Pith. sign in

REVIEW 4 major objections 6 minor 53 references

Pretrained audio language models are near-random at spoofing-aware speaker verification, but task-specific fine-tuning with a composite three-head loss makes them competitive with — and on the primary metric, better than — conventional fusi

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 · deepseek-v4-flash

2026-08-02 01:08 UTC pith:Y4POLWER

load-bearing objection First systematic LALM-for-SASV study with a clean adaptation story, but the claim of beating conventional fusion rests on a re-implemented baseline the authors themselves say is not leaderboard-comparable. the 4 major comments →

arxiv 2607.14753 v1 pith:Y4POLWER submitted 2026-07-16 cs.SD cs.AI

Large Audio Language Models for Spoofing-Aware Speaker Verification

Sofya Savelyeva , Mariia Perunova , Evgeny Kushnir , Artem Dvirniak , Dmitrii Korzh , Oleg Y. Rogov This is my paper
classification cs.SD cs.AI
keywords spoofing-aware speaker verificationlarge audio language modelsspeech deepfake detectionlow-rank adaptationchain-of-thought reasoningreinforcement learningvoice authenticationanti-spoofing
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.

This paper asks whether large audio language models — systems that listen and reply in text — can be taught to decide whether a voice sample is the claimed speaker, a different human, or a synthetic spoof. It finds that these models are near chance without training, but that fine-tuning with a combined objective — a three-way decision loss, a speaker-identity margin loss, and a separate spoof-detection loss — turns them into competitive verification systems. Under the authors' evaluation protocol, the best adapted model reaches 89.3% accuracy and a min a-DCF of 0.19, beating their reproduced conventional fusion pipelines. Reasoning supervision does not beat hard-label training, but it adds interpretable justifications at a small cost to accuracy.

Core claim

The central discovery is that a pretrained audio language model can be converted into a spoofing-aware speaker verifier by supervised adaptation with a composite loss. The three losses act on three different representations: a cross-entropy term over the target/non-target/spoof decision, an additive angular margin loss over speaker embeddings from the audio-conditioned features (applied only to genuine speech), and a binary bona-fide/spoof head. Hard-sample mining, which repeatedly feeds the model the most confusing enrollment-trial pairs, completes the recipe. The resulting model outperforms the reproduced conventional fusion baselines on accuracy and the primary detection-cost metric, alth

What carries the argument

The key mechanism is the three-head composite training objective. A three-class cross-entropy loss supervises the final decision; an additive angular-margin (AAM) loss, applied only to bona fide samples, pulls genuine embeddings together by identity; and a binary spoof-detection loss sharpens sensitivity to synthetic speech. These heads read from different representations — the audio-conditioned feature pooler feeds the speaker head, while a separate linear head classifies spoofiness. Hard-sample mining then concentrates training on boundary pairs. The model's score for threshold metrics is recovered by the probability of the 'target' answer token.

Load-bearing premise

The paper's own limitations section concedes that reasoning traces are not yet validated and that its re-implemented top baseline is not directly comparable to official leaderboard results; beyond those, the load-bearing assumption is that a single stratified sample of 20K evaluation pairs faithfully represents the full evaluation set, and that the authors' re-implemented fusion pipelines fairly represent conventional SASV systems — with no confidence intervals reported.

What would settle it

Evaluate the adapted model on the complete official evaluation set and compare its min a-DCF to the published leaderboard; if the 0.19 a-DCF does not reproduce outside the 20K sample, the paper's core comparison collapses. A cheaper check: resample several 20K subsets and see whether the accuracy and a-DCF vary beyond the reported margin over baselines.

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

If this is right

  • A single audio language model can produce both an accept/reject decision and a natural-language rationale, enabling auditable voice authentication.
  • The three-loss recipe transfers the large-model advantage to other joint discrimination tasks, such as deepfake detection with identity verification.
  • Because conventional fusion systems still achieve better equal-error rates, hybrid designs could pair LALM decisions with classical threshold scoring.
  • Reasoning-oriented training, though not more accurate, gives system builders a way to inspect why a trial was rejected.

