Pith. sign in

REVIEW 3 major objections 5 minor 40 references

A plain Conformer detector, trained with balanced multi-source data and heavy audio augmentation, leads a 14-set speech deepfake arena at 1.454% pooled EER using only open-source data.

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 ↗

T0 review · grok-4.5

2026-07-31 10:26 UTC pith:CCTOBTFT

load-bearing objection Solid systems result: training recipe (especially diverse augmentation) beats architecture chasing on Speech-DF-Arena, with one real caveat on train/eval overlap. the 3 major comments →

arxiv 2607.28351 v1 pith:CCTOBTFT submitted 2026-07-30 cs.SD cs.AI

Teffic-Audio: Tell Fact from Fiction

Wan Lin , Li Wang , Jindong Wang , Kunyu Feng , Zhizheng Wu This is my paper
classification cs.SD cs.AI
keywords Speech Deepfake DetectionSpeech-DF-ArenaGeneralizationConformerMulti-head attentive statistics poolingAudio augmentationAttack-balanced samplingOpen-source training
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.

Speech deepfake detection is no longer about catching one kind of synthesizer. Spoofed audio now comes from synthesis, voice conversion, vocoder rebuilds, and neural codecs, then gets reshaped by rooms, devices, and transmission. This report argues that strong generalization under that mess comes less from exotic detector designs and more from how you build the training distribution. Teffic-Audio keeps a standard pipeline—a Conformer speech encoder, multi-head attentive statistics pooling, and a binary classifier—and puts the work into multi-source open data, attack- and source-balanced sampling, and diverse waveform augmentation. On the 14 Speech-DF-Arena test sets it reaches 1.454% pooled equal error rate, better than the public leaderboard systems in the reported snapshot, with the lowest error on five individual sets and a better accuracy-per-parameter story than much larger models. Ablations say the diverse augmentation step is what most improves hard coding- and channel-heavy conditions, while encoder choice, pooling, and depth still matter. The practical claim is that a simple detector plus a carefully built training mix can serve as a strong reference system for general deepfake detection.

Core claim

With only open-source training data, a straightforward Conformer-based detector using multi-head attentive statistics pooling can outperform currently public Speech-DF-Arena systems, reaching 1.454% pooled EER across 14 heterogeneous test sets, when training uses multi-source corpora, attack- and source-balanced sampling, and diverse audio augmentation rather than heavier architectural complexity.

What carries the argument

The training recipe: multi-source deepfake plus supplementary bonafide corpora; attack- and source-balanced mini-batches (about 1000 samples per spoof unit, matched bonafide); and diverse waveform augmentations (RawBoost, RIR, MUSAN, pitch shift, filters, time masking, codecs, packet loss) applied with probability 0.5. That recipe, not a novel detector topology, is what the paper credits for cross-dataset generalization.

Load-bearing premise

The claim that leaderboard generalization proves a general detector still holds even though training includes official training splits from several of the same benchmark families used in evaluation, so some overlap and a fixed public snapshot partly carry the result.

What would settle it

Retrain with the same recipe but strip every training partition that belongs to any Speech-DF-Arena benchmark family, then re-score the identical 14-set pooled EER protocol; a large rise toward prior public systems would undercut the generalization claim.

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

If this is right

  • Future general deepfake detectors can prioritize training-distribution design over stacking more complex fusion or expert modules.
  • Diverse transmission-style augmentations (codecs, packet loss, filters) become first-class requirements for systems that must survive platform and channel shift.
  • Shallower Conformer stacks (e.g., 4–6 blocks) remain competitive under this recipe, offering a concrete path to lower-parameter reference models.
  • Supplementary real-speech diversity is treated as necessary alongside spoof diversity to stop real audio from being mistaken for fakes.

Where Pith is reading between the lines

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

  • If augmentation and sampling dominate gains, closed commercial systems may be winning mainly on proprietary data mix rather than secret architectures—suggesting open recipes could close much of the gap.
  • The same balanced multi-source recipe is a natural candidate for continual updates as new codecs and TTS families appear, without redesigning the detector head each time.
  • Partial train–test family overlap implies the field still needs a fully generator-held-out arena before ‘general detector’ rankings can be treated as deployment truth.

