Pith. sign in

REVIEW 3 major objections 4 minor 36 references

Time-Frequency Consistency Learning for Robust Speech Deepfake Detection

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read This paper claims that real-time acoustic front-end processing severely degrades speech deepfake detectors, and that a time-frequency consistency training scheme can recover most of the lost accuracy without adding inference cost.

desk verdict Solid, honest robustness study with a genuinely new evaluation angle; the headline numbers are real on its own simulated pipeline, but the transfer to real RTC front-ends is not yet established. read the letter →

arxiv 2607.17761 v2 pith:U3Z6EGA4 submitted 2026-07-20 cs.SD cs.AI

classification cs.SDcs.AI
keywords speechdeepfakedetectionacousticfront-endprocessingrobustnesstime-frequencyconsistencylearningtemporalsoftalignmentfrequencystructuralreal-timecommunicationspoofingrepresentationinvariance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Speech deepfake detectors are usually evaluated on clean or additively noisy audio, but real-time communication platforms pass speech through an acoustic front-end: echo cancellation, noise suppression, automatic gain control, and voice-activity detection. This paper tries to establish that this cascaded processing is a distinct and severe threat to detector accuracy — worse than the sum of its parts — because it disrupts temporal dependencies and distorts frequency structure at the same time. The proposed fix, Time-Frequency Consistency Learning (TFCL), trains the detector on pairs of clean and front-end-processed speech, using attention-based soft temporal alignment and a frequency-domain structural consistency loss so the learned spoof representations stay stable across the pipeline. On a standard spoof-detection benchmark, the training-only scheme lowers the full-pipeline equal-error rate from 17.39% to 9.78% for one backbone and improves several others consistently, with the consistency module removed at inference. If true, deepfake detection can be made robust to real front-ends without extra run-time cost.

What carries the argument

The load-bearing object is the TFCL training module, a dual-branch consistency regularizer attached to any existing front-end feature extractor plus classifier. In the temporal branch, a bidirectional cross-attention soft-alignment computes aligned representations of clean and distorted frames and minimizes cosine distance, capturing non-rigid shifts. In the frequency branch, features are flattened and a linear CKA (centered kernel alignment) score measures second-order structural similarity between the two conditions, with the loss minimizing one minus that score. The two losses are added to the classification cross-entropy with a weight lambda; the whole module exists only during training

What would settle it

Feed the TFCL-trained detector speech that went through a production real-time communication client's native front-end rather than the simulated chain; if the error rate rises far above the simulated 9.78% or no longer tracks the no-TFCL baseline, the AFE-invariance claim would be falsified. A simpler check: swap the module order (for example, run VAD before AGC) in the simulation; if the gains vanish, the model has encoded the specific pipeline rather than a general invariance.

Watch

Extended reading notes

Core claim

The paper's central discovery is that AFE degradation is not a single noise-like corruption but a coupled shift in two representational axes. Time-domain modules such as VAD and AGC create segment-level shifts and non-rigid misalignment, while frequency-domain modules such as noise suppression and echo cancellation weaken or distort spectral structure. TFCL therefore enforces consistency on both axes: a bidirectional cross-attention mechanism softly aligns clean and distorted frame-level features, so the source representation keeps its spoof semantics instead of being forced into frame-wise identity, and a linear CKA loss preserves second-order relationships among frequency-oriented feature

Load-bearing premise

The whole result rests on the simulated echo/noise/gain/VAD chain being a faithful stand-in for real front-ends; if deployed chains use different module order, echo cancellation, or gain behavior, the learned invariance may not transfer to real-world conditions.

Editorial extensions

If this is right

  • If TFCL works as reported, detectors in real-time communication systems can be trained once and remain discriminative behind echo, noise, gain, and VAD processing, with no additional latency or parameters at run time.
  • Training on a mixture of clean and processed audio is not enough; the stage-wise and cascaded results indicate that explicit invariance constraints are what recover accuracy, so future robustness training should include such consistency terms.
  • The gains transfer across multiple backend classifiers, so TFCL functions as a training recipe that can wrap existing detectors rather than requiring a new architecture.
  • Because the consistency module is removed at inference, the robustness improvement is essentially free for deployment, and it should compose with other input augmentation already used in the clean branch.