Where Pith is reading between the lines

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

  • If the 20-thousand-pair sample is representative of the full evaluation set, the margin over conventional fusion suggests LALM-based verification can handle real-world distribution shift; the paper's own caution that official leaderboards may differ leaves this open.
  • The same composite-loss adaptation could be tested on harder protocols with multiple enrollment utterances, where the paper already notes its single-enrollment setup is conservative.
  • The reasoning traces, once human-validated, could provide weak labels for training smaller, faster verifiers that mimic the large model's judgments.

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

4 major / 6 minor

Summary. The paper presents the first systematic study of large audio language models (LALMs) for spoofing-aware speaker verification (SASV). SASV is cast as a pairwise three-way decision (target / non-target / spoof) over enrollment and trial audio, and the authors evaluate zero-shot prompting, LoRA-based supervised fine-tuning, a composite training objective combining three-class cross-entropy with AAM speaker and binary spoof heads, hard-sample mining, chain-of-thought fine-tuning, and GRPO-based optimization. Experiments are conducted on a stratified 20K-pair sample of the ASVspoof5 evaluation set under a single-enrollment protocol. The reported results show near-chance zero-shot performance, large gains after LoRA adaptation, and a final SALMONN-7B configuration reaching 89.30% accuracy and 0.19 min a-DCF, which the authors interpret as competitive with or better than two reproduced fusion baselines. A data-matched comparison on 90K pairs shows that reasoning-supervised variants do not beat hard-label fine-tuning on accuracy, though GRPO improves countermeasure accuracy. The paper argues that LALMs are a promising auditable foundation for unified SASV, while acknowledging computational cost and several protocol caveats.

Significance. If the empirical claims are fully supported, this would be a useful contribution: it is the first comparative study of LALMs for joint speaker/anti-spoofing decisions; it maps out a composite-loss design space (Eq. 2) and shows which components move ASV vs CM accuracy; it includes a data-matched reasoning-supervision comparison; and it is unusually candid in stating limitations (single-enrollment protocol, 20K subset, non-comparable reimplementation of the top ASVspoof5 baseline). The paper also attempts to ground chain-of-thought traces with conventional acoustic/speaker features and filtering, which is a constructive design. However, the headline comparison with conventional SASV currently rests on baselines that the authors themselves describe as not directly comparable to official leaderboard results, and all metrics are reported without confidence intervals. The significance therefore hinges on whether the baseline gap is real or an artifact of the reimplementation/sample; as written, the central claim is conditional rather than established.