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 / 5 minor

Summary. This technical report introduces Teffic-Audio, a speech deepfake detector built from a standard pipeline: a w2v-BERT 2.0 Conformer encoder, multi-head attentive statistics pooling, and an MLP binary classifier. The authors argue that cross-dataset generalization is driven primarily by the training recipe—multi-source open data (deepfake corpora plus supplementary bonafide speech), attack- and source-balanced sampling, and a broad waveform augmentation suite—rather than architectural novelty. On the 14-set Speech-DF-Arena benchmark the system reports a pooled EER of 1.454%, ranking first among currently public leaderboard entries, with best individual EERs on five sets and competitive results from shallower (4–6 block) encoder variants. Staged ablations (Tables 4–5) attribute the largest gain to diverse augmentation and show backbone, pooling, and depth effects under a fixed recipe.

Significance. If the generalization claim holds under stricter held-out analysis, the paper offers a useful reference point for the field: a reproducible open-data recipe that matches or beats larger proprietary systems on a multi-benchmark arena, with clear ablations isolating balanced sampling, bonafide diversity, and especially multi-operator augmentation. The depth study (pooled EER 3.346% at N=4, 106M params) is practically valuable for deployment. Strengths include full EER/ACC/F1 tables, named leaderboard comparisons, and factor-by-factor training ablations under a fixed architecture. The main scientific contribution is empirical systems evidence rather than a new detection principle.

major comments (3)
  1. [§4.1, Table 1, Table 3] §4.1 and Table 1 vs Table 3: The training corpus includes official training partitions of several arena benchmarks (explicitly noted in §4.1), including DFADD and LibriSeVoc, on which Teffic-Audio reports 0.000 EER. Perfect scores on sets whose train splits were seen are consistent with residual protocol/generator overlap and weaken the strongest reading of “general” cross-dataset detection. Because pooled EER merges all scores under one threshold, near-zero sets can materially pull the headline 1.454%. Please (i) list exactly which evaluation sets have any train-partition overlap, (ii) report pooled and average EER after excluding those sets (or a leave-corpus-out protocol), and (iii) state whether DFADD/LSV test generators are disjoint from the training metadata. Without this isolation, the central leaderboard claim is only partially interpretable.
  2. [§3.3, Table 2, Table 4] Table 4: Diverse augmentation produces the dominant jump (pooled EER 5.896%→1.454%), with large gains on ASVspoof24-E and ADD22-T1. The paper does not disentangle which operators (codec, packet loss, filtering, etc.) drive those hard-set gains, nor whether the same operators appear in the construction of those test sets. A leave-one-operator-class ablation (acoustic/recording vs transmission/platform) or a short analysis linking Table 2 operators to the improved sets would make the key causal claim—that the augmentation suite, not leakage or generic regularization, explains generalization—much more secure.
  3. [§4.2, Table 3] §4.2 / Table 3: All primary comparisons are against a single public leaderboard snapshot that mixes proprietary closed systems of unknown training data with open baselines. The manuscript should clarify evaluation fairness: identical audio preprocessing and scoring, whether any proprietary systems may have trained on overlapping public sets, and sensitivity of rank to the pooled vs average metric (Teffic-Audio’s average EER 1.236% is slightly worse than Modulate-VELMA’s 1.104%). Reporting rank under both pooled and macro-average EER, and highlighting open-only comparisons, would better support the “outperforming all currently public systems” claim.
minor comments (5)
  1. [Title] Title/header typography is inconsistent (“TellFact fromFiction” / “TEFFIC-AUDIO: TELLFACT FROMFICTION”). Normalize spacing and hyphenation throughout.
  2. [§2] Eq. (1): yi ∈0,1 should be yi ∈ {0,1}; minor notation polish.
  3. [Table 1, §3.2] Table 1: ADD2023 Track1 lists “–” bonafide; briefly explain how bonafide balance is handled for units without paired real speech.
  4. [§3] Hyperparameters M=1000, augmentation probability 0.5, and operator ranges are stated but training schedule (epochs, LR, freezing of SSL layers, seed variance) is missing; add a short reproducibility subsection.
  5. [Appendix A] Appendix Tables 6–7 are helpful; cross-reference them earlier when claiming consistent ACC/F1 advantage.

