith},
  title        = {Pith review of: NouveauVoice: Generating Novel Pseudo Speakers for Voice Anonymization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S73JAMUB}},
  note         = {Machine review of arXiv:2607.03985}
}
read the original abstract

Advanced neural technologies in speech synthesis and voice conversion (VC) have introduced severe risks to personal privacy, necessitating robust Speaker Anonymization Systems (SAS). Existing SAS approaches modify voice characteristics in the hand-crafted feature space or speaker embedding space, often struggling to provide sufficient identity variance across generated voices. In this paper, we propose NouveauVoice, a novel pseudo-speaker generation framework based on a Hierarchical Deep Variational Autoencoder (NVAE). Integrated as a standalone plug-in module on top of state-of-the-art architectures (FACodec and CosyVoice2), our approach leverages tractable sampling and the Evidence Lower Bound (ELBO) objective to synthesize highly expressive pseudo-speaker embeddings with significantly enhanced speaker diversity. Evaluating our framework under a protocol similar to the VoicePrivacy Challenge alongside Maximum Mean Discrepancy (MMD) analysis, we demonstrate that NouveauVoice achieves strong identity concealment, yielding an Equal Error Rate (EER) exceeding 38% against an automatic speaker verification attacker model. Our system shows a reasonable trade-off between strict anonymity, rich pseudo-speaker diversity, and downstream speech utility, such as intelligibility and emotional expressiveness.

Figures

Figures reproduced from arXiv: 2607.03985 by the authors.

Figure 1
Figure 1. NouveauVoice generator overview. Left side shows the encoder structure, while right side depicts the decoder structure. Dashed lines between encoder [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Effect of the number (N) of randomized groups of the NVAE latents on privacy (EER↑) and utility (WER↓ & UAR↑). Autoencoder. Designed as a standalone plug-in module, it integrates seamlessly with state-of-the-art architectures with minimal training. The framework generates highly expressive, diverse pseudo-speaker embeddings for use in the anonymiza￾tion process. Evaluation results demonstrate that Nouveau￾Voice achi… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 9 linked inside Pith

  1. [16]

    Speaker anonymiza- tion using orthogonal householder neural network,

    X. Miao, Y . Liang, L. Xie, M. Wang, and J. Wei, “Speaker anonymiza- tion using orthogonal householder neural network,”IEEE/ACM Trans- actions on Audio, Speech, and Language Processing, vol. 31, pp. 3681– 3694, 2023

    2023

  2. [1]

    Naturalspeech 3: Zero-shot speech synthesis with factorized codec and diffusion models,

    Zeqian Ju, Yuancheng Wang, Kai Shen, Xu Tan, Detai Xin, Dongchao Yang, Yanqing Liu, Yichong Leng, Kaitao Song, Siliang Tang, et al., “Naturalspeech 3: Zero-shot speech synthesis with factorized codec and diffusion models,”arXiv preprint arXiv:2403.03100, 2024

    arXiv 2024

  3. [2]

    Cosyvoice 2: Scalable streaming speech synthesis with large language models,

    Zhihao Du, Yuxuan Wang, Qian Chen, Xian Shi, Xiang Lv, Tianyu Zhao, Zhifu Gao, Yexin Yang, Changfeng Gao, Hui Wang, et al., “Cosyvoice 2: Scalable streaming speech synthesis with large language models,” arXiv preprint arXiv:2412.10117, 2024

    arXiv 2024

  4. [3]

    Spark-tts: An efficient llm-based text-to-speech model with single-stream decoupled speech tokens,

    Xinsheng Wang, Mingqi Jiang, Ziyang Ma, Ziyu Zhang, Songxiang Liu, Linqin Li, Zheng Liang, Qixi Zheng, Rui Wang, Xiaoqin Feng, Weizhen Bian, Zhen Ye, Sitong Cheng, Ruibin Yuan, Zhixian Zhao, Xinfa Zhu, Jiahao Pan, Liumeng Xue, Pengcheng Zhu, Yunlin Chen, Zhifei Li, Xie Chen, Lei Xie, Yike Guo, and Wei Xue, “Spark-tts: An efficient llm-based text-to-speech...

    2025

  5. [4]

    Controlvc: Zero-shot voice conver- sion with time-varying controls on pitch and speed,

    Meiying Chen and Zhiyao Duan, “Controlvc: Zero-shot voice conver- sion with time-varying controls on pitch and speed,”arXiv preprint arXiv:2209.11866, 2022

    arXiv 2022

  6. [5]

    When large language models meet speech: A survey on integration approaches,

    Zhengdong Yang, Shuichiro Shimizu, Yahan Yu, and Chenhui Chu, “When large language models meet speech: A survey on integration approaches,” inFindings of the Association for Computational Linguis- tics: ACL 2025, 2025, pp. 20298–20315

    2025

  7. [6]

    Kanade: A simple disentangled tokenizer for spoken language modeling,

    Zhijie Huang, Stephen McIntosh, Daisuke Saito, and Nobuaki Mine- matsu, “Kanade: A simple disentangled tokenizer for spoken language modeling,”arXiv preprint arXiv:2602.00594, 2026

    arXiv 2026

  8. [7]

