Pith. sign in

REVIEW 4 major objections 5 minor 42 references

Few-Shot Speech Deepfake Detection Adaptation with Gaussian Processes

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 The paper claims that a Gaussian-process classifier on XLS-R speech embeddings adapts to a new, unseen voice-cloning TTS with as few as 100 samples, reducing equal error rate on 11Labs from 5.76% to 0.54% while preserving in-distribution…

desk verdict Solid few-shot adaptation result with a useful new benchmark, but the headline margin over the best baseline is smaller once you read Table 1 correctly, and the single OOD TTS was chosen post hoc. read the letter →

arxiv 2505.23619 v1 pith:Z3OYMAW3 submitted 2025-05-29 cs.SD cs.LGeess.AS

Neta Glazer , David Chernin , Idan Achituve , Sharon Gannot , Ethan Fetaya This is my paper
classification cs.SDcs.LGeess.AS
keywords audiodeepfakedetectionfew-shotadaptationGaussianprocessclassificationdeepkernellearningvoicecloningXLS-RembeddingsMixProaugmentationpersonalized
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 claims that audio deepfake detection can be adapted to a new, unseen text-to-speech (TTS) voice-cloning model with only a handful of samples, by replacing the usual fine-tuned neural classifier with a Gaussian process (GP) classifier on top of frozen, learned speech embeddings. The proposed system, ADD-GP, trains a deep kernel once on four TTS models, then adapts to a fifth unseen model simply by adding new audio examples to the GP's support set, with no gradient updates. On the unseen 11Labs TTS, 100-shot adaptation brings the equal error rate down to 0.54%, compared with 5.76% for the best baseline. The paper also reports one-shot personalized detection at 0.61% EER and well-calibrated confidence scores, and introduces LibriFake, a benchmark built from LibriSpeech with five voice-cloning TTS models, to make such few-shot evaluations reproducible.

What carries the argument

The central object is ADD-GP: a Dirichlet-based Gaussian process classifier used as the back-end, with an RBF kernel applied to embeddings from XLS-R, a self-supervised speech model that is frozen except for its last block. During training the kernel length scale, output scale, and the last XLS-R block are learned jointly by maximizing the log marginal likelihood over batches of 80 examples. At evaluation time the learned kernel is fixed on 1000 held-out training examples, and few-shot adaptation consists only of appending new TTS samples to that support set and recomputing the GP posterior in closed form, with no gradient steps. MixPro augmentation creates additional support points by interpolating between embeddings of fake samples from the source TTS models and fake samples from the new TTS model, which is what makes the fixed-kernel adaptation work with very few labeled examples.

What would settle it

Run the identical 100-shot protocol on a second previously unseen TTS model, for example a newly released voice-cloning system not among yourTTS, Whisper-Speech, Vall-e-x, F5-TTS, or 11Labs; if ADD-GP MixPro's EER is not substantially below the best continual-learning baseline (RWM MixPro) or remains in double digits, the claim that a fixed XLS-R kernel transfers across generators is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a Gaussian process classifier with a deep RBF kernel over XLS-R embeddings is a better substrate for few-shot adaptation than fine-tuning a large pretrained ADD model, because the kernel encodes a similarity structure over real and fake speech that transfers across TTS models. Once trained on TTS models yourTTS, Whisper-Speech, Vall-e-x, and F5-TTS, ADD-GP keeps the kernel fixed and adapts to the unseen 11Labs model by augmenting its support set with k fake and real examples; with 100 shots and MixPro interpolation this reaches 0.54% EER versus 5.76% for SSL-AASIST-FT with MixPro. The paper further claims this approach is uniquely non-forgetting: after 100-shot adaptation, ADD-GP's in-distribution EER stays at 0.1% while fine-tuned baselines degrade. In the personalized setting, a per-speaker detector built from 20 samples reaches 0.61% EER after one 11Labs example on LibriFake, and GP probabilities are well calibrated, unlike the uncalibrated scores produced by baselines.