major comments (4)
  1. [§IV-C, Table II; §VI RQ2] The central claim that adapted LALMs are competitive with or outperform conventional SASV rests primarily on a reimplementation of the [50] power-fusion system. Section IV-C explicitly states that the original code/checkpoints are unavailable, and that the achieved scores are 'not directly comparable to official leaderboard results.' Published ASVspoof5 top systems typically achieve far lower min a-DCF than the 0.41 reported here for the reimplementation. If the official [50] system or another strong conventional system were evaluated on the same 20K subset and scored below 0.19, the comparison inverts. The authors need to provide a stronger baseline evaluation — either official checkpoints/scores on the same subset or published leaderboard numbers for context — or restrict the abstract and RQ2 conclusion to 'under our reproduced baselines.'
  2. [§IV-B, Tables I–III] All metrics are computed on a single stratified sample of 20K pairs, with no confidence intervals, no repeated sampling, and no full-eval check. The protocol also differs from the official ASVspoof5 Track 2 protocol in using one enrollment instead of three and a truncated 8-s evaluation window. These choices can shift metric values materially. The 0.19 vs 0.31–0.41 differences may be robust, but without error bars or a full-evaluation sanity check the paper cannot rule out that the gap is an artifact of the sample/baseline protocol. Please provide bootstrap confidence intervals, a subset-size sensitivity analysis, or a full-evaluation run on the conventional baselines.
  3. [Table II, §III-B, §IV] The composite objective in Eq. (2) is central to the proposed method, but the weights λ1=1.0, λ2=0.5, λ3=1.0 and the hard-sample-mining thresholds are hand-chosen with no sensitivity analysis. This is load-bearing because the ablation shows a dramatic failure mode: adding only the AAM head drops overall accuracy from 85.84% to 59.82% and CM accuracy from 96.47% to 21.00% (Table II). That suggests the trade-off is sharp and the final configuration may be tuned to the particular eval sample. At minimum, report a small sweep over λ2 and the mining thresholds and show per-seed variance; otherwise the robustness of the 'balanced' recipe is unsubstantiated.
  4. [§III-C, Table III, §VI] The paper motivates reasoning supervision as a route to 'auditable' SASV and claims interpretable rationales as a key benefit, but the reasoning traces are not evaluated at all: there is no human evaluation, no faithfulness/grounding metric, and no comparison of rationales against the conventional features used to construct them. Table III itself shows CoT-SFT and GRPO do not improve overall accuracy over hard-label SFT. Given the paper's framing ('auditable foundation', 'rationales for auditing'), the absence of any reasoning-quality measurement makes the interpretability claim unsupported. The authors acknowledge this in Section VI, but it should be either tested or clearly downgraded from a contribution to a conjecture.
minor comments (6)
  1. [§III-B] Typo: 'an direct SFT' should be 'a direct SFT'.
  2. [§IV-B] The score-extraction procedure uses the target-token posterior, and the threshold is swept to report min a-DCF. Please clarify explicitly whether the threshold is selected on the same 20K eval sample; if so, min a-DCF is an optimistic, oracle-threshold metric and should be labeled as such.
  3. [§IV-A] The reasoning-trace filtering uses a Jaccard threshold of 0.35 and a 2-of-3 correctness rule, but no sensitivity of downstream CoT-SFT/GRPO to these thresholds is given. A sentence on why these values were chosen would help.
  4. [§IV] 'approximately1.8 million pairs spanning500speakers' lacks precision; please report exact counts for the main training set and the 90K reasoning subset.
  5. [Fig. 1] The prompt templates in Fig. 1 are central to reproducibility, but the figure is only described in prose; please ensure the exact prompt strings are included in the paper or appendix.
  6. [§I, §VI] The phrase 'outperforming strong baselines' in the introduction/abstract is stronger than what Section IV-C and Section VI ultimately support; consider aligning the wording with the controlled-protocol caveat.

Circularity Check

0 steps flagged

No significant circularity: the central SASV claims are tested against the external ASVspoof5 eval set, not derived from the paper's inputs; the only self-citation is a minor, non-load-bearing support for the reasoning-supervision branch.

full rationale

The paper's main claim chain is empirical rather than derivational: pretrained LALMs are evaluated zero-shot on an external benchmark, then adapted with composite losses (Eqs. 1-2) and compared on a 20K stratified sample of the ASVspoof5 eval set using the standard min a-DCF metric. None of the reported predictions are defined in terms of the claims they support: the zero-shot result is a direct measurement, the LoRA-SFT and ablation results are trained and evaluated on disjoint data, and the a-DCF score is obtained from a standard threshold sweep over the model's p_target posterior. The reproduced fusion baselines are acknowledged by the paper to be 'not directly comparable to the official leaderboard results' (Sec. IV-C), and the 20K sample is described as a stratified subset; these are external-validity and benchmark-fidelity limitations, not circularity, because the comparison is conditional and falsifiable. The only self-citation is [43], used in Sec. III-C to justify human-cue CoT supervision: 'A complementary line of research shows that human perceptible cues are predictive of synthesized speech and can be used as supervision for LALMs [43].' This is not load-bearing: the reasoning-supervised models do not improve the hard-label result (Table III), and the central claim (adapted LALMs achieve 89.30% accuracy / 0.19 min a-DCF) survives independently of it. Therefore no step reduces by construction to its inputs, and the self-citation concern does not warrant a score above 2.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No invented entities; the paper is empirical. Free parameters are mostly training and annotation hyperparameters rather than fitted constants of a model. The most load-bearing assumptions are the representativeness of the 20K eval subset and the fidelity of the re-implemented baselines.