Reading between the lines

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

  • If the same decomposition holds for codec compression and packet loss, which also mix timing and spectral changes, the temporal-soft-alignment plus frequency-structural-consistency pair could be reused as a general robustness recipe; a direct test is to train with clean/codec pairs and measure stage-wise error.
  • The dual-input clean/distorted setup suggests a self-supervised variant: use a single AFE-degraded sample and let a learned module estimate the clean representation, which would remove the need for paired clean data at deployment.
  • Reporting stage-wise and cascaded results (echo, AEC, NS, AGC, VAD separately) may become the norm for robustness evaluation, since the paper shows that end-to-end numbers hide which front-end stage destroys which cues.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper studies robustness of speech deepfake detection (SDD) under acoustic front-end (AFE) processing, simulating a pipeline of echo, AEC, noise/NS, AGC, and VAD. It reports that existing SDD models degrade substantially under this pipeline and proposes Time-Frequency Consistency Learning (TFCL), a training-time method that enforces temporal consistency via bidirectional cross-attention (Eqs. 1-3) and frequency-domain structural consistency via linear CKA (Eqs. 4-7), jointly with classification loss (Eq. 8). Experiments on ASVspoof2019 LA show that TFCL improves over a matched 'Mix' baseline and across three backbones, with no added inference cost. The paper claims this improves SDD robustness in real-world RTC scenarios.

Significance. The contribution is potentially valuable: it broadens SDD robustness evaluation beyond additive noise to a cascaded AFE pipeline, and the proposed consistency training is simple, architecture-agnostic, and inference-free. The main strength is the controlled ablation in Table 2, where TFCL is compared against the same backbone trained on the same mixed clean+distorted data without TFCL, showing consistent gains (e.g., VAD EER 9.78 vs 17.39 for XLSR-AASIST). The code is promised publicly. However, the validity of the central 'real-world robustness' claim depends on the faithfulness of the AFE simulation, especially the AEC stage, and the current manuscript does not provide enough detail to establish that faithfulness.

major comments (3)
  1. [4.1, Eqs. (9)-(10)] The AEC simulation is under-specified and the equations suggest a likely misconfiguration. Eq. (10) defines x_mic(t)=x(t)+β(x*h)(t-τ), i.e., the same clean utterance x is both the desired near-end speech and the source of the echo. WebRTC AEC requires a separate far-end render reference (the remote playback signal). If the reference supplied to the AEC is x itself, the canceller will treat part of the desired signal as echo, explaining the observed over-suppression, extended silences, and residual artifacts. TFCL may then learn invariance to this particular misconfiguration rather than to a realistic RTC AEC. Please specify exactly what far-end reference was used, or re-run the pipeline with a distinct far-end speech/playback signal and report the resulting degradation and TFCL gains.
  2. [Table 1 vs Table 2] The headline comparison in Table 1 is not apples-to-apples: 'Ours' is trained on clean+distorted mixed data with TFCL, while the listed baselines are largely public pretrained models trained on clean data. The paper should not present Table 1 as evidence of the method's contribution without the matched 'w/o TFCL' comparison. The controlled comparisons in Table 2 (Mix vs Ours, and w/ TFCL vs w/o TFCL) are the appropriate evidence and should be foregrounded in the narrative.
  3. [3.6, Eq. (4)] The frequency-oriented transformation φ is never defined. Eq. (4) says the frame features are reorganized via φ, but no architecture or operation is given. It is also not explained why linear CKA applied to the flattened feature matrix (with dimensions F x T) specifically enforces 'frequency-domain structural consistency' rather than a generic second-order statistic. Please define φ and justify the frequency-domain interpretation, or the ablation 'w/o FSCL' is the only evidence tying the loss to the claimed mechanism.
minor comments (4)
  1. [Table 1] There is an extraneous block of digits ('2 .31 .31 .05 ...') after the table notes. This appears to be a copy-paste artifact and should be removed.
  2. [Figures 1, 4, 5] Minor grammar/presentation: 'Figs. 1' should be 'Fig. 1', 'Figs. 5' should be 'Fig. 5', and 'a ecn sa gcv ad' in Fig. 4 is garbled; axis labels need cleanup.
  3. [Notation] The model name is inconsistently written as 'XLSR_AASIST' and 'XLSR-AASIST' (e.g., Table 1 vs Table 2). Please standardize.
  4. [4.2] Training details omit random seeds and the number of runs; given the large variance often seen in anti-spoofing results, a single run makes it hard to assess whether the reported EER differences are statistically meaningful.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: TFCL's consistency losses are training objectives evaluated on held-out AFE conditions; self-citations are not load-bearing.