    Exploring Disentangled Neural Speech Codecs from Self-Supervised Representations,

    Ryo Aihara, Yoshiki Masuyama, Franc ¸ois G. Germain, Gordon Wich- ern, and Jonathan Le Roux, “Exploring Disentangled Neural Speech Codecs from Self-Supervised Representations,” inIEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), May 2026

    2026

Show all 39 references
  1. [8]

    Dis- crete audio tokens: More than a survey!,

    Pooneh Mousavi, Gallil Maimon, Adel Moumen, Darius Petermann, Jiatong Shi, Haibin Wu, Haici Yang, Anastasia Kuznetsova, Artem Ploujnikov, Ricard Marxer, Bhuvana Ramabhadran, Benjamin Elizalde, Loren Lugosch, Jinyu Li, Cem Subakan, Phil Woodland, Minje Kim, Hung yi Lee, Shinji ...

    2025

  2. [9]

    AutoVC: Zero-shot voice style transfer with only autoencoder loss,

    Kaizhi Qian, Yang Zhang, Shiyu Chang, Xuesong Yang, and Mark Hasegawa-Johnson, “AutoVC: Zero-shot voice style transfer with only autoencoder loss,” inProceedings of the 36th International Conference on Machine Learning. 2019, vol. 97 ofProceedings of Machine Learning Research,...

    2019

  3. [10]

    A comparative study of self-supervised speech representation based voice conversion,

    Wen-Chin Huang, Shu-Wen Yang, Tomoki Hayashi, and Tomoki Toda, “A comparative study of self-supervised speech representation based voice conversion,”IEEE Journal of Selected Topics in Signal Processing, vol. 16, no. 6, pp. 1308–1318, 2022

    2022

  4. [11]

    Vevo: Controllable zero-shot voice imitation with self- supervised disentanglement,

    Xueyao Zhang, Xiaohui Zhang, Kainan Peng, Zhenyu Tang, Vimal Manohar, Yingru Liu, Jeff Hwang, Dangna Li, Yuhao Wang, Julian Chan, et al., “Vevo: Controllable zero-shot voice imitation with self- supervised disentanglement,”arXiv preprint arXiv:2502.07243, 2025

    2025 arXiv

  5. [12]

    Stargan-VC: Non-parallel many-to-many voice conversion using star generative adversarial networks,

    Hirokazu Kameoka, Takuhiro Kaneko, Kou Tanaka, and Nobukatsu Hojo, “Stargan-VC: Non-parallel many-to-many voice conversion using star generative adversarial networks,” in2018 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2018, pp. 266–273

    2018

  6. [13]

    Ace-VC: Adaptive and controllable voice conversion using explicitly disentangled self-supervised speech representations,

    Shehzeen Hussain, Paarth Neekhara, Jocelyn Huang, Jason Li, and Boris Ginsburg, “Ace-VC: Adaptive and controllable voice conversion using explicitly disentangled self-supervised speech representations,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and...

    2023

  7. [14]

    Speaker Anonymisation Using the McAdams Coefficient,

    Jose Patino, Natalia Tomashenko, Massimiliano Todisco, Andreas Nautsch, and Nicholas Evans, “Speaker Anonymisation Using the McAdams Coefficient,” inInterspeech 2021, 2021, pp. 1099–1103

    2021

  8. [15]

    Improving speaker de-identification with functional data analysis of f0 trajectories,

    Lauri Tavi, Tomi Kinnunen, and Rosa Gonz ´alez Hautam¨aki, “Improving speaker de-identification with functional data analysis of f0 trajectories,” Speech Commun., vol. 140, no. C, pp. 1–10, June 2026

    2026

  9. [17]

    The V oicePrivacy 2024 challenge evaluation plan,

    Natalia Tomashenko, Xiaoxiao Miao, Pierre Champion, Sarina Meyer, Xin Wang, Emmanuel Vincent, Michele Panariello, Nicholas Evans, Junichi Yamagishi, and Massimiliano Todisco, “The V oicePrivacy 2024 challenge evaluation plan,” 2024

    2024

  10. [18]

    The third V oicePrivacy challenge: Preserving emotional expressiveness and linguistic content in voice anonymization,

    Natalia Tomashenko, Xiaoxiao Miao, Pierre Champion, Sarina Meyer, Xin Wang, Emmanuel Vincent, Michele Panariello, Nicholas Evans, Junichi Yamagishi, and Massimiliano Todisco, “The third V oicePrivacy challenge: Preserving emotional expressiveness and linguistic content in voic...

    2026

  11. [19]

    Stream-voice-anon: Enhancing utility of real-time speaker anonymization via neural audio codec and language models,

    N. Kuzmin, S. Liu, K. A. Lee, and E. S. Chng, “Stream-voice-anon: Enhancing utility of real-time speaker anonymization via neural audio codec and language models,”arXiv preprint arXiv:2601.13948, 2026

    2026

  12. [20]

    Speech pseudonymisation assessment using voice similarity matrices,

    Paul-Gauthier No ´e, Jean-Franc ¸ois Bonastre, Driss Matrouf, Natalia Tomashenko, Andreas Nautsch, and Nicholas Evans, “Speech pseudonymisation assessment using voice similarity matrices,” in Proceedings of the Annual Conference of the International Speech Communication Associ...