free parameters (5)
  • Composite loss weights (lambda1, lambda2, lambda3) = 1.0, 0.5, 1.0
    Eq. (2) combines CE, AAM and spoof-BCE terms; weights are hand-chosen and no sensitivity analysis is given, yet they directly control the ASV-CM trade-off in Table II.
  • AAM angular margin m = not reported
    Eq. (1) uses ArcFace-style margin; value is not stated, and it determines speaker-separability pressure on bona fide pairs.
  • Hard-sample mining thresholds = not reported
    Sec III-B defines hard-yes/no/spoof pairs qualitatively but no numerical thresholds or update schedule; these materially change the training distribution and final metrics.
  • CoT filtering thresholds = correct in >=2 of 3 traces; Jaccard >=0.35
    Sec IV-A: these thresholds determine which ~90K reasoning traces are kept; no sensitivity analysis is reported.
  • Augmentation probabilities (noise, RIR, RawBoost) = 0.1, 0.07, 0.07
    Hand-picked augmentation schedule; affects robustness and the reported accuracies.
axioms (5)
  • domain assumption The 20K stratified subset of ASVspoof5 eval faithfully represents the full eval distribution.
    All metrics are computed on this subset (Sec IV-B) with no confidence intervals, so the headline numbers assume subset stability.
  • domain assumption Author re-implementations of conventional fusion baselines faithfully represent the published systems.
    Sec IV-C states official [50] code/checkpoints are unavailable and the re-implementation is trained under a different protocol; the 'outperforms baselines' claim rests on this fidelity.
  • domain assumption Softmax over decision-token logits (sasv=p_target) is a valid score for EER/a-DCF.
    Sec IV-B recovers scores from token posteriors; the paper itself notes these are not calibrated, especially for CoT models, yet uses them for threshold metrics.
  • domain assumption Filtered machine-generated reasoning traces are grounded enough to supervise CoT.
    Sec IV-A retains traces only if label-correct and Jaccard-consistent, but human validation is deferred; the paper acknowledges possible hallucination and limited reliability.
  • domain assumption Human-perceptible paralinguistic cues (age, gender, accent, emotion, prosody) are predictive of spoofing and useful as reasoning supervision.
    Inherited from the authors' own prior work [43] and used to build CoT traces; no independent evidence is provided in this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 12024 in / 15124 out tokens · 124771 ms · 2026-08-02T01:08:06.898605+00:00 · methodology

0 comments
read the original abstract

Recent advances in text-to-speech and voice cloning make high-quality spoofing inexpensive and scalable, threatening voice authentication systems, especially automatic speaker verification (ASV). Existing defenses mainly address this threat through binary countermeasures (CMs) for deepfake detection or spoofing-aware speaker verification (SASV), where current systems are dominated by modular ASV-CM fusion and cascaded pipelines. Although large audio language models (LALMs) have shown promise on related audio tasks, including CM and ASV, their use for SASV remains unexplored, despite their capacity to produce natural-language rationales for auditing and robustness beyond discriminative predictions. This work systematically evaluates LALMs for SASV against conventional pipelines under zero-shot prompting, supervised adaptation, reasoning-oriented training, and reinforcement-learning-based optimization. Our results show that pretrained LALMs are near chance in the zero-shot setting, confirming that they are not natively suited to SASV, but that task-specific adaptation closes this gap. We further find that competitive SASV performance can be achieved through several distinct routes. These findings position LALMs as a promising and auditable foundation for unified SASV, while clarifying where conventional cascade systems still lead.

Figures

Figures reproduced from arXiv: 2607.14753 by Artem Dvirniak, Dmitrii Korzh, Evgeny Kushnir, Mariia Perunova, Oleg Y. Rogov, Sofya Savelyeva.