full rationale

The paper's central claim is that adding TFCL to SDD backbones reduces AFE-induced degradation (e.g., full-pipeline EER 9.78 vs 17.39 for the same XLSR-AASIST backbone, Table 2). This is an empirical comparison, not a derivation: the temporal and frequency losses (Eqs. 2-3 and 6-7) are explicit regularizers added to the classification loss (Eq. 8), and the reported gains are measured on the ASVspoof2019 LA evaluation partition, where echo and noise simulation differ from training (RIR/DNS for evaluation vs pyroomacoustics/MUSAN for training). No equation in the paper makes the test metric equal to a fitted parameter or to the consistency loss itself. The hyperparameter lambda is selected using Fig. 6, which is a tuning step that may be optimistic if performed on evaluation-adjacent conditions, but it is not a fitted input renamed as a prediction. The self-citations [30] and [31] appear in motivational and related-work contexts (RTC threats and wav2vec 2.0 adoption); they do not supply the invariance mechanism or the evaluation results, so they are not load-bearing. The paper's own Conclusion limitation that AFE is modeled in isolation from codecs, packet loss, and bandwidth constraints, and the AEC simulation's unspecified far-end reference in Section 4.1, are realism/transfer-fidelity risks rather than circularity. Overall, the TFCL contribution has independent content and the score is at the low end only because of minor self-citations and tuning choices, not because the result reduces to its inputs.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The framework rests on two engineered inputs: the consistency weight λ in Eq. (8) and the simulated AFE pipeline configuration. λ=0.3 is the best-performing point in a sweep across the same evaluation conditions (Fig. 6), so it carries selection-on-eval-set information. The pipeline (WebRTC modules for AEC/NS/AGC/VAD, pyroomacoustics echoes for train/dev, RIR echoes for eval, MUSAN noise at 5–20 dB for train/dev, DNS noise for eval) defines the distribution shift the method is trained and tested on; its realism is the key external assumption.

free parameters (2)
  • λ (consistency loss weight) = 0.3
    Balances L_t + L_f against cross-entropy in Eq. (8); selected via a sweep over AFE evaluation conditions (Fig. 6, §4.6), so it is fit to the test-adjacent distribution.
  • Train/dev SNR sampling range = 5–20 dB
    Noise-level range in §4.1 for train/dev AFE simulation; hand-chosen to define the difficulty of the simulated pipeline.
assumptions (5)
  • domain assumption The WebRTC APM modules (AEC→NS→AGC→VAD) represent the de facto standard acoustic front-end in real-time communication systems.
    Invoked in §3.2 to justify the simulated pipeline; the realism of the benchmark depends on it.
  • domain assumption ASVspoof2019 LA evaluation set is a valid SDD benchmark with unseen spoofing algorithms and speakers.
    Used as the sole evaluation dataset in §4.1; the central numbers inherit its conventions and limitations.
  • ad hoc to paper AFE distortion decomposes into temporal dependency disruption and frequency structural distortion.
    Introduced in §3.3 from spectrogram observation; motivates the two separate loss branches and is not independently proven.
  • ad hoc to paper Enforcing representation consistency between clean and AFE-distorted speech preserves spoof-discriminative semantics.
    Core premise of §3.1; the method assumes that invariant features retain the cues needed for deepfake detection.
  • domain assumption Linear CKA is an adequate surrogate for second-order frequency-domain structural similarity.
    Used in §3.6 to define the frequency consistency loss; standard in representation analysis but a modeling choice here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Time-Frequency Consistency Learning for Robust Speech Deepfake Detection." pith.science (2026). https://pith.science/paper/U3Z6EGA4

