Rapidly Adapting to New Voice Spoofing: Few-Shot Detection of Synthesized Speech Under Distribution Shifts

Ashi Garg; Henry Li Xinyuan; Kevin Duh; Leibny Paola Garc\'ia-Perera; Matthew Wiesner; Nicholas Andrews; Sanjeev Khudanpur; Zexin Cai

REVIEW 3 major objections 4 minor 40 references

Rapidly Adapting to New Voice Spoofing: Few-Shot Detection of Synthesized Speech Under Distribution Shifts

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a self-attentive prototypical network can adapt a voice spoofing detector to unseen synthesis methods, speakers, languages, or audio conditions using as few as 10 labeled in-distribution samples, cutting equal error r

desk verdict Plausible few-shot spoofing detector with an architecture worth a look, but headline gains are 'up to' and lack variance reporting; needs a referee to verify the averages. read the letter →

arxiv 2508.13320 v1 pith:75FNE6ZF submitted 2025-08-18 eess.AS

Ashi Garg , Zexin Cai , Henry Li Xinyuan , Leibny Paola García-Perera , Kevin Duh , Sanjeev Khudanpur , Matthew Wiesner , Nicholas Andrews This is my paper

classification eess.AS

keywords voicespoofingdetectionsynthesizedspeechfew-shotlearningprototypicalnetworksself-attentiondistributionshiftdeepfakeaudioequalerrorrate

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper addresses a practical failure mode of voice spoofing detectors: they are trained on known synthesis methods and degrade when deployed against new ones, new speakers, new languages, or changed audio conditions. It proposes treating this as a few-shot learning problem and introduces a self-attentive prototypical network that adapts using as few as 10 labeled in-distribution samples. In controlled comparisons, the adapted detector beats the zero-shot baseline by up to 32% relative EER on Japanese deepfakes and 20% relative EER on the ASVspoof 2021 Deepfake dataset. The sympathetic reading is that a small, targeted batch of freshly collected samples can keep a spoofing detector current without retraining on large datasets.

What carries the argument

Self-attentive prototypical network: an encoder plus self-attention pooling produces utterance-level embeddings; class prototypes are computed from a labeled support set; queries are scored by distance to prototypes. The prototype recomputation is the mechanism that lets the detector adapt in a few shots.

What would settle it

Take a new synthesis method not in training, collect 10 labeled samples from the same speakers and recording chain as the test set, and compare equal error rate before and after adaptation. If the relative EER reduction over zero-shot is not consistently positive across several such shifts, the central claim fails.

Watch

Extended reading notes

Core claim

The paper claims that few-shot adaptation is a viable response to distribution shifts in synthesized-speech detection. The proposed self-attentive prototypical network computes an embedding for each utterance, pools frames with self-attention, forms class prototypes from the 10-sample support set, and classifies queries by prototype distance. Because adaptation happens by recomputing prototypes, the model can shift to a new attack or condition without gradient updates. The evidence is a systematic zero-shot versus few-shot comparison under controlled training/test separation, with relative EER reductions of up to 32% on Japanese-language deepfakes and 20% on ASVspoof 2021 Deepfake.

Load-bearing premise

The method assumes a labeled, representative support set of 10 in-distribution samples is available at adaptation time for each new attack or condition.

Editorial extensions

If this is right

Deployed spoofing detectors could be updated on-site with a small labeled batch whenever a new synthesis method appears.
The few-shot formulation gives a standard evaluation protocol: controlled training/test splits that isolate distribution shifts, useful for comparing future detectors.
Adaptation with 10 samples is enough in the tested conditions, so retraining from scratch is not required for these shifts.
Relative EER reductions translate to fewer missed deepfakes at the same false-alarm rate, which matters for biometric authentication.

Reading between the lines

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

The support set is assumed representative; selecting which 10 samples to label (e.g., by diversity or confidence) is a natural next test that could change the gains.
If prototype computation is cheap, the method could be extended to continual adaptation, refreshing prototypes as new attacks stream in.
The same prototype-adaptation idea should transfer to other audio deepfake tasks, such as voice conversion or full audio deepfakes, and to other languages, though the paper only reports Japanese and ASVspoof 2021.

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.

