Pith. sign in

REVIEW 3 major objections 6 minor 41 references

Fixed non-semantic embeddings from TRILLsson detect fake speech as accurately as state-of-the-art systems in-domain and far better on unseen datasets.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Frozen non-semantic TRILLson embeddings with a lightweight backend beat prior spoofing detectors on out-of-domain datasets while staying competitive in-domain.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Worth a careful referee: the DF21 and ablation results are real, but the ItW comparison is not reproducible as written. the 3 major comments →

arxiv 2509.00186 v1 pith:UMMVLVIO submitted 2025-08-29 cs.SD cs.AIeess.AS

Generalizable Audio Spoofing Detection using Non-Semantic Representations

Arnab Das , Yassine El Kheir , Carlos Franzreb , Tim Herzig , Tim Polzehl , Sebastian M\"oller This is my paper
classification cs.SD cs.AIeess.AS
keywords audio spoofing detectiondeepfake detectionnon-semantic audio representationTRILLssonself-supervised embeddingscross-dataset generalizationanti-spoofingASVspoof
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.

The reading

This paper tries to show that audio spoofing detection can generalize to unseen fake-speech generators, codecs, and real-world noise if the detector is built on non-semantic speech representations instead of semantic embeddings, hand-crafted features, or raw waveforms. The authors extract frozen frame-level embeddings from TRILL and TRILLsson, feed them through a simple backend of convolutions, LSTM layers, and attention pooling, and train only the backend. On the in-domain ASVspoof 2019 LA test set the best configuration matches state-of-the-art performance, while on out-of-domain sets it clearly beats prior systems. The point matters because current deepfake detectors drop sharply when tested on new data, and the paper offers a representation choice that appears to fix that drop without fine-tuning.

Core claim

The paper claims that non-semantic universal audio representations, specifically TRILLsson embeddings extracted from 200 ms chunks, carry enough generator-specific artifacts to separate bonafide from fake speech while discarding the semantic content that causes other models to overfit. With a fixed extractor and a small trainable backend, the best model achieves 0.96% EER on LA19, comparable to AASIST's 0.83%, and outperforms all compared state-of-the-art systems on unseen data: 13.27% EER on DF21 versus 20.26% for SE-Rawformer, and 20.08% EER on In-the-Wild versus 24.27% for the best prior fusion of semantic embeddings. An ablation replacing non-semantic features with XLS-R embeddings in th

What carries the argument

The central object is the non-semantic frame embedding stack from TRILL and TRILLsson, chunked at 200 ms to capture local artifacts without losing global temporal patterns. The chunked embeddings are stacked into a 2D representation, passed through residual 1D convolutions, two LSTM layers, multi-head attention pooling over time, and an MLP head that outputs bonafide/fake logits. The fixed, frozen embeddings do the representation work; the backend only learns to aggregate them.

Load-bearing premise

That TRILLsson embeddings, frozen from paralinguistic training, retain enough generator-specific artifacts to separate fake from real speech while stripping speaker, channel, and semantic content that cause other models to overfit.

What would settle it

Take the same backend and compare TRILLsson features against semantic features on a new unseen dataset with different speakers, codecs, and generators from any training set; if semantic features match or beat TRILLsson there, the claimed non-semantic advantage fails. Alternatively, measure how much speaker identity is recoverable from TRILLsson embeddings: if a downstream classifier can identify speakers from them with high accuracy, the 'non-semantic discard' premise is false.

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

If this is right

  • If this holds, spoofing countermeasures can be built from frozen public embeddings plus a lightweight backend, avoiding expensive fine-tuning of large self-supervised models.
  • Non-semantic features would transfer across codecs, telephony channels, and unseen voice-conversion algorithms better than semantic or end-to-end models, which is the main practical failure of current detectors.
  • The 200 ms chunking result suggests an interpretable design principle: detector temporal resolution should match about one syllable of speech, not the 25 ms frame used by ASR-oriented extractors.
  • Combining non-semantic frontends with stronger backends, such as graph neural networks, could push out-of-domain performance further.
  • The ablation indicates that semantic embeddings like XLS-R actively hurt generalization even with an identical detector, so feature choice matters more than architecture for cross-dataset robustness.