Load-bearing premise

The whole few-shot benefit rests on the assumption that the similarity structure (RBF kernel) learned on XLS-R features from four TTS models transfers to a new, unseen TTS model without being refit; the paper tests this on only a single out-of-distribution model, 11Labs.

Editorial extensions

If this is right

  • At 100 labeled examples from a new TTS model, ADD-GP with MixPro reports 0.54% EER on 11Labs, versus 5.76% for the best adapted baseline.
  • Because adaptation only appends support points, a deployed system can update on new TTS samples without gradient updates, and after 100-shot adaptation its in-distribution EER remains at 0.1% while fine-tuned baselines degrade.
  • A personalized detector trained on 20 samples of a single speaker reaches 0.61% EER after one 11Labs fake sample on LibriFake, roughly matching the non-personalized 100-shot result.
  • The GP returns calibrated probabilities, so users can set confidence thresholds rather than relying on uncalibrated binary scores.

Reading between the lines

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

  • The paper does not test sequential adaptation across multiple new TTS models one after another; a natural stress test would be to keep adding support points from successive unseen generators and measure whether old-model performance stays flat, which would show whether the non-parametric design truly prevents catastrophic forgetting in deployment.
  • Because the kernel is never refit, the method's promise hinges on XLS-R embedding geometry being universal across generators; a cheap probe is to apply the same 100-shot protocol to a different kind of synthetic speech, such as voice conversion or vocoder-only artifacts, and see whether the fixed kernel still separates them from real speech.
  • The MixPro trick suggests a broader recipe for non-parametric detectors generally: interpolate source- and target-domain embeddings of the same class to synthesize support points, which could transfer to other spoofing-detection settings beyond TTS.
  • Calibrated probabilities open a deployment route the paper only mentions briefly: in fraud triage, low-confidence predictions could be routed to human review or additional verification instead of forcing a binary accept or reject decision.
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 introduces ADD-GP, a Gaussian Process classifier with an XLS-R deep kernel for audio deepfake detection, and evaluates its few-shot adaptation to a previously unseen TTS model (11Labs) on a new LibriFake benchmark. The method keeps the learned kernel fixed and adapts by augmenting the support set with target-domain samples, optionally using MixPro interpolation. Experiments report strong improvements over fine-tuned baselines, including 0.54% EER after 100-shot adaptation with MixPro on 11Labs, and additional gains in a personalized detection setting. The authors release code and data.

Significance. If the claims hold, the paper makes a useful contribution: a non-parametric back-end that adapts without gradient updates is an appealing direction for audio deepfake detection, where new TTS models appear rapidly and labeled data are scarce. The LibriFake dataset and the focus on calibration are also valuable. The main experimental result is plausible and the approach is well motivated. However, the strength of the central generalization claim is currently limited by the use of a single out-of-distribution TTS that was selected post hoc because it was the hardest, and by the absence of statistical comparison against the strongest baseline.