Desk Editor's Note

read the letter

Here's my quick take on arXiv:2508.13320. The paper does two things: it introduces a self-attentive prototypical network for few-shot spoof detection, and it evaluates that against zero-shot baselines under controlled distribution shifts. The framing is sensible—anti-spoofing under unseen synthesis methods is a real problem, and 10 labeled samples as an adaptation budget is a concrete, practical setting. I also give credit for comparing zero-shot and few-shot under the same controlled conditions, which is the right way to isolate the effect of adaptation.

That said, the evidence in the abstract doesn't establish the strength of the result. The headline is "up to 32% relative EER reduction" on Japanese deepfakes and 20% on ASVspoof 2021 Deepfake. "Up to" is doing real work here. With only 10 support samples, the gain may be heavily dependent on which 10 you draw. The abstract reports no variance, number of trials, or whether the support set was random or curated. Without that, I can't tell if the average improvement is 32% or much smaller—possibly negative in some draws. That's a legitimate soft spot, not a fatal one. The stress-test note about labeled support sets is also valid but it's a limitation of any few-shot method, not a flaw specific to this paper.

Another practical concern: the support set requires ground-truth labels at adaptation time. If a new attack just appeared, labeled samples may not exist. That's a deployment caveat the authors should discuss.

I also have to say that the full text we received is corrupted—garbled text, impossible to verify the equations, tables, or baseline numbers. That's a pipeline problem, but it caps what I can endorse from this read. A referee with a clean copy should check whether the prototypical network with self-attention is meaningfully better than a standard prototypical net, and whether the baselines are the same architecture without self-attention.

Net: this is a plausible incremental contribution to the anti-spoofing subfield. I'd accept it for review, and I'd ask the authors to report mean and variance over multiple support-set draws, and to separate the benefit of the architecture from the benefit of adaptation. Not something I'd cite yet, but I'd bring it to reading group if someone wants a concrete few-shot spoofing example.

Referee Report

3 major / 4 minor

Summary. The paper addresses voice spoofing detection under distribution shifts caused by unseen synthesis methods, speakers, languages, or audio conditions. It proposes a self-attentive prototypical network for few-shot adaptation and claims that with as few as 10 in-distribution samples it can achieve up to 32% relative EER reduction on Japanese-language deepfakes and 20% relative EER reduction on ASVspoof 2021 Deepfake, in conditions where zero-shot performance is degraded. The supplied full text is heavily corrupted and unreadable, so I cannot verify the proposed architecture, the experimental protocol, or the numerical results; my assessment rests mainly on the abstract and the surrounding review materials.

Significance. If the empirical claims are substantiated, the paper would make a practically useful contribution: adapting a spoofing detector with a handful of labeled samples could mitigate distribution shifts better than pure zero-shot detection. The focus on few-shot adaptation for anti-spoofing is timely, and the proposed self-attentive prototypical network is a plausible approach. However, the evidence as currently presented is not sufficient for verification: the abstract gives only relative EER improvements in selected conditions, with no absolute numbers, no error bars, no details on support-set selection, and no comparison protocol. No code or reproducible artifacts are provided, so the strength of the paper depends entirely on the experimental report, which is unreadable in the supplied manuscript.

major comments (3)

[Full text (entire manuscript)] The supplied full text is not readable: it consists of corrupted/mojibake text with no legible sections, equations, or tables. I cannot verify the proposed self-attentive prototypical network, the exact training/evaluation protocol, the datasets used, or the reported numerical results. This is a blocking issue for any substantive review. Please provide a cleanly encoded manuscript.
[Abstract] The central quantitative claims are reported as 'up to 32%' and '20%' relative EER reductions, but no absolute EER values, baseline zero-shot EERs, or confidence intervals are given. With only 10 support samples, few-shot adaptation is likely to be high-variance depending on the random draw of the support set; 'up to' suggests possible selection of the best condition or best draw. The paper should report the mean and variance over multiple support-set draws and over all tested evaluation conditions, not only favorable ones, to establish that the improvement is reliable.
[Abstract / experimental setup] The method assumes that a labeled support set of 10 in-distribution samples is available at adaptation time. In real deployment, labels for a newly emerging spoofing attack are often unavailable, so this assumption is load-bearing for the practical claim. The manuscript should explicitly state this assumption, discuss how such labels could be obtained, and ideally compare with unsupervised or self-supervised adaptation or at least analyze sensitivity to support-set quality and composition. Without this, the claimed practical advantage over zero-shot detection is not fully established.