Where Pith is reading between the lines

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

  • A testable extension would be to see whether the non-semantic advantage grows as the generator becomes more realistic: if synthetic speech still leaves paralinguistic rhythm and prosody artifacts, TRILLsson-like features should keep separating them even when phoneme-level content is nearly perfect.
  • The paper's syllable-duration explanation suggests an experiment the authors did not run: sweep chunk durations continuously from 20 ms to 500 ms and check whether the EER curve bottoms out near 150-250 ms across datasets, which would directly test the claimed mechanism.
  • Because TRILLsson is trained for paralinguistic tasks, it may encode speaker identity as well; if that is true, the generalization claim would partly rest on speaker-channel robustness, and a detector that explicitly removes speaker information might generalize even further.
  • The fixed-extractor design means the method can be updated cheaply as new synthetic generators appear, so it could serve as a continuously deployed baseline that only the backend needs retraining.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a spoofing countermeasure that uses frozen non-semantic audio embeddings (TRILL and TRILLsson variants) extracted from fixed-duration chunks, followed by a compact backend composed of a residual convolutional block, optional frame-wise delta, two LSTM layers, multi-head attention pooling, and an MLP classifier. Models are trained only on the ASVspoof 2019 LA training set and evaluated on the LA19, LA21, DF21, and In-the-Wild datasets. The authors report in-domain LA19 EER comparable to state-of-the-art (0.96% for MT3 vs. 0.83% for AASIST), improved out-of-domain EER on DF21 (13.27% vs. 20.26% for SE-Rawformer), and improved EER on In-the-Wild (20.08% for MT1 vs. 24.27% for the best prior fusion system). An ablation replacing TRILLsson features with XLS-R features under the same backend is used to attribute the generalization gain to non-semantic representations. Public code and model links are provided.

Significance. If the reported results are reproducible and the evaluation protocols match those of the baselines, the paper makes a useful empirical contribution: it shows that simple, frozen, non-semantic frontends may generalize better to unseen spoofing algorithms and real-world conditions than end-to-end systems, semantic SSL embeddings, and hand-crafted features. A particular strength is the controlled ablation in Table 5, which holds the detector backend fixed and directly compares semantic vs. non-semantic frontends, and the release of code and use of publicly available pre-trained models. However, the central claim depends on protocol details that are currently underspecified, especially for the In-the-Wild evaluation, and on the model-selection procedure for LA19.

