Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

Towards Neural Audio Codec Source Parsing

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper establishes that codecfake audio can be attributed to its generating neural audio codec by regressing the codec's numeric parameters (quantizers, bits per second, sampling rate), and that a hyperbolic multi-subspace network…

desk verdict Genuinely new task framing, but the paper's own Table 2 contradicts the central consistency claim and the 'trace back to the exact NAC' promise is never tested. read the letter →

arxiv 2506.12627 v1 pith:JZ4DKYDZ submitted 2025-06-14 eess.AS cs.SD

Orchid Chetia Phukan , Girish , Mohd Mujtaba Akhtar , Arun Balaji Buduru , Rajesh Sharma This is my paper
classification eess.AScs.SD
keywords audiodeepfakedetectioncodecfakesourceattributionneuralcodecstructuredregressionhyperbolicgeometrymulti-tasklearningforensic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces Neural Audio Codec Source Parsing (NACSP), a new way to attribute codecfake audio by predicting the numeric parameters of the neural audio codec that generated it, rather than classifying which codec produced it. The authors argue that open-set attribution, which can only label audio as unknown, cannot describe unseen codecs, so they recast the problem as multi-task regression over three parameters: number of quantizers, bits per second, and sampling rate. To make this work, they propose HYDRA, a framework that maps pretrained speech representations into several hyperbolic subspaces with task-specific attention, and they report that HYDRA outperforms Euclidean-space baselines and prior source-attribution systems on the ST-codecfake and CodecFake benchmarks, in both closed-set and open-set settings. The ultimate promise is that estimated parameters allow tracing audio back to the exact codec that generated it.

What carries the argument

The load-bearing mechanism is HYDRA, which projects frozen pretrained-model representations into K separate Poincaré ball subspaces, each with its own learned curvature, via the exponential map; for each target parameter it aggregates these subspaces with learned attention weights using Möbius addition and scalar multiplication, and penalizes pairwise mutual information between subspace latents through a hyperbolic total correlation loss. This task-specific hyperbolic disentanglement is what the paper credits for separating the acoustic cues relevant to each codec parameter and for the consistent gains over Euclidean baselines.

What would settle it

Test a held-out codec that shares its (Q, BPS, SR) triple with a training codec but uses a different architecture; if the predicted parameters cannot distinguish the two, the triple does not uniquely identify the generating codec, and NACSP's tracing promise collapses. Alternatively, convert predicted parameters into a codec identity and measure attribution accuracy on open-set codecs; chance-level accuracy would show that low regression error does not translate into source identification.

Watch

Extended reading notes

Core claim

The central claim is that the limitation of open-set codecfake attribution, namely that unseen codecs are lumped together as unknown, can be overcome by treating the source-identification problem as structured regression over codec parameters (quantizers Q, bits per second BPS, sampling rate SR). The paper establishes the first benchmark for this task and shows that a hyperbolic multi-subspace network, HYDRA, consistently achieves lower RMSE and MAE than Euclidean baseline models on both closed-set and open-set evaluations across eight pretrained speech models. The paper further claims that no single pretrained model dominates, so the architectural choice matters more than the choice of feature extractor.

Load-bearing premise

The argument assumes that the three codec parameters, number of quantizers, bits per second, and sampling rate, can be recovered from audio in a way that transfers to unseen codecs, and that different codecs have distinguishable parameter combinations.

Editorial extensions

If this is right

  • Open-set attribution no longer needs an explicit unknown class; unseen codecs can be described by their predicted (Q, BPS, SR) instead of being discarded as generic unknowns.
  • The same framework handles in-distribution and open-set attribution in one model, removing the two-stage detection-plus-classification pipeline used by earlier methods.
  • Because the parameters are continuous, NACSP can resolve fine-grained variants within a codec family (different bandwidth or quantizer settings) that discrete classification would collapse.
  • The reported invariance across pretrained models suggests that downstream architecture, not the choice of PTM, drives NACSP performance.
  • If parameter estimation is accurate enough, forensic tools could in principle reconstruct the generating codec's configuration directly from a fake audio clip.