Circularity Check

0 steps flagged

No derivation circularity: empirical systems result on an external multi-dataset leaderboard, not a self-defining prediction chain.

full rationale

Teffic-Audio is a standard train-and-evaluate speech deepfake detector. The load-bearing claim is an empirical pooled EER of 1.454% on Speech-DF-Arena’s 14 test sets under a fixed Conformer + MHASP architecture, attributed to multi-source data, attack/source-balanced sampling, and diverse augmentation (Abstract; §3; §4.2; Table 3). Ablations (Tables 4–5) vary one recipe or architecture factor at a time and report external metrics; nothing is defined in terms of the quantity it purports to predict. There is no fitted scalar renamed as a first-principles forecast, no uniqueness theorem imported from overlapping authors, and no ansatz smuggled in via self-citation. Partial inclusion of official training partitions of some arena benchmarks (§4.1; Table 1) and perfect EERs on DFADD/LSV raise evaluation-contamination / generalization-validity concerns, but those are protocol risks, not circular reduction of a derivation to its inputs. Against an external leaderboard snapshot the paper is self-contained; score 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The central claim is empirical performance under a chosen protocol. It rests on standard SSL/detection modeling choices, hand-set sampling and augmentation hyperparameters, and the assumption that Speech-DF-Arena pooled EER on the listed 14 sets is the right generalization yardstick. No new physical entities; free parameters are recipe knobs, not fitted physical constants.

free parameters (4)
  • spoof-unit sample count M = 1000
    Attack/source-balanced sampling draws M utterances per spoofing unit each epoch; M is set by authors, not derived.
  • augmentation apply probability and operator count = p=0.5; 1–2 operators
    Each sample is augmented with probability 0.5 using one or two randomly chosen operators; controls effective training distribution.
  • MHASP head count and MLP widths = 4 heads; 2048→1536→1024; clf 512
    4 attention heads, 2048-d pool, MLP 2048→1536→1024 and classifier hidden 512 are architectural choices that affect the reported EER.
  • augmentation numeric ranges = Table 2 ranges
    SNR bands, pitch ±1 st, filter cutoffs, packet-loss rates, codec list, etc. (Table 2) are hand-chosen simulation ranges that drive the key ablation gain.
axioms (5)
  • domain assumption Binary utterance-level BCE on spoof vs bonafide with sigmoid score and EER/ACC/F1 under Speech-DF-Arena pooled thresholding is an adequate measure of general deepfake detection.
    Entire evaluation and ranking (§4.1–4.2) assume this protocol and metrics without alternative decision costs or calibration analysis.
  • domain assumption w2v-BERT 2.0 Conformer representations plus temporal attentive statistics retain spoof artifacts across TTS/VC/vocoder/codec and channel conditions.
    System overview §2 treats the SSL encoder and MHASP as sufficient feature path; ablations support but do not prove universality.
  • domain assumption Open-source multi-corpus mixing with generator/source balancing approximates the heterogeneity of the 14-set arena without needing proprietary data.
    Claim of open-source-only training and generalization (§3, Abstract) depends on this coverage assumption.
  • standard math Standard neural training (end-to-end SGD-style optimization of encoder, pool, and classifier) converges to a stable detector under the stated recipe.
    Ordinary supervised learning assumptions; no exotic proof obligations.
  • ad hoc to paper Waveform augmentations in Table 2 are representative enough of real recording/transmission chains that gains on arena hard sets transfer to practical deployment.
    Key performance jump is attributed to Diverse Audio Augmentation (§5.1); representativeness is postulated, not independently validated outside the arena.
invented entities (1)
  • Teffic-Audio (named system / training recipe bundle) no independent evidence
    purpose: Package the specific encoder+pooling+multi-source balanced+diverse-augmentation stack as a reference detector.
    Brand-level system name; not a new scientific object beyond the engineering combination. No independent existence outside this report's configuration.

pith-pipeline@v1.2.0-daily-grok45 · 29568 in / 3794 out tokens · 81088 ms · 2026-07-31T10:26:30.952141+00:00 · methodology