major comments (4)
  1. [§3.1] The out-of-distribution TTS is chosen post hoc: the authors write that they 'found that detection performance was significantly worse on 11Labs. Therefore, we selected it as the unseen TTS.' This selection bias means the few-shot adaptation results, including the 0.54% EER in Table 1, are not necessarily representative of performance on a randomly held-out TTS. The central claim of adapting to 'previously unseen generation models' needs validation on at least one additional held-out TTS, or a pre-registered random holdout, to support generalization.
  2. [§4.2, Table 1] The improvement of ADD-GP over the strongest baseline is smaller than the paper's narrative suggests. At 100-shot with MixPro, RWM MixPro achieves 1.24% EER, not the 5.76% of SSL-AASIST-FT MixPro. ADD-GP at 0.54% is still better, but the relative gain is about 2.3x, not the larger margin implied by comparing to SSL-AASIST. The paper should explicitly compare against the best-performing baseline and report pairwise significance tests; the standard deviations (e.g., 0.03 vs 0.05) suggest the difference is likely significant, but this is not stated.
  3. [§3.2, §3.3] The adaptation mechanism relies on the assumption that the RBF kernel learned on four training TTS models transfers to a new TTS without any kernel re-fitting. This is a load-bearing assumption, and it is tested on only one unseen TTS, 11Labs, which was chosen for its difficulty. It remains unclear whether the method would work on a TTS whose embedding-space distribution is farther from the training set (e.g., a different language or a different generation paradigm). An additional OOD TTS or a synthetic distribution-shift experiment would substantially strengthen the claim.
  4. [§3.2, Algorithm 1] The training and evaluation protocol is underspecified. The number of training iterations T, the learning rate, the sampling distribution over the four TTS models, and the exact composition of the 1000 held-out support examples (class balance, TTS mix, how they were drawn) are not given. Algorithm 1 also references xeval and zeval without defining them in the input line. These details are necessary for reproducibility and for assessing whether the reported EERs are robust to support-set composition.
minor comments (5)
  1. [§4.3] The text refers to 'LibreFake', but the dataset is named 'LibriFake'; please fix the typo.
  2. [§4.4] The abbreviation 'SSL-ASSIST' is inconsistent with 'SSL-AASIST' used elsewhere; please unify the spelling.
  3. [Table 1] The last column header '100-Shots ID Params' is ambiguous; it appears to report the number of trainable parameters after 100-shot adaptation, but the header should be split (e.g., '100-Shots ID' and 'Params') for clarity.
  4. [Algorithm 1] The pseudocode uses T in the loop condition but does not specify its value or how it is chosen; please define it in the algorithm description.
  5. [Figure 1] The caption describes calibration curves but does not specify the number of bins or the binning strategy; adding this would make the figure reproducible.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: ADD-GP's few-shot result is an empirical posterior evaluation with a fixed kernel; the 11Labs OOD choice is post-hoc but not circular.

full rationale

The paper's derivation chain is not circular. Section 3.2 and Algorithm 1 learn the RBF deep kernel (Eq. 4) by maximizing the log-marginal likelihood of a Dirichlet-based GP classifier on XLS-R embeddings from the four in-distribution TTS models; the hyperparameters and last XLS-R block are trained on those labels, not on 11Labs. Section 3.3 keeps the learned kernel fixed and performs few-shot adaptation by augmenting the 1000 held-out support examples with new-TTS samples and recomputing the closed-form GP predictive posterior (Eqs. 1-3). The 0.54% EER in Table 1 is therefore an empirical test-set measurement, not an optimized or fitted quantity and not a renamed input. MixPro (Sec. 3.4, Eq. 5) is an external augmentation rule and does not make the target result part of the training objective. The self-citations ([11], [14]) are background or a post-hoc note ('we note that this aligns with previous works...') and are not load-bearing; the adaptation mechanism itself relies on standard GP/DKL theory and external backbones. The main caveat is evaluative rather than circular: Section 3.1 states that 11Labs was chosen as the unseen TTS because 'detection performance was significantly worse on 11Labs,' i.e., the OOD test model was selected after observing its difficulty. That post-hoc selection limits how strongly the single-model result supports 'unseen TTS' generalization, but it does not reduce the result to its inputs. Also, the claimed margin over 'the best baseline' should cite RWM MixPro (1.24%) rather than SSL-AASIST-FT MixPro (5.76%) in Table 1; the method still leads, but with a smaller margin.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on standard GP theory, the transferability of a learned kernel to a new TTS, and the representativeness of the generated dataset. The kernel hyperparameters and the size of the support set are free parameters. No physically invented entities are introduced.