Figure 1
Figure 1. Figure 1: Prompt variants for LALM-based SASV. Top: direct decision over [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed SALMONN-based SASV model. [PITH_FULL_IMAGE:figures/full_fig_p004_2.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

53 extracted references · 22 linked inside Pith

  1. [1]

    Xtts: a massively multilin- gual zero-shot text-to-speech model,

    E. Casanova, K. Davis, E. G ¨olge, G. G ¨oknar, I. Gulea, L. Hart, A. Alja- fari, J. Meyer, R. Morais, S. Olayemiet al., “Xtts: a massively multilin- gual zero-shot text-to-speech model,”arXiv preprint arXiv:2406.04904, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    Cosyvoice 3: Towards in-the-wild speech generation via scaling-up and post-training,

    Z. Du, C. Gao, Y . Wang, F. Yu, T. Zhao, H. Wang, X. Lv, H. Wang, C. Ni, X. Shiet al., “Cosyvoice 3: Towards in-the-wild speech generation via scaling-up and post-training,”arXiv preprint arXiv:2505.17589, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Add 2023: the second audio deepfake detection challenge,

    J. Yiet al., “Add 2023: the second audio deepfake detection challenge,” arXiv preprint arXiv:2305.13774, 2023

    Pith/arXiv arXiv 2023

  4. [4]

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

    X. Wang, H. Delgado, H. Tak, J.-w. Jung, H.-j. Shim, M. Todisco, I. Kukanov, X. Liu, M. Sahidullah, T. Kinnunenet al., “Asvspoof 5: Crowdsourced speech data, deepfakes, and adversarial attacks at scale,” arXiv preprint arXiv:2408.08739, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Sasv 2022: The first spoofing-aware speaker verification challenge,

    J.-w. Jung, H. Tak, H.-j. Shim, H.-S. Heo, B.-J. Lee, S.-W. Chung, H.-J. Yu, N. Evans, and T. Kinnunen, “Sasv 2022: The first spoofing-aware speaker verification challenge,”Interspeech 2022, pp. 2893–2897, 2022

    2022

  6. [6]

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

    B. Desplanques, J. Thienpondt, and K. Demuynck, “Ecapa-tdnn: Em- phasized channel attention, propagation and aggregation in tdnn based speaker verification,”Interspeech 2020, 2020

    2020

  7. [7]

    Baseline systems for the first spoofing-aware speaker verification challenge: Score and embedding fusion,

    H.-j. Shim, H. Tak, X. Liu, H.-S. Heo, J.-w. Jung, J. S. Chung, S.-W. Chung, H.-J. Yu, B.-J. Lee, M. Todiscoet al., “Baseline systems for the first spoofing-aware speaker verification challenge: Score and embedding fusion,” inProc. Odyssey 2022, 2022, pp. 330–337

    2022

  8. [8]

    Two methods for spoofing-aware speaker verification: Multi-layer perceptron score fusion model and integrated embedding projector,

    J. Heo, J.-H. Kim, and H.-s. Shin, “Two methods for spoofing-aware speaker verification: Multi-layer perceptron score fusion model and integrated embedding projector,” inProc. Interspeech 2022, 2022, pp. 2878–2882

    2022

  9. [9]

    The dku-oppo system for the 2022 spoofing-aware speaker verification challenge,

    X. Wang, X. Qin, Y . Wang, Y . Xu, and M. Li, “The dku-oppo system for the 2022 spoofing-aware speaker verification challenge,” inProc. Interspeech 2022, 2022, pp. 4396–4400

    2022

  10. [10]

    Joint optimiza- tion of speaker and spoof detectors for spoofing-robust automatic speaker verification,

    O. Kurnaz, J. Mishra, T. H. Kinnunen, and C. Hanilc ¸i, “Joint optimiza- tion of speaker and spoof detectors for spoofing-robust automatic speaker verification,”IEEE Transactions on Audio, Speech and Language Pro- cessing, 2026

    2026

  11. [11]

    Integrated spoofing-robust automatic speaker verification via a three-class formulation and llr,

    K. Tan, L. Zhang, R. Zhang, J. Rohdin, L. P. Garc ´ıa-Perera, Z. Cai, S. Khudanpur, M. Wiesner, and N. Andrews, “Integrated spoofing-robust automatic speaker verification via a three-class formulation and llr,” arXiv preprint arXiv:2603.13780, 2026

    arXiv 2026

  12. [12]