0 comments
read the original abstract

Speech deepfake detection has expanded in scope with increasingly heterogeneous spoofing mechanisms, including speech synthesis, voice conversion, vocoder reconstruction, and neural-codec resynthesis. The resulting spoofing artifacts can be further shaped by variability in source speech, recording environments, and transmission channels. This variability makes robust generalization across heterogeneous conditions a central requirement for practical detection systems. This report presents Teffic-Audio, a general speech deepfake detection system designed for comprehensive evaluation environment. Teffic-Audio adopts a straightforward detector architecture consisting of a Conformer-based speech encoder, multi-head attentive statistics pooling, and a binary classifier. Rather than relying on additional architectural complexity, the system improves generalization through its training recipe, which integrates multi-source data, attack- and source-balanced sampling, and diverse audio augmentation. Trained only with open-source data, Teffic-Audio achieves a pooled EER of 1.454% on the 14 test sets of Speech-DF-Arena, outperforming all currently public systems on the leaderboard. It also obtains the lowest EER on five individual test sets and shows a favorable performance-complexity trade-off compared with larger leading systems. Overall, Teffic-Audio provides a strong and practical reference system for general speech deepfake detection.

Figures

Figures reproduced from arXiv: 2607.28351 by Jindong Wang, Kunyu Feng, Li Wang, Wan Lin, Zhizheng Wu.

Figure 1
Figure 1. Figure 1: Overview of the Teffic-Audio system architecture and the waveform-to-score training pipeline. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] 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

40 extracted references · 1 canonical work pages

  1. [6]

    Harder or different? understanding generalization of audio deepfake detection

    Nicolas M Müller, Nicholas Evans, Hemlata Tak, Philip Sperl, and Konstantin Böttinger. Harder or different? understanding generalization of audio deepfake detection. InProc. Interspeech 2024, pages 2705–2709, 2024b. Nuria Alina Chandra, Ryan Murtfeldt, Lin Qiu, Arnab Karmakar, Hannah Lee, Emmanuel Tanumihardja, Kevin Farhat, Ben Caffee, Sejin Paik, Changy...

    Pith/arXiv arXiv 2024

  2. [10]

    Does audio deepfake detection generalize?arXiv preprint arXiv:2203.16263,

    10 Teffic-Audio: Tell Fact from FictionTECHNICALREPORT Nicolas M Müller, Pavel Czempin, Franziska Dieckmann, Adam Froghyar, and Konstantin Böttinger. Does audio deepfake detection generalize?arXiv preprint arXiv:2203.16263,

    arXiv

  3. [11]

    Asvspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale

    Xin Wang, Héctor Delgado, Hemlata Tak, Jee-weon Jung, Hye-jin Shim, Massimiliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidullah, Tomi H Kinnunen, et al. Asvspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale. InProc. ASVspoof 2024, pages 1–8,

    2024

  4. [12]

    Light convolutional neural network with feature genuinization for detection of synthetic speech attacks

    Zhenzong Wu, Rohan Kumar Das, Jichen Yang, and Haizhou Li. Light convolutional neural network with feature genuinization for detection of synthetic speech attacks. InProc. Interspeech 2020, pages 1101–1105,

    2020

  5. [13]

    Replay and synthetic speech detection with res2net architecture

    Xu Li, Na Li, Chao Weng, Xunying Liu, Dan Su, Dong Yu, and Helen Meng. Replay and synthetic speech detection with res2net architecture. InICASSP 2021-2021 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 6354–6358. IEEE,

    2021

  6. [14]

    End-to-end anti-spoofing with rawnet2

    Hemlata Tak, Jose Patino, Massimiliano Todisco, Andreas Nautsch, Nicholas Evans, and Anthony Larcher. End-to-end anti-spoofing with rawnet2. InICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6369–6373. IEEE, 2021a. Jee-weon Jung, Hee-Soo Heo, Hemlata Tak, Hye-jin Shim, Joon Son Chung, Bong-Jin Lee,...

    2021

  7. [15]

    Xls-r: Self-supervised cross-lingual speech representation learning at scale.arXiv preprint arXiv:2111.09296,

    Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick V on Platen, Yatharth Saraf, Juan Pino, et al. Xls-r: Self-supervised cross-lingual speech representation learning at scale.arXiv preprint arXiv:2111.09296,

    Pith/arXiv arXiv

  8. [16]

    Mixture of experts fusion for fake audio detection using frozen wav2vec 2.0

    Zhiyong Wang, Ruibo Fu, Zhengqi Wen, Jianhua Tao, Xiaopeng Wang, Yuankun Xie, Xin Qi, Shuchen Shi, Yi Lu, Yukun Liu, et al. Mixture of experts fusion for fake audio detection using frozen wav2vec 2.0. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE,

    2025

  9. [17]