free parameters (4)
  • kernel length scale ℓ = not reported
    Learned during training via log-marginal likelihood optimization (Alg. 1, line 7). Value not given in the paper.
  • kernel output scale σ = not reported
    Learned during training alongside ℓ. Value not given in the paper.
  • XLS-R last block parameters = not reported
    Fine-tuned during the training phase; all other XLS-R layers are frozen. The number of parameters is context-dependent.
  • number of held-out support examples (1000) = 1000
    Chosen by the authors as the fixed support set for GP evaluation; this is a design choice that influences results.
assumptions (3)
  • standard math Standard Gaussian Process formulation and Dirichlet-based GP classification provide a valid approximate posterior with closed-form updates.
    The paper relies on the GP posterior formulas (Eqs. 2-3) and the Dirichlet-based GP classifier of Milios et al. [16] without re-deriving them. This is a standard background assumption.
  • domain assumption XLS-R embeddings provide a feature space in which real and fake speech are separable, and where the RBF kernel similarity transfers to unseen TTS models.
    The entire method depends on the learned kernel being discriminative for a new TTS. The paper tests this only for 11Labs, leaving a gap in generality.
  • domain assumption The LibriFake dataset constructed from LibriSpeech and five selected TTS models is representative of realistic deepfake audio for evaluating adaptation.
    The paper does not validate that the generated fake audio covers a broad enough distribution, nor does it analyze potential dataset biases (e.g., TTS-specific artifacts).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Few-Shot Speech Deepfake Detection Adaptation with Gaussian Processes." pith.science (2026). https://pith.science/paper/Z3OYMAW3

@misc{pith2026250523619,
  author       = {Pith},
  title        = {Pith review of: Few-Shot Speech Deepfake Detection Adaptation with Gaussian Processes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z3OYMAW3}},
  note         = {Machine review of arXiv:2505.23619}
}
read the original abstract

Recent advancements in Text-to-Speech (TTS) models, particularly in voice cloning, have intensified the demand for adaptable and efficient deepfake detection methods. As TTS systems continue to evolve, detection models must be able to efficiently adapt to previously unseen generation models with minimal data. This paper introduces ADD-GP, a few-shot adaptive framework based on a Gaussian Process (GP) classifier for Audio Deepfake Detection (ADD). We show how the combination of a powerful deep embedding model with the Gaussian processes flexibility can achieve strong performance and adaptability. Additionally, we show this approach can also be used for personalized detection, with greater robustness to new TTS models and one-shot adaptability. To support our evaluation, a benchmark dataset is constructed for this task using new state-of-the-art voice cloning models.

Figures

Figures reproduced from arXiv: 2505.23619 by the authors.

Figure 1
Figure 1. Calibration curves for different models. On the left is the calibration on the 10-shot experiment. On the right, calibration for personalized 5-shot VoxCeleb experiment. ADD-GP, show a decrease in performance on the original TTSs, even when using methods designed to mitigate “catastrophic forgetting”. We note that this aligns with previous works that showed good continual learning results with GP models [14]. 4.3. P… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 25 canonical work pages

  1. [1]

    Recent models also excel in voice cloning, enabling the generation of speech that closely mimics a specific speaker’s voice with just a single reference ut- terance [3, 4]

    Introduction The rapid advancements in text-to-speech (TTS) generation have significantly improved the ability to synthesize highly real- istic speech from textual input [1, 2]. Recent models also excel in voice cloning, enabling the generation of speech that closely mimics a specific speaker’s voice with just a single reference ut- terance [3, 4]. This b...

    work page

  2. [2]

    Few-Shot Speech Deepfake Detection Adaptation with Gaussian Processes

    Background We provide a brief introduction to Gaussian Process (GP) clas- sifiers, which form the foundation of our proposed method. Scalars are denoted with lowercase letters (e.g.,x), vectors with boldface lowercase letters (e.g., x), and matrices with boldface uppercase letters (e.g., X). In general, y = [y1, ..., yN ]T repre- sents the vector of label...

    work page Pith review arXiv 2025

  3. [3]

    Methods We introduce our framework that uses a Dirichlet-based GP classifier [16] as a back-end, on features extracted from XLS-R [7], a Wav2Vec2 based self-supervised model, as the front-end. 3.1. Datasets Generation To evaluate our method, we introduce LibriFake, a large-scale deepfake benchmark derived from LibriSpeech [28]. For each sample in LibriSpe...