    Salmonn: Towards generic hearing abilities for large language models,

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. Ma, and C. Zhang, “Salmonn: Towards generic hearing abilities for large language models,”arXiv preprint arXiv:2310.13289, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Biet al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Sakura: On the multi-hop reasoning of large audio-language models based on speech and audio information,

    C.-K. Yang, N. Ho, Y .-T. Piao, and H.-y. Lee, “Sakura: On the multi-hop reasoning of large audio-language models based on speech and audio information,”arXiv preprint arXiv:2505.13237, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Speechllm-as-judges: Towards general and in- terpretable speech quality evaluation,

    H. Wang, J. Zhao, Y . Yang, S. Liu, J. Chen, Y . Zhang, S. Zhao, J. Li, J. Zhou, H. Sunet al., “Speechllm-as-judges: Towards general and in- terpretable speech quality evaluation,”arXiv preprint arXiv:2510.14664, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Holiantispoof: Audio llm for holistic speech anti-spoofing,

    X. Xu, Y . Ren, L. Liu, W. Wu, B. Li, C. Lu, S. Wang, and C. Zhang, “Holiantispoof: Audio llm for holistic speech anti-spoofing,”arXiv preprint arXiv:2602.04535, 2026

    arXiv 2026

  17. [17]

    Audio language model for deepfake detection grounded in acoustic chain-of-thought,

    R. Chen, Y . Fang, P. Chang, Y . Li, M. Baali, and B. Raj, “Audio language model for deepfake detection grounded in acoustic chain-of-thought,” arXiv preprint arXiv:2603.28021, 2026

    arXiv 2026

  18. [18]

    Can audio large language models verify speaker identity?

    Y . Ren, X. Xu, B. Li, S. Wang, and C. Zhang, “Can audio large language models verify speaker identity?”arXiv preprint arXiv:2509.19755, 2025

    arXiv 2025

  19. [19]

    Speakerllm: A speaker-specialized audio-llm for speaker understanding and verification reasoning,

    K. Nam, J. Heo, S. Bae, H.-J. Yu, and J. S. Chung, “Speakerllm: A speaker-specialized audio-llm for speaker understanding and verification reasoning,”arXiv preprint arXiv:2605.15044, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    Wildspoof: advancing in-the-wild data in text-to-speech generation and spoofing- aware automatic speaker verification,

    Y . Wu, J.-W. Jung, H.-J. Shim, X. Cheng, and X. Wang, “Wildspoof: advancing in-the-wild data in text-to-speech generation and spoofing- aware automatic speaker verification,” inICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026, pp. 21 922–21 924

    2026

  21. [21]

    At-add: All-type audio deepfake detection challenge evaluation plan,

    Y . Xie, H. Cheng, J. Zhou, X. Guo, T. Wang, J. Liu, W. Wang, R. Fu, X. Wang, H. Huanget al., “At-add: All-type audio deepfake detection challenge evaluation plan,”arXiv preprint arXiv:2604.08184, 2026

    Pith/arXiv arXiv 2026

  22. [22]

    Radar challenge 2026: Robust audio deepfake recognition under media trans- formations,

    H.-T. Luong, X. Liu, I. Kukanov, Z. X. Chai, and K. A. Lee, “Radar challenge 2026: Robust audio deepfake recognition under media trans- formations,”arXiv preprint arXiv:2605.09568, 2026

    Pith/arXiv arXiv 2026

  23. [23]

    Overview of esdd2: Environment-aware speech and sound deepfake detection challenge,

    X. Zhang, H. Yin, Y . Xiao, L. Zhang, T. Dang, R. K. Das, and M. Li, “Overview of esdd2: Environment-aware speech and sound deepfake detection challenge,”arXiv preprint arXiv:2606.10791, 2026

    Pith/arXiv arXiv 2026

  24. [24]

    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,” inICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2022, pp. 6367–6371

    2022

  25. [25]

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

    H. Tak, M. Kamble, J. Patino, M. Todisco, and N. Evans, “Rawboost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing,” inICASSP 2022-2022 IEEE In- ternational Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 6382–6386