minor comments (4)

[Abstract] Typo: 'upto' should be 'up to'.
[Abstract] 'deepfakes in Japanese language' is vague; please identify the specific corpus/subset and the spoofing attack types used.
[Abstract] The phrase 'traditional zero-shot detectors' should be defined; the full text should state which baseline architectures, features, and training data are used.
[Full text tables] The garbled tables in the full text need careful reformatting; even after fixing the encoding, ensure row/column headers and metric definitions are clear and consistent across tables.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; few-shot adaptation is evaluated on held-out support/test splits.

full rationale

The paper's central claim is an empirical comparison between zero-shot detectors and a few-shot adaptive detector. The few-shot method uses a labeled support set of 10 in-distribution samples to adapt, and is then evaluated on disjoint test data. This is standard supervised few-shot adaptation: the support-set labels are supervision, not the labels being predicted, and no test-set information is used to fit the reported EER reductions. The claimed 32%/20% relative EER improvements are reported as evaluation results rather than being forced by construction. There is no visible self-citation chain, no imported uniqueness theorem, and no fitted parameter that is later renamed as a prediction. The 'up to' wording and the reliance on labeled in-distribution support samples are practical limitations or correctness risks, but they are not circularity. The derivation chain is self-contained as an experimental protocol: train on source conditions, adapt on a small labeled support set from the target condition, test on held-out target data.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The central claim rests primarily on the availability of the support set and the validity of the simulated distribution-shift protocol. No new physical entities are introduced.

free parameters (1)

support set size = 10
The paper uses 10 in-distribution samples for few-shot adaptation; this is a hand-chosen number of examples, not derived from theory.

assumptions (2)

domain assumption Availability of a labeled support set from the target distribution at evaluation time
The few-shot adaptation protocol requires 10 labeled in-distribution samples for each test condition, as stated in the abstract.
domain assumption Controlled training-condition changes faithfully emulate real-world distribution shifts
The evaluation compares zero-shot and few-shot under artificial distribution shifts created by controlling training conditions; whether this generalizes to real unseen spoofing attacks is an unproven premise.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Rapidly Adapting to New Voice Spoofing: Few-Shot Detection of Synthesized Speech Under Distribution Shifts." pith.science (2026). https://pith.science/paper/75FNE6ZF

@misc{pith2026250813320,
  author       = {Pith},
  title        = {Pith review of: Rapidly Adapting to New Voice Spoofing: Few-Shot Detection of Synthesized Speech Under Distribution Shifts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/75FNE6ZF}},
  note         = {Machine review of arXiv:2508.13320}
}

read the original abstract

We address the challenge of detecting synthesized speech under distribution shifts -- arising from unseen synthesis methods, speakers, languages, or audio conditions -- relative to the training data. Few-shot learning methods are a promising way to tackle distribution shifts by rapidly adapting on the basis of a few in-distribution samples. We propose a self-attentive prototypical network to enable more robust few-shot adaptation. To evaluate our approach, we systematically compare the performance of traditional zero-shot detectors and the proposed few-shot detectors, carefully controlling training conditions to introduce distribution shifts at evaluation time. In conditions where distribution shifts hamper the zero-shot performance, our proposed few-shot adaptation technique can quickly adapt using as few as 10 in-distribution samples -- achieving upto 32% relative EER reduction on deepfakes in Japanese language and 20% relative reduction on ASVspoof 2021 Deepfake dataset.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 37 canonical work pages

[1]

Y. Yang, Y. Kartynnik, Y. Li, J. Tang, X. Li, G. Sung, and M. Grundmann, `` StreamVC: Real-Time Low-Latency Voice Conversion ,'' in IEEE International Conference on Acoustics, Speech and Signal Processing, 2024, pp. 11\,016--11\,020

work page 2024
[2]

Kameoka , T

H. Kameoka , T. Kaneko , K. Tanaka , and N. Hojo , `` StarGAN-VC: Non-Parallel Many-to-Many Voice Conversion Using Star Generative Adversarial Networks ,'' in IEEE Spoken Language Technology Workshop, 2018, pp. 266--273

work page 2018
[3]