major comments (3)
  1. [§3.1, Table 4] The headline out-of-domain result on In-the-Wild is not verifiable as written. The manuscript states only the total size of the ItW dataset (17.2h fake, 20.7h bonafide, 31,779 utterances), but does not specify which subset was used for evaluation, how utterances were selected, whether the split follows the original train/test partition of Müller et al. [2] or some other protocol, how many utterances per class were scored, or how scores were fused across chunks/utterances. Since ItW is not a standardized benchmark with a single protocol, the comparison of 20.08% EER (MT1) against 24.27% (XLS-R/WavLM/Hubert fusion) is only meaningful if the same evaluation set and scoring method are used. Without this information, the central claim of superior generalization on public-domain data cannot be assessed or reproduced.
  2. [§3.2, §4, Tables 1 and 3] The model-selection procedure is ambiguous and potentially circular. Section 3.2 states that the best models are chosen based on performance on the LA19 dev set, but Section 4 and Table 1 use the LA19 evaluation set to select the top two configurations (200ms chunk, Direct) and then report in-domain performance on that same evaluation set. If the configuration (model variant, chunk size, delta flag) was selected on the LA19 eval set, the reported in-domain EERs are optimistic and the comparison to SOTA is not an honest out-of-sample evaluation. In addition, Section 4 says for OOD tables the best score from three seeds is reported, without variance or confidence intervals. Please clarify whether selection used dev or eval, and report mean ± std over seeds for all reported EERs.
  3. [§4, Table 5] The ablation claims to isolate semantic vs. non-semantic features by using the same detector backend, but the temporal resolution and feature dimensionality differ between TRILLsson and XLS-R: TRILLsson embeddings are extracted from 200ms chunks, whereas XLS-R is described as using 25ms frames with 20ms stride. It is therefore unclear whether the large OOD drop for XLS-R (28.78% on LA21) is caused by semantic content or by the different temporal context available to the backend. To support the claim that non-semantic features are inherently better for generalization, the authors should either match the temporal context more carefully or vary both representation type and chunk length in a full factorial manner.
minor comments (6)
  1. [Throughout] Many typographical errors and inconsistent spacing: 'T able', 'W e', 'Y assine', 'archives' should be 'achieves', 'underperformant' should be 'underperform'. The paper would benefit from a careful proofread.
  2. [§3.1 ItW] The ItW description gives total hours and utterance counts, but not class-wise utterance counts or the duration distribution for the evaluation subset actually used. Please state the number of bonafide and fake test utterances and the scoring procedure (e.g., utterance-level or chunk-level aggregation).
  3. [§2, Figure 1] The figure shows 'TRILL or TRILLsson' and 'Detector Backend', but the optional delta and chunking are not visible in the schematic. A small annotation clarifying where chunking/delta occurs would improve readability.
  4. [§4] The phrase 'almost a five-times performance drop from LA21' refers to EER rising from 4.31% to 20.26% for SE-Rawformer; this is a relative increase of about 4.7x, so 'almost five times worse' is acceptable but should be phrased carefully to avoid implying an absolute percentage-point ratio.
  5. [§4, Table 4] The comparison to 'XLS-R, WavLM, Hubert & Fusion' from [31] should state explicitly whether the cited system was trained on LA19 only or used additional training data. If additional data were used, the comparison is not apples-to-apples and should be discussed.
  6. [§4, Discussion] The claim that TRILLsson embeddings are 'non-semantic' is inherited from the model's training objective, but the paper does not verify that the embeddings lack speaker or channel information. A short analysis or a caveat would strengthen the argument, especially since speaker/channel leakage could also explain OOD generalization.

Circularity Check

0 steps flagged

No significant circularity: the generalization claims rest on frozen external embeddings and held-out test sets, not on fitted quantities or self-citations.

full rationale

The paper's central chain is: (1) use TRILL/TRILLsson as frozen feature extractors (public weights, from non-overlapping authors), (2) train a fixed backend only on the ASVspoof 2019 LA training split, and (3) evaluate on LA21, DF21, and ItW without training or fine-tuning on them. These out-of-domain evaluations are genuinely external: no parameter in the detector is fitted to the target EER values, and the ablation in Table 5 holds the backend fixed while swapping non-semantic TRILLsson features for semantic XLS-R features, directly testing the paper's hypothesis. There is no load-bearing self-citation chain; citations [23,24] establish the properties of TRILL/TRILLsson, [15] supplies class weights, and [28] independently documents XLS-R's generalization decline. Two caveats should be flagged but are not circularity: the ItW evaluation split is not specified in Section 3.1/Table 4, making the headline comparison to prior ItW results hard to verify, and the text is inconsistent about whether the best configuration was chosen on the LA19 dev set (Section 3.2) or the LA19 evaluation set (Section 4, 'Based on the results in Table 1'). If the latter, the reported in-domain LA19 numbers are optimistically selected, but this is benchmark hygiene/selection bias rather than a case where a result equals its input by construction. Neither caveat affects the core out-of-domain comparisons, which use datasets not touched during model selection. Hence no circular dependency is exhibited.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method's central claim relies entirely on pretrained TRILLson embeddings and the choice of configuration; no first-principles derivation, no new mathematical objects, and no newly postulated physical entities are introduced.