    Temporal-channel modeling in multi-head self-attention for synthetic speech detection.arXiv preprint arXiv:2406.17376,

    Duc-Tuan Truong, Ruijie Tao, Tuan Nguyen, Hieu-Thi Luong, Kong Aik Lee, and Eng Siong Chng. Temporal-channel modeling in multi-head self-attention for synthetic speech detection.arXiv preprint arXiv:2406.17376,

    Pith/arXiv arXiv

  10. [18]

    Frame-level temporal difference learning for partial deepfake speech detection.IEEE Signal Processing Letters, 2025b

    Menglu Li, Xiao-Ping Zhang, and Lian Zhao. Frame-level temporal difference learning for partial deepfake speech detection.IEEE Signal Processing Letters, 2025b. Hoan My Tran, Damien Lolive, David Guennec, Aghilas Sini, Arnaud Delhay, and Pierre-François Marteau. Leveraging ssl speech features and mamba for enhanced deepfake detection. InInterspeech 2025, ...

    2025

  11. [19]

    Leveraging mixture of experts for improved speech deepfake detection

    Viola Negroni, Davide Salvi, Alessandro Ilic Mezza, Paolo Bestagini, and Stefano Tubaro. Leveraging mixture of experts for improved speech deepfake detection. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE,

    2025

  12. [20]

    Molex: Mixture of lora experts in speech self-supervised models for audio deepfake detection.arXiv preprint arXiv:2509.09175,

    Zihan Pan, Sailor Hardik Bhupendra, and Jinyang Wu. Molex: Mixture of lora experts in speech self-supervised models for audio deepfake detection.arXiv preprint arXiv:2509.09175,

    Pith/arXiv arXiv

  13. [21]

    Rawboost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing

    Hemlata Tak, Madhu Kamble, Jose Patino, Massimiliano Todisco, and Nicholas Evans. Rawboost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing. InICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6382–6386. IEEE, 2022a. Wanying Ge, Xin Wang, Xuechen Li...

    arXiv 2022

  14. [22]

    A data-centric approach to generalizable speech deepfake detection

    Wen Huang, Yuchen Mao, and Yanmin Qian. A data-centric approach to generalizable speech deepfake detection. arXiv preprint arXiv:2512.18210, 2025a. Mingru Yang, Yanmei Gu, Qianhua He, Yanxiong Li, Peirong Zhang, Yongqiang Chen, Zhiming Wang, Huijia Zhu, Jian Liu, and Weiqiang Wang. Generalizable audio deepfake detection via hierarchical structure learning...

    arXiv 2025

  15. [23]

    Generalizable audio deepfake detection via latent space refinement and augmentation

    11 Teffic-Audio: Tell Fact from FictionTECHNICALREPORT Wen Huang, Yanmei Gu, Zhiming Wang, Huijia Zhu, and Yanmin Qian. Generalizable audio deepfake detection via latent space refinement and augmentation. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2025b. Pu Huang, Shouguang Wang, ...

    2025

  16. [24]

    Generalizable speech deepfake detection via meta-learned lora

    Janne Laakkonen, Ivan Kukanov, and Ville Hautamäki. Generalizable speech deepfake detection via meta-learned lora. InICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 19092–19096. IEEE,

    2026

  17. [25]

    Conformer: Convolution-augmented transformer for speech recognition

    Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, et al. Conformer: Convolution-augmented transformer for speech recognition. InProc. Interspeech 2020, pages 5036–5040,

    2020

  18. [26]

    Attentive statistics pooling for deep speaker embedding.arXiv preprint arXiv:1803.10963,

    Koji Okabe, Takafumi Koshinaka, and Koichi Shinoda. Attentive statistics pooling for deep speaker embedding.arXiv preprint arXiv:1803.10963,

    Pith/arXiv arXiv

  19. [28]

    Replay attacks against audio deepfake detection

    Nicolas Müller, Piotr Kawa, Wei-Herng Choong, Adriana Stan, Aditya Tirumala Bukkapatnam, Karla Pizzi, Alexander Wagner, and Philip Sperl. Replay attacks against audio deepfake detection. InProc. Interspeech 2025, pages 2245–2249,

    2025

  20. [29]

    Aishell-3: A multi-speaker mandarin tts corpus and the baselines

    Yao Shi, Hui Bu, Xin Xu, Shaoji Zhang, and Ming Li. Aishell-3: A multi-speaker mandarin tts corpus and the baselines. arXiv preprint arXiv:2010.11567,