Choi, S.-H

H.-Y. Choi, S.-H. Lee, and S.-W. Lee, `` DDDM-VC: Decoupled Denoising Diffusion Models with Disentangled Representation and Prior Mixup for Verified Robust Voice Conversion ,'' in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 16, 2024, pp. 17\,862--17\,870

work page 2024
[4]

Casanova, J

E. Casanova, J. Weber, C. D. Shulby, A. C. Junior, E. G \"o lge, and M. A. Ponti, `` YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot Voice Conversion for Everyone ,'' in International Conference on Machine Learning, 2022, pp. 2709--2720

work page 2022
[5]

K. Qian, Y. Zhang, S. Chang, X. Yang, and M. Hasegawa-Johnson, `` A uto VC : Zero-Shot Voice Style Transfer with Only Autoencoder Loss ,'' in International Conference on Machine Learning, vol. 97, 2019, pp. 5210--5219

work page 2019
[6]

Kaneko, H

T. Kaneko, H. Kameoka, K. Tanaka, and N. Hojo, `` CycleGAN-VC2: Improved CycleGAN-Based Non-Parallel Voice Conversion ,'' in IEEE International Conference on Acoustics, Speech and Signal Processing, 2019, pp. 6820--6824

work page 2019
[7]

A. R. Bargum, S. Serafin, and C. Erkut, `` Reimagining Speech: A Scoping Review of Deep Learning-Powered Voice Conversion ,'' Frontiers in signal processing, vol. 4, p. 1339159, 2024

work page 2024
[8]

Cheng and J

Q. Cheng and J. Sorensen, `` Spread Spectrum Signaling for Speech Watermarking ,'' in IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings, vol. 3, 2001, pp. 1337--1340

work page 2001

Show all 40 references

[9]

Faundez-Zanuy, M

M. Faundez-Zanuy, M. Hagm \"u ller, and G. Kubin, ``Speaker verification security improvement by means of speech watermarking,'' Speech communication, vol. 48, no. 12, pp. 1608--1619, 2006

2006
[10]

Dathathri, A

S. Dathathri, A. See, S. Ghaisas, P.-S. Huang, R. McAdam, J. Welbl, V. Bachani, A. Kaskasoli, R. Stanforth, T. Matejovicova et al., `` Scalable Watermarking for Identifying Large Language Model Outputs ,'' Nature, vol. 634, no. 8035, pp. 818--823, 2024

2024
[11]

H. Liu, M. Guo, Z. Jiang, L. Wang, and N. Gong, `` Audiomarkbench: Benchmarking Robustness of Audio Watermarking ,'' Advances in Neural Information Processing Systems, vol. 37, pp. 52\,241--52\,265, 2024

2024
[12]

D'Amour, K

A. D'Amour, K. Heller, D. Moldovan, B. Adlam, B. Alipanahi, A. Beutel, C. Chen, J. Deaton, J. Eisenstein, M. D. Hoffman et al., ``Underspecification presents challenges for credibility in modern machine learning,'' Journal of Machine Learning Research, vol. 23, no. 226, pp. 1-...

2022
[13]

A. Garg, Z. Cai, L. Zhang, H. L. Xinyuan, L. P. García-Perera, K. Duh, S. Khudanpur, M. Wiesner, and N. Andrews, `` ShiftySpeech: A Large-Scale Synthetic Speech Dataset with Distribution Shifts ,'' 2025. [Online]. Available: https://arxiv.org/abs/2502.05674

2025 arXiv
[14]

Vinyals, C

O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstra et al., `` Matching Networks for One Shot Learning ,'' Advances in Neural Information Processing Systems, vol. 29, 2016

2016
[15]

C. Finn, P. Abbeel, and S. Levine, `` Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks ,'' in International Conference on Machine Learning, 2017, pp. 1126--1135

2017
[16]

Snell, K

J. Snell, K. Swersky, and R. Zemel, `` Prototypical Networks for Few-shot Learning ,'' Advances in Neural Information Processing Systems, vol. 30, 2017

2017
[17]

X. Wang, J. Yamagishi, M. Todisco, H. Delgado, A. Nautsch, N. Evans, M. Sahidullah, V. Vestman, T. Kinnunen, K. A. Lee et al., `` ASVspoof 2019: A Large-Scale Public Database of Synthesized, Converted and Replayed Speech ,'' Computer Speech & Language, vol. 64, p. 101114, 2020

2019
[18]

X. Liu, X. Wang, M. Sahidullah, J. Patino, H. Delgado, T. Kinnunen, M. Todisco, J. Yamagishi, N. Evans, A. Nautsch, and K. A. Lee, `` ASVspoof 2021: Towards Spoofed and Deepfake Speech Detection in the Wild ,'' IEEE/ACM Transactions on Audio, Speech, and Language Processing, v...