free parameters (4)
  • chunk window size = 200ms
    Selected by scanning 50, 100, 200, 300ms and taking the best EER on LA19 evaluation (Table 1).
  • model variant = MT1, MT3 (TRILLsson1, TRILLsson3)
    Chosen based on Table 1 results on LA19 evaluation; the top two models at 200ms chunk size.
  • delta flag = Direct (no delta)
    Selected by comparing Direct vs Delta variants on LA19 evaluation in Table 1.
  • seed reporting = best of 3 seeds
    Section 3.2: 'the best score is reported from three different seed runs'; this inflates reported EER and is not the best practice for variance reporting.
axioms (4)
  • domain assumption TRILL/TRILLsson pretrained embeddings are suitable frozen feature extractors for spoofing detection.
    Section 2: 'pre-trained model weights are publicly available' and used as fixed feature extractors; the entire method depends on these embeddings carrying artifact information.
  • ad hoc to paper The LA19 development set is representative for selecting the best model configuration.
    Section 3.2 says 'Best models are chosen based on performance on LA19 dev set', yet the configuration is chosen by inspecting LA19 evaluation results in Table 1, which introduces selection-on-test-set bias.
  • domain assumption LA21, DF21, and ItW are valid out-of-domain test sets.
    Section 3.1 classifies them as external evaluations; however LA21 shares the attack algorithms with LA19 and only adds channel/codec variation, so calling it out-of-domain is only partially true.
  • domain assumption Cited baseline results were produced under the same training and evaluation protocol (LA19 train-only, same ItW split).
    Tables 3 and 4 compare against numbers from other papers without documenting the exact ItW split and scoring, so the comparison may not be apples-to-apples.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalizable Audio Spoofing Detection using Non-Semantic Representations." pith.science (2026). https://pith.science/paper/UMMVLVIO

@misc{pith2026250900186,
  author       = {Pith},
  title        = {Pith review of: Generalizable Audio Spoofing Detection using Non-Semantic Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UMMVLVIO}},
  note         = {Machine review of arXiv:2509.00186}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Rapid advancements in generative modeling have made synthetic audio generation easy, making speech-based services vulnerable to spoofing attacks. Consequently, there is a dire need for robust countermeasures more than ever. Existing solutions for deepfake detection are often criticized for lacking generalizability and fail drastically when applied to real-world data. This study proposes a novel method for generalizable spoofing detection leveraging non-semantic universal audio representations. Extensive experiments have been performed to find suitable non-semantic features using TRILL and TRILLsson models. The results indicate that the proposed method achieves comparable performance on the in-domain test set while significantly outperforming state-of-the-art approaches on out-of-domain test sets. Notably, it demonstrates superior generalization on public-domain data, surpassing methods based on hand-crafted features, semantic embeddings, and end-to-end architectures.

Figures

Figures reproduced from arXiv: 2509.00186 by Arnab Das, Carlos Franzreb, Sebastian M\"oller, Tim Herzig, Tim Polzehl, Yassine El Kheir.

Figure 1
Figure 1. Figure 1: Schematic architecture diagram of the proposed framework. cross-dataset evaluations. Experimental results demonstrate perfor￾mance comparable to state-of-the-art (SOTA) models for in-domain evaluation. Moreover, the proposed method leveraging these representations exhibits improved generalization to out-of-domain test datasets. Notably, it achieves superior performance on noisy real-world public-domain dat… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 31 canonical work pages · 3 internal anchors

  1. [1]

    In the Wild

    Introduction In the past decade, significant strides in generative speech research, particularly in the context of text-to-speech (TTS) and voice conversion (VC) systems, have enabled the generation of synthetic speech that is more natural sounding and of higher quality. These developments have effectively narrowed the distinction between authentic (real)...

    work page

  2. [2]

    Initially, the input audio waveform is chunked into frames and each chunk is processed through TRILL or TRILLs- son models to extract audio representations

    Methodology A high-level architecture diagram of the proposed framework is il- lustrated in Figure 1. Initially, the input audio waveform is chunked into frames and each chunk is processed through TRILL or TRILLs- son models to extract audio representations. The resulting frame- wise representations are then stacked to form a 2D representation X ∈Rd×t, wh...