    2022

  26. [26]

    Samo: Speaker attractor multi-center one-class learning for voice anti-spoofing,

    S. Ding, Y . Zhang, and Z. Duan, “Samo: Speaker attractor multi-center one-class learning for voice anti-spoofing,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    2023

  27. [27]

    End-to-end anti-spoofing with rawnet2,

    H. Tak, J. Patino, M. Todisco, A. Nautsch, N. Evans, and A. Larcher, “End-to-end anti-spoofing with rawnet2,” inICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 6369–6373

    2021

  28. [28]

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

    Pith/arXiv arXiv 2022

  29. [29]

    Alethia: A foun- dational encoder for voice deepfakes,

    Y . Zhu, B. Dwivedi, J. Raghuram, and S. Koppisetti, “Alethia: A foun- dational encoder for voice deepfakes,”arXiv preprint arXiv:2605.00251, 2026

    Pith/arXiv arXiv 2026

  30. [30]

    A subnetwork approach for spoofing aware speaker verification,

    A. Alenin, N. Torgashov, A. Okhotnikov, R. Makarov, and I. Yakovlev, “A subnetwork approach for spoofing aware speaker verification,” in Proc. Interspeech (submitted), 2022

    2022

  31. [31]

    Spoofing- aware speaker verification robust against domain and channel mis- matches,

    C. Zeng, X. Miao, X. Wang, E. Cooper, and J. Yamagishi, “Spoofing- aware speaker verification robust against domain and channel mis- matches,” in2024 IEEE Spoken Language Technology Workshop (SLT). IEEE, 2024, pp. 1150–1157

    2024

  32. [32]

    Spoofceleb: Speech deepfake detection and sasv in the wild,

    J.-w. Jung, Y . Wu, X. Wang, J.-H. Kim, S. Maiti, Y . Matsunaga, H.- j. Shim, J. Tian, N. Evans, J. S. Chunget al., “Spoofceleb: Speech deepfake detection and sasv in the wild,”IEEE Open Journal of Signal Processing, 2025

    2025

  33. [33]

    But systems for wildspoof challenge: Sasv in the wild,

    J. Peng, J. Li, J. Rohdin, L. Zhang, M. Hlav ´aˇcek, and O. Plchot, “But systems for wildspoof challenge: Sasv in the wild,”arXiv preprint arXiv:2512.12851, 2025

    arXiv 2025

  34. [34]

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

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

    2022

  35. [35]

    Scaling up masked audio encoder learning for general audio classification,

    H. Dinkel, Z. Yan, Y . Wang, J. Zhang, Y . Wang, and B. Wang, “Scaling up masked audio encoder learning for general audio classification,”arXiv preprint arXiv:2406.06992, 2024

    Pith/arXiv arXiv 2024

  36. [36]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.” Iclr, vol. 1, no. 2, p. 3, 2022

    2022

  37. [37]

    Qformer: An efficient quaternion transformer for image denoising

    B. Jiang, Y . Lu, G. Lu, and B. Zhang, “Qformer: An efficient quaternion transformer for image denoising.” inIJCAI, 2024, pp. 4237–4245

    2024

  38. [38]

    The llama 3 herd of models,

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughanet al., “The llama 3 herd of models,”arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Allm4add: Unlocking the capabilities of audio large language models for audio deepfake detection,

    H. Gu, J. Yi, C. Wang, J. Tao, Z. Lian, J. He, Y . Ren, Y . Chen, and Z. Wen, “Allm4add: Unlocking the capabilities of audio large language models for audio deepfake detection,” inProceedings of the 33rd ACM International Conference on Multimedia, 2025, pp. 11 736–11 745

    2025

  40. [40]

    SALMONN: Towards generic hearing abilities for large language models,

    C. Tang, W. Yu, G. Sun, X. Chen, T. Tan, W. Li, L. Lu, Z. MA, and C. Zhang, “SALMONN: Towards generic hearing abilities for large language models,” inThe Twelfth International Conference on Learning Representations, 2024. [Online]. Available: https://openreview.net/forum?id=14rn7HpKVk