    Pith/arXiv arXiv 2010

  21. [30]

    Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio.arXiv preprint arXiv:2106.06909,

    Guoguo Chen, Shuzhou Chai, Guanbo Wang, Jiayu Du, Wei-Qiang Zhang, Chao Weng, Dan Su, Daniel Povey, Jan Trmal, Junbo Zhang, et al. Gigaspeech: An evolving, multi-domain asr corpus with 10,000 hours of transcribed audio.arXiv preprint arXiv:2106.06909,

    Pith/arXiv arXiv

  22. [31]

    Cn-celeb: a challenging chinese speaker recognition dataset

    Yue Fan, JW Kang, LT Li, KC Li, HL Chen, ST Cheng, PY Zhang, ZY Zhou, YQ Cai, and Dong Wang. Cn-celeb: a challenging chinese speaker recognition dataset. InICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7604–7608. IEEE,

    2020

  23. [32]

    Bona fide cross testing reveals weak spot in audio deepfake detection systems.arXiv preprint arXiv:2509.09204,

    Chin Yuen Kwok, Jia Qi Yip, Zhen Qiu, Chi Hung Chi, and Kwok Yan Lam. Bona fide cross testing reveals weak spot in audio deepfake detection systems.arXiv preprint arXiv:2509.09204,

    Pith/arXiv arXiv

  24. [33]

    Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining

    12 Teffic-Audio: Tell Fact from FictionTECHNICALREPORT Hyung Won Chung, Xavier Garcia, Adam Roberts, Yi Tay, Orhan Firat, Sharan Narang, and Noah Constant. Unimax: Fairer and more effective language sampling for large-scale multilingual pretraining. InThe Eleventh International Conference on Learning Representations. Jiasheng Ye, Peiju Liu, Tianxiang Sun,...

    2025

  25. [35]

    High fidelity neural audio compression.arXiv preprint arXiv:2210.13438,

    Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High fidelity neural audio compression.arXiv preprint arXiv:2210.13438,

    Pith/arXiv arXiv

  26. [36]

    Bicrossmamba-st: speech deepfake detection with bidirectional mamba spectro-temporal cross-attention.arXiv preprint arXiv:2505.13930,

    Yassine El Kheir, Tim Polzehl, and Sebastian Möller. Bicrossmamba-st: speech deepfake detection with bidirectional mamba spectro-temporal cross-attention.arXiv preprint arXiv:2505.13930,

    Pith/arXiv arXiv

  27. [37]

    Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation.arXiv preprint arXiv:2202.12233, 2022b

    Hemlata Tak, Massimiliano Todisco, Xin Wang, Jee-weon Jung, Junichi Yamagishi, and Nicholas Evans. Automatic speaker verification spoofing and deepfake detection using wav2vec 2.0 and data augmentation.arXiv preprint arXiv:2202.12233, 2022b. Piotr Kawa, Marcin Plata, Micha´L Czuba, Piotr Syga, et al. Improved deepfake detection using whisper features.arXi...

    Pith/arXiv arXiv

  28. [38]

    Exploring generalization to unseen audio data for spoofing: Insights from ssl models

    Atharva Kulkarni, Hoan My Tran, Ajinkya Kulkarni, Sandipana Dowerah, Damien Lolive, and Mathew Magimai Doss. Exploring generalization to unseen audio data for spoofing: Insights from ssl models. InASVSpoof workshop 2024,

    2024

  29. [39]

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

    Hemlata Tak, Jee-weon Jung, Jose Patino, Madhu Kamble, Massimiliano Todisco, and Nicholas Evans. End-to-end spectro-temporal graph attention networks for speaker verification anti-spoofing and speech deepfake detection. arXiv preprint arXiv:2107.12710, 2021b. Yang Xiao, Ting Dang, and Rohan Kumar Das. Rawtfnet: A lightweight cnn architecture for speech an...