    work page

  3. [3]

    Datasets W e conducted extensive experiments using four distinct English datasets

    Experimental setup 3.1. Datasets W e conducted extensive experiments using four distinct English datasets. ASVspoof 2019 Logical Access (LA19):ASVspoof 2019 challenge made this dataset publicly available [1], comprising12,483 genuine and 108,978 spoofed utterances sampled at16 kHz. The dataset features male and female speakers sourced from the VCTK corpus...

    work page 2019

  4. [4]

    The results are presented in T able 1

    Results & discussion T o determine the most suitable representation for spoofing detection among TRILL and TRILLsson models, we evaluate our models on the LA19 evaluation set. The results are presented in T able 1. For each model, two rows are presented based on whether the frame-wise delta is performed or not as mentioned in Section 2. W e define that in...

    work page

  5. [5]

    W e perform extensive experiments to find the most suitable TRILLsson model and optimal chunking duration to balance the local and global temporal features

    Conclusion & outlook In this paper, we propose a spoofing detection method based on non- semantic audio representation extracted from TRILL and various TRILLsson models. W e perform extensive experiments to find the most suitable TRILLsson model and optimal chunking duration to balance the local and global temporal features. Results show that the proposed...

    work page

  6. [6]

    Acknowledgements This research has been partly funded by the Federal Ministry of Education and Research, Germany (BMBF 03RU2U151C, project news-polygraph) and partly by the V olkswagen Foundation

    work page

  7. [7]

    Relative phase information for detecting human speech and spoofed speech

    L. W ang, Y . Y oshida, Y . Kawakami, and S. Nakagawa, “Relative phase information for detecting human speech and spoofed speech.” in INTERSPEECH, 2015, pp. 2092–2096

    work page 2015

  8. [8]

    Generalizable Audio Spoofing Detection using Non-Semantic Representations

    and mel-frequency cepstral coefficient (MFCC) [ 9]. Early DNN-based methods by [10] introduced ResNet and Inception- based binary classifiers, outperforming prior speech representation frameworks. The authors suggest that shallow networks are better suited for detecting artifacts from spoofing algorithms, rather than focusing on high-level, respectively d...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech,

    X. W ang, J. Y amagishi, M. T odisco, H. Delgado, A. Nautsch, N. Evans, M. Sahidullah, V . V estman, T . Kinnunen, K. A. Leeet al., “ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech,”Computer Speech & Language, vol. 64, p. 101114, 2020

    work page 2019

  10. [10]

    Does audio deepfake detection generalize?

    N. M. M ¨uller, P . Czempin, F . Dieckmann, A. Froghyar, and K. B ¨ottinger, “Does audio deepfake detection generalize?” arXiv preprint arXiv:2203.16263, 2022

    arXiv 2022

  11. [11]

    Resnet and model fusion for automatic spoofing detection

    Z. Chen, Z. Xie, W . Zhang, and X. Xu, “Resnet and model fusion for automatic spoofing detection. ” inInterspeech, 2017, pp. 102–106

    work page 2017

  12. [12]

    Replay and synthetic speech detection with res2net architecture,

    X. Li, N. Li, C. W eng, X. Liu, D. Su, D. Y u, and H. Meng, “Replay and synthetic speech detection with res2net architecture,” inICASSP 2021-2021 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2021, pp. 6354–6358

    work page 2021

  13. [13]

    Fake speech detection using residual network with transformer encoder,

    Z. Zhang, X. Yi, and X. Zhao, “Fake speech detection using residual network with transformer encoder, ” inProceedings of the 2021 ACM workshop on information hiding and multimedia security, 2021, pp. 13–22

    work page 2021

  14. [14]

    Audio deepfake detection: A survey,

    J. Yi, C. W ang, J. T ao, X. Zhang, C. Y . Zhang, and Y . Zhao, “ Audio deepfake detection: A survey, ”arXiv preprint arXiv:2308.14970, 2023