Reading between the lines

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

  • Beyond the paper: converting predicted parameter triples back into codec identities would test whether low regression error actually yields attribution; the paper does not report that step.
  • Beyond the paper: the same parameter-regression logic could extend to other generative components (vocoder type, tokenizer settings) in spoofed speech synthesis, not just codecs.
  • Beyond the paper: because HYDRA's gains appear to come from geometry rather than a specific pretrained model, the approach could transfer to music, singing-voice, or visual codecfake forensics.
  • Beyond the paper: open-set gains may mostly interpolate within a small finite parameter grid; testing with codec parameter values outside the training range would reveal true extrapolation.
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

5 major / 4 minor

Summary. The manuscript introduces Neural Audio Codec Source Parsing (NACSP), which reframes codec-fake source attribution as multi-task regression over three generative codec parameters: number of quantizers (Q), bits per second (BPS), and sampling rate (SR). The authors benchmark eight frozen speech pre-trained models (PTMs) on ST-Codecfake (closed-set and open-set) and CodecFake, using a CNN-based Euclidean baseline, a hyperbolic baseline, and their proposed HYDRA architecture. HYDRA projects PTM features into task-specific Poincaré balls with learned curvatures, aggregates the subspaces with task attention, and adds a hyperbolic total-correlation (HTC) loss for disentanglement. The paper claims that HYDRA consistently outperforms Euclidean baselines and sets a new state of the art, and that estimating these parameters makes it possible to trace an audio back to the exact NAC that generated it.

Significance. The NACSP formulation is a timely and potentially useful reframing of open-set audio source attribution: predicting interpretable codec parameters instead of discrete class labels is a natural step toward forensic interpretability, and the first PTM benchmark for this task has value for the community. The evaluation design is not circular: the regression targets are external dataset labels, and the ST-Codecfake open-set split contains held-out codecs. HYDRA's idea of per-task hyperbolic subspaces with attention is interesting. However, the paper's central comparative claim is contradicted by its own Table 2, and the paper's forensic promise of exact codec tracing is never actually tested. The method section also omits details needed for reproducibility. If these issues are fixed, the contribution could be significant; as written, the evidence does not support the claimed state of the art.

major comments (5)
  1. [Table 2 (CodecFake closed-set)] The BPS column directly contradicts the Abstract's claim that HYDRA 'consistently outperforms' Euclidean baselines. For every PTM in Table 2, the Euclidean baseline has lower BPS RMSE and MAE than HYDRA; for example, with x-vector the baseline gives 8.85/7.34 while HYDRA gives 11.09/9.02, and with WavLM the baseline gives 9.81/8.43 while HYDRA gives 11.98/9.77. Since CodecFake is one of only two datasets and BPS is one of three tasks, the claim of consistent superiority fails as stated. The authors should report error bars and significance tests, and should either revise the claim to be dataset- and task-specific or explain why BPS is excluded from the consistency claim.
  2. [Conclusion and Evaluation (Tables 1-3)] The Conclusion states that 'by estimating the parameters, we can trace back to the exact NAC that generated a particular audio,' but no experiment in the paper tests this mapping. Tables 1-3 report only RMSE/MAE on Q, BPS, and SR; there is no closed-set codec-identification accuracy, no experiment that converts predicted parameters into a codec identity, and no identification of which of the five open-set codecs in ST-Codecfake was generated. Moreover, the regression target is not shown to be unique across codec families: different NACs can share identical (Q, BPS, SR) configurations, in which case exact tracing is impossible by construction. The paper should either provide a codec-identity evaluation (for example, nearest-configuration matching) or explicitly restrict the claim to parameter estimation rather than exact source identification.
  3. [Section Modeling (HYDRA)] The attention mechanism is a named contribution, but the attention weights α_t(k) are never defined. The text only states that the weights are non-negative and sum to one over subspaces, and then uses them in the Möbius combination. No scoring function, softmax normalization, or task-query construction is given. Without an equation describing how α_t(k) is computed from the PTM representation or the hyperbolic subspaces, the method is not reproducible as written and the claimed benefit of task-specific attention cannot be assessed.
  4. [Section Modeling (HTC loss)] The hyperbolic total-correlation loss is defined as a sum of KL divergences between joint and marginal densities, but the paper does not specify how p(z_i, z_j) and p(z_i)p(z_j) are estimated (e.g., histogram binning, kernel density estimation, or a variational bound), nor what batch size, number of bins, or weighting is used relative to the task MSE losses. In addition, the ablation described in Section Experiments removes the attention modules and the HTC loss together with the task-specific subspaces, so there is no isolated ablation showing that HTC contributes to the result. This is a load-bearing component for the claimed disentanglement and needs to be specified and tested.
  5. [Tables 1-4] No error bars, confidence intervals, or significance tests are reported anywhere. The text itself says that PTM differences are 'neither statistically significant nor consistent' across datasets, but no statistical test is performed to support this statement. Given that some HYDRA-vs-baseline margins are small and that Table 2 contains a case where the baseline is better, the state-of-the-art claim is not supported without uncertainty quantification or repeated-seed experiments.
minor comments (4)
  1. [Throughout] Cross-references are incomplete: the text refers to 'Section .' and 'Figure 3' without giving the equations or architecture details in the text, and equations are not numbered. This makes it hard to follow the method.
  2. [Abstract and Introduction] There are several typos and grammatical errors, including 'evalaution', 'an need of the hour', 'baseline models that dwells in euclidean space', and 'generated form'. These should be corrected in revision.
  3. [Table 4 and SOTA comparison] The reimplemented SOTA baselines (MiO and AASIST) are not described in enough detail to confirm that the same training setup, epochs, inputs, and hyperparameters were used. Please provide the reimplementation settings.
  4. [Section Pre-trained Models] For Whisper, the text says 'the encoder outputs are used' but does not specify which encoder layer or whether the features are pooled over time before the downstream network; this should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: NACSP is supervised regression against externally defined codec parameters, HYDRA's components are regularizers trained jointly, and self-citations are not load-bearing.