    work page

  4. [4]

    catastrophic forgetting

    Experiments 4.1. Baselines We evaluated our approach against SSL-AASIST [34], a widely used baseline known for its strong performance on publicly available datasets. SSL-AASIST employs the AASIST [21] ar- chitecture as the back-end while using XLS-R as the front-end, making it a relevant benchmark for comparison with our XLS- R-based framework. To adapt S...

    work page

  5. [5]

    As demonstrated by our exper- iments, this can be a significant security concern, as ADD per- formance on seen TTS is almost perfect while degrading con- siderably on unseen TTSs

    Conclusion In this work, we address the issue of adapting audio deepfake detection to new unseen TTSs. As demonstrated by our exper- iments, this can be a significant security concern, as ADD per- formance on seen TTS is almost perfect while degrading con- siderably on unseen TTSs. We tackle this issue by introduc- ing a novel ADD approach based on Gaussi...

    work page

  6. [6]

    Lora: Low-rank adaptation of large language models,

    E. J. Hu, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen et al. , “Lora: Low-rank adaptation of large language models,” in International Conference on Learning Representa- tions (ICLR), 2022

    work page 2022

  7. [7]

    Neural codec language mod- els are zero-shot text to speech synthesizers,

    C. Wang, S. Chen, Y . Wu, Z. Zhang, L. Zhou, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Li et al. , “Neural codec language mod- els are zero-shot text to speech synthesizers,” arXiv preprint arXiv:2301.02111, 2023

    arXiv 2023

  8. [8]

    V oicebox: Text-guided multilingual universal speech generation at scale,

    M. Le, A. Vyas, B. Shi, B. Karrer, L. Sari, R. Moritz, M. Williamson, V . Manohar, Y . Adi, J. Mahadeokar et al. , “V oicebox: Text-guided multilingual universal speech generation at scale,” Advances in neural information processing systems , vol. 36, 2024

    work page 2024

Show all 42 references
  1. [9]

    Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,

    E. Casanova, J. Weber, C. D. Shulby, A. C. Junior, E. G ¨olge, and M. A. Ponti, “Yourtts: Towards zero-shot multi-speaker tts and zero-shot voice conversion for everyone,” in International Con- ference on Machine Learning. PMLR, 2022, pp. 2709–2720

    2022

  2. [10]

    F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,

    Y . Chen, Z. Niu, Z. Ma, K. Deng, C. Wang, J. Zhao, K. Yu, and X. Chen, “F5-tts: A fairytaler that fakes fluent and faithful speech with flow matching,” arXiv preprint arXiv:2410.06885, 2024

    2024 arXiv

  3. [11]

    Audio deepfake approaches,

    O. A. Shaaban, R. Yildirim, and A. A. Alguttar, “Audio deepfake approaches,” IEEE Access, vol. 11, pp. 132 652–132 682, 2023

    2023

  4. [12]

    Bayesian inference for logistic models using P ´olya–Gamma latent variables,

    N. G. Polson, J. G. Scott, and J. Windle, “Bayesian inference for logistic models using P ´olya–Gamma latent variables,” Journal of the American Statistical Association, pp. 1339–1349, 2013

    2013

  5. [13]

    Xls-r: Self- supervised cross-lingual speech representation learning at scale,

    A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. V on Platen, Y . Saraf, J. Pino et al. , “Xls-r: Self- supervised cross-lingual speech representation learning at scale,” arXiv preprint arXiv:2111.09296, 2021

    2021 arXiv

  6. [14]

    C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for Machine Learning. The MIT Press, 2006

    2006

  7. [15]

    Scalable vari- ational gaussian process classification,

    J. Hensman, A. Matthews, and Z. Ghahramani, “Scalable vari- ational gaussian process classification,” in Artificial Intelligence and Statistics. PMLR, 2015, pp. 351–360

    2015

  8. [16]

    Stochas- tic variational deep kernel learning,

    A. G. Wilson, Z. Hu, R. Salakhutdinov, and E. P. Xing, “Stochas- tic variational deep kernel learning,” in Proceedings of the 30th International Conference on Neural Information Processing Sys- tems, 2016, pp. 2594–2602

    2016

  9. [17]