    Pith/arXiv arXiv 2025

  30. [40]

    Teffic-Audio achieves the best pooled ACC and pooled F1-score among the listed systems

    14 Teffic-Audio: Tell Fact from FictionTECHNICALREPORT A Additional Leaderboard Performances Tables 6 and 7 report additional ACC and F1-score comparisons on the Speech-DF-Arena. Teffic-Audio achieves the best pooled ACC and pooled F1-score among the listed systems. Together with the EER results in the main text, these results show a consistent performanc...

    arXiv 2026

  31. [2015]

    Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Hector Delgado, Andreas Nautsch, Nicholas Evans, Md Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, et al

    doi:10.21437/Interspeech.2015-462. Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Hector Delgado, Andreas Nautsch, Nicholas Evans, Md Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, et al. ASVspoof 2019: A large-scale pub- lic database of synthesized, converted and replayed speech.Computer Speech & Language, 64:101114,

    work page doi:10.21437/interspeech.2015-462 2015

  32. [2017]

    Musan: A music, speech, and noise corpus.arXiv preprint arXiv:1510.08484,

    David Snyder, Guoguo Chen, and Daniel Povey. Musan: A music, speech, and noise corpus.arXiv preprint arXiv:1510.08484,

    Pith/arXiv arXiv

  33. [2018]

    Seamless: Multilingual expressive and streaming speech translation.arXiv preprint arXiv:2312.05187,

    Loïc Barrault, Yu-An Chung, Mariano Coria Meglioli, David Dale, Ning Dong, Mark Duppenthaler, Paul-Ambroise Duquenne, Brian Ellis, Hady Elsahar, Justin Haaheim, et al. Seamless: Multilingual expressive and streaming speech translation.arXiv preprint arXiv:2312.05187,

    Pith/arXiv arXiv

  34. [2020]

    Nicolas M Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Böttinger

    doi:10.1016/j.csl.2020.101114. Nicolas M Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Böttinger. Mlaad: The multi-language audio anti-spoofing dataset. In2024 International Joint Conference on Neural Networks (IJCNN), pages 1–7. IEEE, 2024a. Menglu Li, Yasaman Ahmadiadli, an...

    arXiv 2020

  35. [2021]

    Add 2022: the first audio deep synthesis detection challenge

    Jiangyan Yi, Ruibo Fu, Jianhua Tao, Shuai Nie, Haoxin Ma, Chenglong Wang, Tao Wang, Zhengkun Tian, Ye Bai, Cunhang Fan, et al. Add 2022: the first audio deep synthesis detection challenge. InICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 9216–9220. IEEE,

    2022

  36. [2022]

    Add 2023: the second audio deepfake detection challenge.arXiv preprint arXiv:2305.13774,

    Jiangyan Yi, Jianhua Tao, Ruibo Fu, Xinrui Yan, Chenglong Wang, Tao Wang, Chu Yuan Zhang, Xiaohui Zhang, Yan Zhao, Yong Ren, et al. Add 2023: the second audio deepfake detection challenge.arXiv preprint arXiv:2305.13774,

    Pith/arXiv arXiv 2023

  37. [2023]

    Wavefake: A data set to facilitate audio deepfake detection.arXiv preprint arXiv:2111.02813,

    Joel Frank and Lea Schönherr. Wavefake: A data set to facilitate audio deepfake detection.arXiv preprint arXiv:2111.02813,

    Pith/arXiv arXiv

  38. [2024]

    Cosyvoice 2: Scalable streaming speech synthesis with large language models.arXiv preprint arXiv:2412.10117, 2024a

    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, 2024a. Sanyuan Chen, Shujie Liu, Long Zhou, Yanqing Liu, Xu Tan, Jinyu Li, Sheng Zhao, Yao Qian, and Furu Wei. Vall-e 2: Neur...

    Pith/arXiv arXiv

  39. [2025]

    Asvspoof 2015: the first automatic speaker verification spoofing and countermeasures challenge

    Zhizheng Wu, Tomi Kinnunen, Nicholas Evans, Junichi Yamagishi, Cemal Hanilçi, and Md Sahidullah. Asvspoof 2015: the first automatic speaker verification spoofing and countermeasures challenge. 09

    2015

  40. [2026]

    Asvspoof 2021: accelerating progress in spoofed and deepfake speech detection

    Junichi Yamagishi, Xin Wang, Massimiliano Todisco, Md Sahidullah, Jose Patino, Andreas Nautsch, Xuechen Liu, Kong Aik Lee, Tomi Kinnunen, Nicholas Evans, et al. Asvspoof 2021: accelerating progress in spoofed and deepfake speech detection. InProc. ASVSPOOF 2021, pages 47–54,

    2021