@misc{pith2026260717761,
  author       = {Pith},
  title        = {Pith review of: Time-Frequency Consistency Learning for Robust Speech Deepfake Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U3Z6EGA4}},
  note         = {Machine review of arXiv:2607.17761}
}
read the original abstract

Recently, speech deepfake detection (SDD) has achieved significant progress. However, its robustness evaluation remains largely confined to controlled additive noise scenarios, lacking systematic investigation of the complex distortions introduced by acoustic front-end (AFE) processing pipelines in real-world deployments. In this work, we simulate a unified AFE pipeline comprising acoustic echo cancellation, noise suppression, automatic gain control, and voice activity detection (VAD), and conduct a comprehensive evaluation of current state-of-the-art models. The results show that the nonlinear and time-frequency coupled distortions introduced by AFE significantly degrade detection performance. To address this issue, we propose a Time-Frequency Consistency Learning (TFCL) framework, which aims to learn invariant spoofing representations that remain stable before and after AFE processing. We observe that AFE not only introduces temporal misalignment (e.g., segment-level shifts caused by VAD), but also weakens or distorts critical frequency-domain cues. To this end, TFCL employs an attention-driven soft alignment mechanism to capture cross-temporal dependencies, along with frequency-domain structural consistency constraints to enforce feature invariance. As a result, the model is able to maintain stable representations under both temporal perturbations and spectral distortions. Extensive experimental results demonstrate that the proposed method effectively mitigates the performance degradation caused by AFE processing, significantly improving the robustness of SDD in real-world scenarios. The code is available at https://github.com/JunXue-tech/TFCL.

Figures

Figures reproduced from arXiv: 2607.17761 by the authors.

Figure 1
Figure 1. Spectrogram visualization of the utterance [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework. The distorted speech is generated by passing clean speech through the AFE [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. General acoustic front-end processing pipeline [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance Comparison of Different Models under [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: t-SNE [26] visualization of feature distributions across different AFE stages, with XLSR+AASIST as the baseline model. 0.1 0.3 0.5 0.7 0.9 0 3 6 12 EER (%) clean echo aec noisy ns agc vad [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Results of hyperparameter 𝜆 under various AFE conditions. intensified, these structures rapidly deform and undergo notice￾able distribution shifts. The feature clusters become increasingly stretched and even overlap, leading to a significant degradation in class separa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 4 linked inside Pith

  1. [1]

    Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, et al

  2. [2]

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems33 (2020), 12449–12460

  3. [3]

    Muhammad Usman Bashir. [n. d.].WebRTC Audio Processing Overview. https: //github.com/mail2chromium/Android-Audio-Processing-Using-WebRTC

  4. [4]

    Jialu Cao, Hui Tian, Peng Tian, Haizhou Li, and Jianzong Wang. 2025. Robust De- tection of Partially Spoofed Audio Using Semantic-Aware Inconsistency Learning. IEEE Transactions on Audio, Speech and Language Processing(2025)

  5. [5]

    2025.Company Finance Director Nearly Loses Over US$499,000 to Scammers Using Deepfake to Impersonate CEO

    Channel News Asia. 2025.Company Finance Director Nearly Loses Over US$499,000 to Scammers Using Deepfake to Impersonate CEO. https://www.channelnewsasia.com/singapore/deepfake-scam-impersonate- ceo-company-finance-director-5048706 Accessed: 2025-11-6

  6. [6]

    Qixian Chen, Yuxiong Xu, Sara Mandelli, Sheng Li, and Bin Li. 2025. Adaptive Mixture of Low-Rank Experts for Robust Audio Spoofing Detection.IEEE Signal Processing Letters(2025)

  7. [7]

    Xuanjun Chen, I-Ming Lin, Lin Zhang, Jiawei Du, Haibin Wu, Hung yi Lee, and Jyh-Shing Roger Jang. 2025. Codec-Based Deepfake Source Tracing via Neural Audio Codec Taxonomy. InInterspeech 2025. 1538–1542. doi:10.21437/ Interspeech.2025-1297