    Guided deep kernel learning,

    I. Achituve, G. Chechik, and E. Fetaya, “Guided deep kernel learning,” in Uncertainty in Artificial Intelligence. PMLR, 2023, pp. 11–21

    2023

  10. [18]

    Deep kernel learning,

    A. G. Wilson, Z. Hu, R. Salakhutdinov, and E. P. Xing, “Deep kernel learning,” in International Conference on Artificial Intelli- gence and Statistics (AISTATS), 2016

    2016

  11. [19]

    Multi-class Gaussian process classification made conjugate: Efficient infer- ence via data augmentation,

    T. Galy-Fajou, F. Wenzel, C. Donner, and M. Opper, “Multi-class Gaussian process classification made conjugate: Efficient infer- ence via data augmentation,” in Uncertainty in Artificial Intelli- gence. PMLR, 2020, pp. 755–765

    2020

  12. [20]

    GP-Tree: A Gaussian process classifier for few-shot incremen- tal learning,

    I. Achituve, A. Navon, Y . Yemini, G. Chechik, and E. Fetaya, “GP-Tree: A Gaussian process classifier for few-shot incremen- tal learning,” in International Conference on Machine Learning . PMLR, 2021, pp. 54–65

    2021

  13. [21]

    Personalized federated learning with Gaussian processes,

    I. Achituve, A. Shamsian, A. Navon, G. Chechik, and E. Fetaya, “Personalized federated learning with Gaussian processes,” Ad- vances in Neural Information Processing Systems , vol. 34, pp. 8392–8406, 2021

    2021

  14. [22]

    and RawNet3 [23], learn directly from raw waveforms, ex- tracting features and classifying them within the same model without a separate feature extraction step. Wav2Vec2 [24], HuBERT [25], and XLS-R [7] have be- come standard backbone models in modern ADD systems due to their...

  15. [23]

    Dirichlet-based Gaussian processes for large-scale cali- brated classification,

    D. Milios, R. Camoriano, P. Michiardi, L. Rosasco, and M. Fil- ippone, “Dirichlet-based Gaussian processes for large-scale cali- brated classification,”Advances in Neural Information Processing Systems, vol. 31, 2018

    2018

  16. [24]

    Manifold Gaussian processes for regression,

    R. Calandra, J. Peters, C. E. Rasmussen, and M. P. Deisenroth, “Manifold Gaussian processes for regression,” in 2016 Interna- tional Joint Conference on Neural Networks (IJCNN) . IEEE, 2016, pp. 3338–3345

    2016

  17. [25]

    Audio deepfake detection: A survey,

    J. Yi, C. Wang, J. Tao, X. Zhang, C. Y . Zhang, and Y . Zhao, “Audio deepfake detection: A survey,” arXiv preprint arXiv:2308.14970, 2023

    2023 arXiv

  18. [26]

    Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,

    J.-w. Jung, H.-S. Heo, H. Tak, H.-j. Shim, J. S. Chung, B.-J. Lee, H.-J. Yu, and N. Evans, “Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” in arXiv preprint arXiv:2110.01200, 2021

    2021 arXiv

  19. [27]

    Aasist: Audio anti-spoofing using integrated spectro- temporal graph attention networks,

    ——, “Aasist: Audio anti-spoofing using integrated spectro- temporal graph attention networks,” inIEEE international confer- ence on acoustics, speech and signal processing (ICASSP), 2022, pp. 6367–6371

    2022

  20. [28]

    End-to-end spectro-temporal graph attention networks for speaker verification anti-spoofing and speech deepfake detec- tion,

    H. Tak, J. weon Jung, J. Patino, M. Kamble, M. Todisco, and N. Evans, “End-to-end spectro-temporal graph attention networks for speaker verification anti-spoofing and speech deepfake detec- tion,” in Proc. 2021 Edition of the Automatic Speaker Verification and Spoofing Counter...