2021
[19]

H. Tak, M. Todisco, X. Wang, J. weon Jung, J. Yamagishi, and N. Evans, `` Automatic Speaker Verification Spoofing and Deepfake Detection Using Wav2vec 2.0 and Data Augmentation ,'' in Speaker and Language Recognition Workshop, 2022, pp. 112--119

2022
[20]

R. A. R. Soto, K. Koch, A. Khan, B. Y. Chen, M. Bishop, and N. Andrews, `` Few-Shot Detection of Machine-Generated Text using Style Representations ,'' in The Twelfth International Conference on Learning Representations, 2024

2024
[21]

Heggan, S

C. Heggan, S. Budgett, T. Hospedales, and M. Yaghoobi, `` MetaAudio: A Few-Shot Audio Classification Benchmark ,'' in International Conference on Artificial Neural Networks, 2022, pp. 219--230

2022
[22]

Okabe, T

K. Okabe, T. Koshinaka, and K. Shinoda, `` Attentive Statistics Pooling for Deep Speaker Embedding ,'' in Interspeech 2018, 2018, pp. 2252--2256

2018
[23]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, . Kaiser, and I. Polosukhin, `` Attention Is All You Need ,'' Advances in neural information processing systems, vol. 30, 2017

2017
[24]

Conneau, A

A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli, `` Unsupervised Cross-Lingual Representation Learning for Speech Recognition ,'' in Interspeech 2021, 2021, pp. 2426--2430

2021
[25]

J. Ren, S. Fort, J. Liu, A. G. Roy, S. Padhy, and B. Lakshminarayanan, `` A Simple Fix to Mahalanobis Distance for Improving Near-OOD Detection ,'' arXiv preprint arXiv:2106.09022, 2021

2021 arXiv
[26]

Kukanov, J

I. Kukanov, J. Laakkonen, T. Kinnunen, and V. Hautam \"a ki, `` Meta-Learning Approaches For Improving Detection of Unseen Speech Deepfakes ,'' in IEEE Spoken Language Technology Workshop (SLT), 2024, pp. 1173--1178

2024
[27]

2783--2787

Nicolas Müller and Pavel Czempin and Franziska Diekmann and Adam Froghyar and Konstantin Böttinger , `` Does Audio Deepfake Detection Generalize? '' in Interspeech 2022 , 2022 , pp. 2783--2787

2022
[28]

H. Wu, Y. Tseng, and H. yi Lee, `` CodecFake: Enhancing Anti-Spoofing Models Against Deepfake Audios from Codec-Based Speech Synthesis Systems ,'' in Interspeech 2024, 2024, pp. 1770--1774

2024
[29]

Snyder, D

D. Snyder, D. Garcia-Romero, G. Sell, D. Povey, and S. Khudanpur, `` X-Vectors: Robust DNN Embeddings for Speaker Recognition ,'' in IEEE International Conference on Acoustics, Speech and Signal Processing, 2018, pp. 5329--5333

2018
[30]

J. S. Chung, A. Nagrani, and A. Zisserman, `` VoxCeleb2: Deep Speaker Recognition ,'' in Interspeech 2018, 2018, pp. 1086--1090

2018
[31]

X. Xiao, X. Tian, S. Du, H. Xu, E. S. Chng, and H. Li, `` Spoofing Speech Detection Using High Dimensional Magnitude and Phase Features: the NTU Approach for ASVspoof 2015 Challenge ,'' in Interspeech 2015, 2015, pp. 2052--2056

2015
[32]

Wolters, C

P. Wolters, C. Careaga, B. Hutchinson, and L. Phillips, `` A Study of Few-Shot Audio Classification ,'' arXiv preprint arXiv:2012.01573, 2020

2012 arXiv
[33]

Y. Chen, T. Ko, L. Shang, X. Chen, X. Jiang, and Q. Li, `` An Investigation of Few-Shot Learning in Spoken Term Classification ,'' in Interspeech 2020, 2020, pp. 2582--2586