full rationale

The derivation chain is self-contained. NACSP labels (Q, BPS, SR) are properties of the generating codec assigned from the dataset metadata, not derived from the model's own outputs; training a regressor on these labels and measuring RMSE/MAE on held-out closed-set and open-set codec configurations (Tables 1-3) is a standard external generalization test. HYDRA's hyperbolic projections, task-specific attention, and HTC loss are learned transformations and a mutual-information regularizer, not quantities that encode the target labels by construction, so the reported gains over Euclidean baselines are empirical rather than definitional. Self-citations to Chetia Phukan et al. (2024) appear as related-work motivation and as one of the reimplemented SOTA comparison systems (Table 4, 'M'); they do not supply the load-bearing premise that codec parameters are recoverable or that the triple uniquely identifies a codec. The conclusion's claim that 'by estimating the parameters, we can trace back to the exact NAC' is not directly validated (no experiment converts predicted parameters into a codec identity, and codecs sharing identical triples would break the mapping), but that is an evidentiary gap in the paper's forensic claim, not a circular reduction of the regression to its inputs. No equation or fitted value is reused as its own prediction, and no uniqueness theorem is imported from the authors' prior work. Therefore no circular step can be exhibited under the quoting rule.

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

The central claim depends on the empirical premise that codec parameters are recoverable from audio and that they are sufficient for codec identity, plus several unstated training hyperparameters (HTC weight, dropout, curvature init). The math used is standard hyperbolic geometry.

free parameters (4)
  • HTC loss weight = not reported
    The paper proposes the hyperbolic total correlation loss as a disentanglement penalty but never states the total objective or the weight used to combine it with the MSE terms. This weight controls the strength of disentanglement and is a free hyperparameter.
  • Curvature c_k initialization = not reported (learned)
    Each hyperbolic subspace has a learned curvature c_k but no initialization or range is given; the learned curvature values are fitted to the training data and affect the geometry of the subspaces.
  • Dropout rate = not reported
    Dropout is mentioned for overfitting prevention but the rate is not specified; it affects generalization and results.
  • Task loss weighting = equal (assumed, not stated)
    The three regression losses (Q, BPS, SR) are summed without any stated weighting; if a different weighting were used, results would change.
assumptions (4)
  • domain assumption The three NAC parameters (Q, BPS, SR) are sufficient to characterize a neural audio codec for attribution purposes.
    Introduced in the Introduction and Conclusion, where the paper claims that estimating these parameters lets one 'trace back to the exact NAC'. No evidence is provided that this mapping is unique or sufficient.
  • domain assumption Audio traces of these parameters generalize from seen to unseen codecs.
    The open-set evaluation on ST-Codecfake assumes that regression trained on closed-set codecs transfers to novel codecs; the paper reports RMSE but does not test whether predicted parameters actually identify the unseen codec.
  • ad hoc to paper Hyperbolic spaces with learned curvature are better suited than Euclidean spaces for disentangling PTM representations in this task.
    Motivated by hierarchy in the HYDRA section, but no theoretical or empirical isolation of this assumption; the ablation (single hyperbolic space) is worse than task-specific subspaces but the comparison is confounded with attention and HTC loss.
  • standard math Poincaré exponential/log maps and Möbius operations as defined in the HYDRA section are the correct geometry-preserving operations.
    Standard definitions from hyperbolic geometry, used without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Neural Audio Codec Source Parsing." pith.science (2026). https://pith.science/paper/JZ4DKYDZ