    2024

  41. [41]

    Qwen2-audio technical report,

    Y . Chu, J. Xu, Q. Yanget al., “Qwen2-audio technical report,”arXiv preprint arXiv:2407.10759, 2024

    Pith/arXiv arXiv 2024

  42. [42]

    ArcFace: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “ArcFace: Additive angular margin loss for deep face recognition,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 4690– 4699

    2019

  43. [43]

    Towards robust speech deepfake detection via human-inspired reasoning,

    A. Dvirniak, E. Kushnir, D. Tarasov, A. Iudin, O. Kiriukhin, M. Pautov, D. Korzh, and O. Y . Rogov, “Towards robust speech deepfake detection via human-inspired reasoning,”arXiv preprint arXiv:2603.10725, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    Colmbo: Speaker language model for descriptive profiling,

    M. Baali, S. Han, S. A. Hannan, P. Samal, K. Singh, S. Deshmukh, R. Singh, and B. Raj, “Colmbo: Speaker language model for descriptive profiling,”arXiv preprint arXiv:2506.09375, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    emotion2vec: Self-supervised pre-training for speech emotion repre- sentation,

    Z. Ma, Z. Zheng, J. Ye, J. Li, Z. Gao, S. Zhang, and X. Chen, “emotion2vec: Self-supervised pre-training for speech emotion repre- sentation,” inFindings of the Association for Computational Linguistics: ACL 2024, 2024, pp. 15 747–15 760

    2024

  46. [46]

    Introducing Parselmouth: A Python interface to Praat,

    Y . Jadoul, B. Thompson, and B. de Boer, “Introducing Parselmouth: A Python interface to Praat,”Journal of Phonetics, vol. 71, pp. 1–15, 2018

    2018

  47. [47]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shaoet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,” 2024. [Online]. Available: https://arxiv.org/abs/2402.03300

    Pith/arXiv arXiv 2024

  48. [48]

    V oxceleb: Large- scale speaker verification in the wild,

    A. Nagrani, J. S. Chung, W. Xie, and A. Zisserman, “V oxceleb: Large- scale speaker verification in the wild,”Computer Science and Language, 2019

    2019

  49. [49]

    Step-audio-r1 technical report,

    F. Tian, X. T. Zhang, Y . Zhang, H. Zhang, Y . Li, D. Liu, Y . Deng, D. Wu, J. Chen, L. Zhaoet al., “Step-audio-r1 technical report,”arXiv preprint arXiv:2511.15848, 2025

    arXiv 2025

  50. [50]

    Intema system description for the ASVspoof5 Challenge: power weighted score fusion,

    A. Aliyev and A. Kondratev, “Intema system description for the ASVspoof5 Challenge: power weighted score fusion,” inThe Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), 2024, pp. 152–157

    2024

  51. [51]

    a-dcf: an architecture agnostic metric with application to spoofing-robust speaker verification,

    H.-j. Shim, J.-w. Jung, T. Kinnunen, N. Evans, J.-F. Bonastre, and I. Lapidot, “a-dcf: an architecture agnostic metric with application to spoofing-robust speaker verification,”arXiv preprint arXiv:2403.01355, 2024

    Pith/arXiv arXiv 2024

  52. [52]

    Ecapa2: A hybrid neural network architecture and training strategy for robust speaker embeddings,

    J. Thienpondt and K. Demuynck, “Ecapa2: A hybrid neural network architecture and training strategy for robust speaker embeddings,” in 2023 IEEE automatic speech recognition and understanding workshop (ASRU). IEEE, 2023, pp. 1–8

    2023

  53. [53]

    Analyzing reasoning shifts in audio deepfake detection under adversarial attacks: The reasoning tax versus shield bifurcation,

    B. Nguyen and T. Le, “Analyzing reasoning shifts in audio deepfake detection under adversarial attacks: The reasoning tax versus shield bifurcation,”arXiv preprint arXiv:2601.03615, 2026

    Pith/arXiv arXiv 2026