  8. [8]

    Yujie Chen, Jiangyan Yi, Jun Xue, Chenglong Wang, Xiaohui Zhang, Shunbo Dong, Siding Zeng, Jianhua Tao, Zhao Lv, and Cunhang Fan. 2024. RawBMamba: End-to-End Bidirectional State Space Model for Audio Deepfake Detection. In Proc. Interspeech 2024. 2720–2724

Show all 36 references
  1. [9]

    Zhihao Du, Qian Chen, Shiliang Zhang, Kai Hu, Heng Lu, Yexin Yang, Hangrui Hu, Siqi Zheng, Yue Gu, Ziyang Ma, et al. 2024. Cosyvoice: A scalable multilingual zero-shot text-to-speech synthesizer based on supervised semantic tokens.arXiv preprint arXiv:2407.05407(2024)

  2. [10]

    Cunhang Fan, Mingming Ding, Jianhua Tao, Ruibo Fu, Jiangyan Yi, Zhengqi Wen, and Zhao Lv. 2024. Dual-branch knowledge distillation for noise-robust syn- thetic speech detection.IEEE/ACM Transactions on Audio, Speech, and Language Processing32 (2024), 2453–2466

  3. [11]

    Eduardo Fonseca, Jordi Pons Puig, Xavier Favory, Frederic Font Corbera, Dmitry Bogdanov, Andres Ferraro, Sergio Oramas, Alastair Porter, and Xavier Serra. 2017. Freesound datasets: a platform for the creation of open audio datasets. InHu X, Cunningham SJ, Turnbull D, Duan Z, e...

  4. [12]

    Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. 2017. Audio set: An ontology and human-labeled dataset for audio events. In2017 IEEE international conference on acoustics, speech and signal process...

  5. [13]

    Google. [n. d.].WebRTC Audio Processing Module. https://webrtc. googlesource.com/src/+/refs/heads/main/modules/audio_processing/g3doc/ audio_processing_module.md

  6. [14]

    Hao Gu, Jiangyan Yi, Chenglong Wang, Jianhua Tao, Zheng Lian, Jiayi He, Yong Ren, Yujie Chen, and Zhengqi Wen. 2025. Allm4add: Unlocking the capabilities of audio large language models for audio deepfake detection. InProceedings of the 33rd ACM International Conference on Mult...

  7. [15]

    Jee-weon Jung, Hee-Soo Heo, Hemlata Tak, Hye-jin Shim, Joon Son Chung, Bong- Jin Lee, Ha-Jin Yu, and Nicholas Evans. 2022. Aasist: Audio anti-spoofing using integrated spectro-temporal graph attention networks. InICASSP 2022-2022 IEEE international conference on acoustics, spe...

  8. [16]

    Tianchi Liu, Duc-Tuan Truong, Rohan Kumar Das, Kong Aik Lee, and Haizhou Li. 2025. Nes2net: A lightweight nested architecture for foundation model driven speech anti-spoofing.IEEE Transactions on Information Forensics and Security20 (2025), 12005–12018

  9. [17]

    Xuechen Liu, Xin Wang, Md Sahidullah, Jose Patino, Héctor Delgado, Tomi Kinnunen, Massimiliano Todisco, Junichi Yamagishi, Nicholas Evans, Andreas Nautsch, et al . 2023. Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild.IEEE/ACM Transactions on Audio, Sp...

  10. [18]

    Robin Scheibler, Eric Bezzam, and Ivan Dokmanić. 2018. Pyroomacoustics: A python package for audio room simulation and array processing algorithms. In2018 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 351–355

  11. [19]

    Udayon Sen, Alka Luqman, and Anupam Chattopadhyay. 2025. Toward Noise- Aware Audio Deepfake Detection: Survey, SNR-Benchmarks, and Practical Recipes.arXiv preprint arXiv:2512.13744(2025)

  12. [20]

    David Snyder, Guoguo Chen, and Daniel Povey. 2015. Musan: A music, speech, and noise corpus.arXiv preprint arXiv:1510.08484(2015)

  13. [21]

    Hemlata Tak, Madhu Kamble, Jose Patino, Massimiliano Todisco, and Nicholas Evans. 2022. Rawboost: A raw data boosting and augmentation method applied to automatic speaker verification anti-spoofing. InICASSP 2022-2022 IEEE Inter- national Conference on Acoustics, Speech and Si...