    Pith/arXiv arXiv 2023

  15. [15]

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

    J.-w. Jung, H.-S. Heo, H. T ak, H.-j. Shim, J. S. Chung, B.-J. Lee, H.-J. Y u, and N. Evans, “ AASIST: Audio anti-spoofing using integrated spectro-temporal graph attention networks,” inICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2022, pp. 6367–6371

    work page 2022

  16. [16]

    Ur channel-robust synthetic speech detection system for asvspoof 2021,

    X. Chen, Y . Zhang, G. Zhu, and Z. Duan, “Ur channel-robust synthetic speech detection system for asvspoof 2021,”arXiv preprint arXiv:2107.12018, 2021

    work page arXiv 2021

  17. [17]

    A comparison of features for synthetic speech detection,

    M. Sahidullah, T . Kinnunen, and C. Hanil c ¸i, “A comparison of features for synthetic speech detection,” inProc. Interspeech 2015, 2015, pp. 2087–2091

    work page 2015

  18. [18]

    T owards end-to-end synthetic speech detection,

    G. Hua, A. B. J. Teoh, and H. Zhang, “T owards end-to-end synthetic speech detection,” IEEE Signal Processing Letters , vol. 28, pp. 1265–1269, 2021

    work page 2021

  19. [19]

    Multi-task learning improves synthetic speech detection,

    Y . Mo and S. W ang, “Multi-task learning improves synthetic speech detection,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6392–6396

    work page 2022

  20. [20]

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

    H. T ak, J.-w. Jung, J. Patino, M. Kamble, M. T odisco, and N. Evans, “End-to-end spectro-temporal graph attention networks for speaker verification anti-spoofing and speech deepfake detection,” arXiv preprint arXiv:2107.12710, 2021

    Pith/arXiv arXiv 2021

  21. [21]

    Speaker recognition from raw waveform with sincnet,

    M. Ravanelli and Y . Bengio, “Speaker recognition from raw waveform with sincnet,” in2018 IEEE spoken language technology workshop (SLT). IEEE, 2018, pp. 1021–1028

    2018

  22. [22]

    Advanced rawnet2 with attention- based channel masking for synthetic speech detection,

    J. Li, Y . Long, Y . Li, and D. Xu, “ Advanced rawnet2 with attention- based channel masking for synthetic speech detection,” in Proc. INTERSPEECH, vol. 2023, 2023, pp. 2788–2792

    work page 2023

  23. [23]

    T owards learning a universal non-semantic representation of speech,

    J. Shor, A. Jansen, R. Maor, O. Lang, O. Tuval, F . d. C. Quitry, M. T agliasacchi, I. Shavitt, D. Emanuel, and Y . Haviv, “T owards learning a universal non-semantic representation of speech,”arXiv preprint arXiv:2002.12764, 2020

    Pith/arXiv arXiv 2002

  24. [24]

    A robust audio deepfake detection system via multi-view feature,

    Y . Y ang, H. Qin, H. Zhou, C. W ang, T . Guo, K. Han, and Y . W ang, “ A robust audio deepfake detection system via multi-view feature,” in ICASSP 2024-2024 IEEE International Conference on Acous- tics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 13 131–13 135

    work page 2024

  25. [25]

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

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

    Pith/arXiv arXiv 2021

  26. [26]

    W avLM: Large-scale self-supervised pre-training for full stack speech processing,

    S. Chen, C. W ang, Z. Chen, Y . Wu, S. Liu, Z. Chen, J. Li, N. Kanda, T . Y oshioka, X. Xiaoet al., “W avLM: Large-scale self-supervised pre-training for full stack speech processing, ”IEEE Journal of Selected T opics in Signal Processing, vol. 16, no. 6, pp. 1505–1518, 2022

    work page 2022

  27. [27]

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

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

    work page 2021

  28. [28]

    Exploring generalization to unseen audio data for spoof- ing: Insights from ssl models,