    2021

  21. [29]

    Pushing the limits of raw waveform speaker recogni- tion,

    J.-w. Jung, Y . J. Kim, H.-S. Heo, B.-J. Lee, Y . Kwon, and J. S. Chung, “Pushing the limits of raw waveform speaker recogni- tion,” Proc. Interspeech, 2022

    2022

  22. [30]

    wav2vec: Unsupervised pre-training for speech recognition,

    S. Schneider, A. Baevski, R. Collobert, and M. Auli, “wav2vec: Unsupervised pre-training for speech recognition,”arXiv preprint arXiv:1904.05862, 2019

    1904 arXiv

  23. [31]

    Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdi- nov, and A. Mohamed, “Hubert: Self-supervised speech represen- tation learning by masked prediction of hidden units,”IEEE/ACM transactions on audio, speech, and language processing, vol. 29, pp. 3451–3460, 2021

    2021

  24. [32]

    Cross- domain audio deepfake detection: Dataset and analysis,

    Y . Li, M. Zhang, M. Ren, M. Ma, D. Wei, and H. Yang, “Cross- domain audio deepfake detection: Dataset and analysis,” arXiv preprint arXiv:2404.04904, 2024

    2024 arXiv

  25. [33]

    Sonar: A synthetic ai- audio detection framework and benchmark,

    X. Li, P.-Y . Chen, and W. Wei, “Sonar: A synthetic ai- audio detection framework and benchmark,” arXiv preprint arXiv:2410.04324, 2024

    2024 arXiv

  26. [34]

    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,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, pp. 5206–5210

    2015

  27. [35]

    Speak, read and prompt: High-fidelity text-to-speech with minimal supervision,

    E. Kharitonov, D. Vincent, Z. Borsos, R. Marinier, S. Girgin, O. Pietquin, M. Sharifi, M. Tagliasacchi, and N. Zeghidour, “Speak, read and prompt: High-fidelity text-to-speech with minimal supervision,” 2023. [Online]. Available: https://arxiv. org/abs/2302.03540

    2023 arXiv

  28. [36]

    Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,

    Z. Zhang, L. Zhou, C. Wang, S. Chen, Y . Wu, S. Liu, Z. Chen, Y . Liu, H. Wang, J. Liet al., “Speak foreign languages with your own voice: Cross-lingual neural codec language modeling,”arXiv preprint arXiv:2303.03926, 2023

    2023 arXiv

  29. [37]

    V oxceleb: a large-scale speaker identification dataset,

    A. Nagrani, J. S. Chung, and A. Zisserman, “V oxceleb: a large-scale speaker identification dataset,” arXiv preprint arXiv:1706.08612, 2017

    2017 arXiv

  30. [38]

    Few-shot adaption to distribution shifts by mixing source and target embed- dings,

    Y . Xue, A. Payani, Y . Yang, and B. Mirzasoleiman, “Few-shot adaption to distribution shifts by mixing source and target embed- dings,” International Conference on Machine Learning (ICML) , 2024

    2024

  31. [39]

    mixup: Beyond empirical risk minimization,

    H. Zhang, “mixup: Beyond empirical risk minimization,” arXiv preprint arXiv:1710.09412, 2017

    2017 arXiv

  32. [40]

    Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,

    H. Tak, M. Todisco, X. Wang, J.-w. Jung, J. Yamagishi, and N. Evans, “Automatic speaker verification spoofing and deep- fake detection using wav2vec 2.0 and data augmentation,” arXiv preprint arXiv:2202.12233, 2022

    2022 arXiv

  33. [41]

    Continual learning of context-dependent processing in neural networks,

    G. Zeng, Y . Chen, B. Cui, and S. Yu, “Continual learning of context-dependent processing in neural networks,” Nature Ma- chine Intelligence, vol. 1, no. 8, pp. 364–372, 2019

    2019

  34. [42]

    What to remember: Self-adaptive continual learning for audio deepfake detection,

    X. Zhang, J. Yi, C. Wang, C. Y . Zhang, S. Zeng, and J. Tao, “What to remember: Self-adaptive continual learning for audio deepfake detection,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 17, 2024, pp. 19 569–19 577

    2024

Pith tools

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