    2020

  13. [21]

    Nvae: A deep hierarchical variational autoencoder,

    Arash Vahdat and Jan Kautz, “Nvae: A deep hierarchical variational autoencoder,”Advances in neural information processing systems, vol. 33, pp. 19667–19679, 2020

    2020

  14. [22]

    Robust stochastic optimization via gradient quantile clipping,

    Ibrahim Merad and St ´ephane Ga¨ıffas, “Robust stochastic optimization via gradient quantile clipping,”arXiv preprint arXiv:2309.17316, 2023

    2023 arXiv

  15. [23]

    Improved variational inference with inverse autoregressive flow,

    Durk P Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling, “Improved variational inference with inverse autoregressive flow,”Advances in neural information processing systems, vol. 29, 2016

    2016

  16. [24]

    beta-vae: Learning basic visual concepts with a constrained variational framework,

    Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner, “beta-vae: Learning basic visual concepts with a constrained variational framework,” inInternational Conference on Learning Representations, 2017

    2017

  17. [25]

    Neural discrete representation learning,

    Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu, “Neural discrete representation learning,” inAdvances in Neural Information Processing Systems, 2017, vol. 30, pp. 6306–6315

    2017

  18. [26]

    Unsupervised domain adaptation by backpropagation,

    Yaroslav Ganin and Victor Lempitsky, “Unsupervised domain adaptation by backpropagation,” inInternational conference on machine learning. PMLR, 2015, pp. 1180–1189

    2015

  19. [27]

    Libritts: A corpus derived from librispeech for text-to-speech,

    Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu, “Libritts: A corpus derived from librispeech for text-to-speech,”arXiv preprint arXiv:1904.02882, 2019

    1904 arXiv

  20. [28]

    Cam++: A fast and efficient network for speaker verification using context-aware masking,

    Hui Wang, Siqi Zheng, Yafeng Chen, Luyao Cheng, and Qian Chen, “Cam++: A fast and efficient network for speaker verification using context-aware masking,”arXiv preprint arXiv:2303.00332, 2023

    2023 arXiv

  21. [29]

    Cosyvoice,

    FunAudioLLM, “Cosyvoice,” 2024, Accessed: 2024-12-05

    2024

  22. [30]

    Gaussian mixture models.,

    Douglas A Reynolds et al., “Gaussian mixture models.,”Encyclopedia of biometrics, vol. 741, no. 659-663, pp. 3, 2009

    2009

  23. [31]

    Speaker generation,

    Daisy Stanton, Matt Shannon, Soroosh Mariooryad, RJ Skerry-Ryan, Eric Battenberg, Tom Bagby, and David Kao, “Speaker generation,” in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 7897–7901

    2022

  24. [32]

    Scikit-learn: Machine learning in python,

    Fabian Pedregosa, Ga ¨el Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Pret- tenhofer, Ron Weiss, Vincent Dubourg, et al., “Scikit-learn: Machine learning in python,”Journal of machine learning research, vol. 12, pp. 28...

    2011

  25. [33]

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

    Brecht Desplanques, Jenthe Thienpondt, and Kris Demuynck, “Ecapa- tdnn: Emphasized channel attention, propagation and aggregation in tdnn based speaker verification,”arXiv preprint arXiv:2005.07143, 2020

    2005 arXiv

  26. [34]

    V oxceleb: A large-scale speaker identification dataset,

    Arsha Nagrani, Joon Son Chung, and Andrew Zisserman, “V oxceleb: A large-scale speaker identification dataset,” inINTERSPEECH, 2017

    2017

  27. [35]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,”Advances in neural information processing systems, vol. 33, pp. 12449–12460, 2020

    2020

  28. [36]

    Librispeech: An asr corpus based on public domain audio books,

    V . Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur, “Librispeech: An asr corpus based on public domain audio books,” in 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2015, pp. 5206–5210

    2015

  29. [37]

    SUPERB: Speech Processing Universal PERfor- mance Benchmark,

    Shu wen Yang et al., “SUPERB: Speech Processing Universal PERfor- mance Benchmark,” inProc. Interspeech 2021, 2021, pp. 1194–1198

    2021

  30. [38]

    IEMOCAP: Interactive emotional dyadic motion capture database,

    Carlos Busso, Murtaza Bulut, Chi-Chun Lee, Abe Kazemzadeh, Emily Mower, Samuel Kim, Jeannette N Chang, Sungbok Lee, and Shrikanth S Narayanan, “IEMOCAP: Interactive emotional dyadic motion capture database,”Language Resources and Evaluation, vol. 42, no. 4, pp. 335– 359, 2008

    2008

  31. [39]

    A kernel two-sample test,

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Sch¨olkopf, and Alexander Smola, “A kernel two-sample test,”The journal of machine learning research, vol. 13, no. 1, pp. 723–773, 2012

    2012

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.