2020
[34]

Liang, B

J. Liang, B. Meyer, I. N. Lee, and T.-T. Do, `` Self-Supervised Learning for Acoustic Few-Shot Classification ,'' in IEEE International Conference on Acoustics, Speech and Signal Processing, 2025, pp. 1--5

2025
[35]

M. Pal, A. Raikar, A. Panda, and S. K. Kopparapu, `` Synthetic Speech Detection Using Meta-Learning with Prototypical Loss ,'' arXiv preprint arXiv:2201.09470, 2022

2022 arXiv
[36]

C. Finn, P. Abbeel, and S. Levine, `` Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks ,'' in International conference on machine learning, 2017, pp. 1126--1135

2017
[37]

J. Yang, R. K. Das, and H. Li, `` Significance of Subband Features for Synthetic Speech Detection ,'' IEEE Transactions on Information Forensics and Security, vol. 15, pp. 2160--2170, 2019

2019
[38]

D. Paul, M. Pal, and G. Saha, `` Spectral Features for Synthetic Speech Detection ,'' IEEE Journal of Selected Topics in Signal Processing, vol. 11, no. 4, pp. 605--617, 2017

2017
[39]

T. B. Patel and H. A. Patil, `` Combining Evidences from Mel Cepstral, Cochlear Filter Cepstral and Instantaneous Frequency Features for Detection of Natural vs. Spoofed Speech ,'' in Interspeech, 2015, pp. 2062--2066

2015
[40]

fl`iw ێTC7eR f T 4 Of L1t/ +.7 g ; & r?lޛ Ӽo޵N焘Ԧ@9k߳4ѴrJL0i L X !*<LR \ Bahs bK)si i N(1 At Q9f[ D i i圼zK J V(@

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

1976

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Y. Yang, Y. Kartynnik, Y. Li, J. Tang, X. Li, G. Sung, and M. Grundmann, `` StreamVC: Real-Time Low-Latency Voice Conversion ,'' in IEEE International Conference on Acoustics, Speech and Signal Processing, 2024, pp. 11\,016--11\,020

work page 2024

[2] [2]

Kameoka , T

H. Kameoka , T. Kaneko , K. Tanaka , and N. Hojo , `` StarGAN-VC: Non-Parallel Many-to-Many Voice Conversion Using Star Generative Adversarial Networks ,'' in IEEE Spoken Language Technology Workshop, 2018, pp. 266--273

work page 2018

[3] [3]

Choi, S.-H

H.-Y. Choi, S.-H. Lee, and S.-W. Lee, `` DDDM-VC: Decoupled Denoising Diffusion Models with Disentangled Representation and Prior Mixup for Verified Robust Voice Conversion ,'' in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 16, 2024, pp. 17\,862--17\,870

work page 2024

[4] [4]

Casanova, J

E. Casanova, J. Weber, C. D. Shulby, A. C. Junior, E. G \"o lge, and M. A. Ponti, `` YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot Voice Conversion for Everyone ,'' in International Conference on Machine Learning, 2022, pp. 2709--2720

work page 2022

[5] [5]

K. Qian, Y. Zhang, S. Chang, X. Yang, and M. Hasegawa-Johnson, `` A uto VC : Zero-Shot Voice Style Transfer with Only Autoencoder Loss ,'' in International Conference on Machine Learning, vol. 97, 2019, pp. 5210--5219

work page 2019

[6] [6]

Kaneko, H

T. Kaneko, H. Kameoka, K. Tanaka, and N. Hojo, `` CycleGAN-VC2: Improved CycleGAN-Based Non-Parallel Voice Conversion ,'' in IEEE International Conference on Acoustics, Speech and Signal Processing, 2019, pp. 6820--6824

work page 2019

[7] [7]

A. R. Bargum, S. Serafin, and C. Erkut, `` Reimagining Speech: A Scoping Review of Deep Learning-Powered Voice Conversion ,'' Frontiers in signal processing, vol. 4, p. 1339159, 2024

work page 2024

[8] [8]

Cheng and J

Q. Cheng and J. Sorensen, `` Spread Spectrum Signaling for Speech Watermarking ,'' in IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings, vol. 3, 2001, pp. 1337--1340

work page 2001