    A. Kulkarni, H. M. Tran, A. Kulkarni, S. Dowerah, D. Lolive, and M. M. Doss, “Exploring generalization to unseen audio data for spoof- ing: Insights from ssl models, ” inASVSpoof workshop 2024, 2024

    work page 2024

  29. [29]

    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, ”arXiv preprint arXiv:2005.07143, 2020

    Pith/arXiv arXiv 2005

  30. [30]

    MLAAD: The multi-language audio anti-spoofing dataset,

    N. M. M ¨uller, P . Kawa, W . H. Choong, E. Casanova, E. G ¨olge, T . M¨uller, P . Syga, P . Sperl, and K. B ¨ottinger, “MLAAD: The multi-language audio anti-spoofing dataset,” arXiv preprint arXiv:2401.09512, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Audio deepfake detection with self-supervised xls-r and sls classifier,

    Q. Zhang, S. W en, and T . Hu, “ Audio deepfake detection with self-supervised xls-r and sls classifier,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 6765–6773

    work page 2024

  32. [32]

    This gives us insight that features extracted over a duration of average syllable length are more beneficial for spoofing detection tasks

    reports that the maximum duration of phonemes in standard English speech is approximately 200ms, while [33] empirically finds the average syllable duration to be also around 200ms. This gives us insight that features extracted over a duration of average syllable length are more beneficial for spoofing detection tasks. Similarly, end-to-end models that dir...

    work page

  33. [33]

    TRILLsson: Distilled Universal Paralinguistic Speech Representations

    J. Shor and S. V enugopalan, “Trillsson: Distilled universal paralinguis- tic speech representations, ”arXiv preprint arXiv:2203.00236, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  34. [34]

    Self- normalizing neural networks,

    G. Klambauer, T . Unterthiner, A. Mayr, and S. Hochreiter, “Self- normalizing neural networks,” Advances in neural information processing systems, vol. 30, 2017

    work page 2017

  35. [35]

    CSTR VCTK Corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),

    J. Y amagishi, C. V eaux, K. MacDonaldet al., “CSTR VCTK Corpus: English multi-speaker corpus for CSTR voice cloning toolkit (version 0.92),”University of Edinburgh. The Centre for Speech T echnology Research (CSTR), 2019

    work page 2019

  36. [36]

    ASVspoof 2021: T owards spoofed and deepfake speech detection in the wild,

    X. Liu, X. W ang, M. Sahidullah, J. Patino, H. Delgado, T . Kinnunen, M. T odisco, J. Y amagishi, N. Evans, A. Nautschet al., “ ASVspoof 2021: T owards spoofed and deepfake speech detection in the wild,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023

    work page 2021

  37. [37]

    Investigating self-supervised front ends for speech spoofing countermeasures,

    X. W ang and J. Y amagishi, “Investigating self-supervised front ends for speech spoofing countermeasures, ”arXiv preprint arXiv:2111.07725, 2021

    Pith/arXiv arXiv 2021

  38. [38]

    Lever- aging positional-related local-global dependency for synthetic speech detection,

    X. Liu, M. Liu, L. W ang, K. A. Lee, H. Zhang, and J. Dang, “Lever- aging positional-related local-global dependency for synthetic speech detection,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    work page 2023

  39. [39]

    Df40: T oward next-generation deepfake detection,

    Z. Y an, T . Y ao, S. Chen, Y . Zhao, X. Fu, J. Zhu, D. Luo, C. W ang, S. Ding, Y . Wu et al., “Df40: T oward next-generation deepfake detection, ”arXiv preprint arXiv:2406.13495, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    SpeechFormer: A Hierarchical Efficient Framework Incorporating the Characteristics of Speech

    W . Chen, X. Xing, X. Xu, J. Pang, and L. Du, “Speechformer: A hierarchical efficient framework incorporating the characteristics of speech, ”arXiv preprint arXiv:2203.03812, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  41. [41]

    Syllable-level duration determination

    W . N. Campbell, “Syllable-level duration determination.” in EUROSPEECH, 1989, pp. 2698–2701

    work page 1989

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.