  14. [22]

    Hemlata Tak, Massimiliano Todisco, Xin Wang, Jee-weon Jung, Junichi Yamagishi, and Nicholas Evans. 2022. Automatic Speaker Verification Spoofing and Deepfake Detection Using Wav2vec 2.0 and Data Augmentation. InProc. Odyssey 2022. 112–119

  15. [23]

    Massimiliano Todisco, Xin Wang, Ville Vestman, Md Sahidullah, Hector Del- gado, Andreas Nautsch, Junichi Yamagishi, Nicholas Evans, Tomi Kinnunen, and Kong Aik Lee. 2019. ASVspoof 2019: Future Horizons in Spoofed and Fake Audio Detection. InInterspeech 2019. International Spee...

  16. [24]

    Hoan My Tran, Damien Lolive, Aghilas Sini, Arnaud Delhay, Pierre-François Marteau, and David Guennec. 2025. Multi-level SSL feature gating for audio deepfake detection. InProceedings of the 33rd ACM International Conference on Multimedia. 11766–11775

  17. [25]

    Duc-Tuan Truong, Ruijie Tao, Tuan Nguyen, Hieu-Thi Luong, Kong Aik Lee, and Eng Siong Chng. 2024. Temporal-Channel Modeling in Multi-head Self-Attention for Synthetic Speech Detection. InProc. Interspeech 2024. 537–541

  18. [26]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research9, 11 (2008)

  19. [27]

    Xin Wang, Héctor Delgado, Hemlata Tak, Jee-Weon Jung, Hye-Jin Shim, Massi- miliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidullah, Tomi Kinnunen, et al. 2024. ASVspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale. InThe Automatic Speaker Verifica...

  20. [28]

    Ziteng Wang, Yueyue Na, Biao Tian, and Qiang Fu. 2022. NN3A: Neural network supported acoustic echo cancellation, noise suppression and automatic gain control for real-time communications. InICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Process...

  21. [29]

    Yuankun Xie, Yi Lu, Ruibo Fu, Zhengqi Wen, Zhiyong Wang, Jianhua Tao, Xin Qi, Xiaopeng Wang, Yukun Liu, Haonan Cheng, et al . 2025. The codecfake dataset and countermeasures for the universally detection of deepfake audio. IEEE Transactions on Audio, Speech and Language Proces...

  22. [30]

    Jun Xue, Zhuolin Yi, Yihuan Huang, Yanzhen Ren, Yujie Chen, Cunhang Fan, Zicheng Su, Yongcheng Zhang, and Bo Cai. 2026. Rtcfake: Speech deepfake detec- tion in real-time communication. InFindings of the Association for Computational Linguistics: ACL 2026. 5763–5775

  23. [31]

    Jun Xue, Tong Zhang, Zhuolin Yi, Yihuan Huang, Yi Chai, Yiyang Zhang, and Yanzhen Ren. 2026. Profiling the Voice: Speaker-Specific Phoneme Fingerprinting for Speech Deepfake Detection.arXiv preprint arXiv:2605.17737(2026)

  24. [32]

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

  25. [33]

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

  26. [34]

    Qishan Zhang, Shuangbing Wen, and Tao Hu. 2024. Audio deepfake detection with self-supervised xls-r and sls classifier. InProceedings of the 32nd ACM Inter- national Conference on Multimedia. 6765–6773

  27. [35]

    Yixuan Zhou, Guoyang Zeng, Xin Liu, Xiang Li, Renjie Yu, Ziyang Wang, Runchuan Ye, Weiyue Sun, Jiancheng Gui, Kehan Li, et al . 2025. VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning.arXiv preprint arXiv:2509.24650(2025)

  28. [2022]

    XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale. InProc. Interspeech 2022. 2278–2282

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.