@misc{pith2026250612627,
  author       = {Pith},
  title        = {Pith review of: Towards Neural Audio Codec Source Parsing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JZ4DKYDZ}},
  note         = {Machine review of arXiv:2506.12627}
}
read the original abstract

A new class of audio deepfakes-codecfakes (CFs)-has recently caught attention, synthesized by Audio Language Models that leverage neural audio codecs (NACs) in the backend. In response, the community has introduced dedicated benchmarks and tailored detection strategies. As the field advances, efforts have moved beyond binary detection toward source attribution, including open-set attribution, which aims to identify the NAC responsible for generation and flag novel, unseen ones during inference. This shift toward source attribution improves forensic interpretability and accountability. However, open-set attribution remains fundamentally limited: while it can detect that a NAC is unfamiliar, it cannot characterize or identify individual unseen codecs. It treats such inputs as generic ``unknowns'', lacking insight into their internal configuration. This leads to major shortcomings: limited generalization to new NACs and inability to resolve fine-grained variations within NAC families. To address these gaps, we propose Neural Audio Codec Source Parsing (NACSP) - a paradigm shift that reframes source attribution for CFs as structured regression over generative NAC parameters such as quantizers, bandwidth, and sampling rate. We formulate NACSP as a multi-task regression task for predicting these NAC parameters and establish the first comprehensive benchmark using various state-of-the-art speech pre-trained models (PTMs). To this end, we propose HYDRA, a novel framework that leverages hyperbolic geometry to disentangle complex latent properties from PTM representations. By employing task-specific attention over multiple curvature-aware hyperbolic subspaces, HYDRA enables superior multi-task generalization. Our extensive experiments show HYDRA achieves top results on benchmark CFs datasets compared to baselines operating in Euclidean space.

Figures

Figures reproduced from arXiv: 2506.12627 by the authors.

Figure 1
Figure 1. Demonstration of NACSP; Fake U represents a un [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. HYDRA; FCN stands for fully connected network HYDRA In this section, we introduce our proposed framework, HYDRA, for NACSP, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Baseline (Hyperbolic); FCN stands for fully con [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Towards Generalized Source Tracing for Codec-Based Deepfake Speech

    cs.SD 2025-06 conditional novelty 5.0 of 10

    SASTNet, which fuses Whisper semantic features with Wav2Vec2 and AudioMAE acoustic features, improves source tracing for codec-based deepfake speech on CodecFake+, while exposing that prior models overfit to silence.

Reference graph

Works this paper leans on

38 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Babu, A.; Wang, C.; Tjandra, A.; Lakhotia, K.; Xu, Q.; Goyal, N.; Singh, K.; von Platen , P.; Saraf, Y.; Pino, J.; Baevski, A.; Conneau, A.; and Auli, M. 2022. XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale. In Interspeech 2022, 2278--2282

    work page 2022

  4. [4]

    Baevski, A.; Zhou, Y.; Mohamed, A.; and Auli, M. 2020. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33: 12449--12460

    2020

  5. [5]

    Bhagtani, K.; Yadav, A. K. S.; Bestagini, P.; and Delp, E. J. 2024. Attribution of Diffusion Based Deepfake Speech Generators. In 2024 IEEE International Workshop on Information Forensics and Security (WIFS), 1--6. IEEE

    work page 2024

  6. [6]

    Chen, S.; Wang, C.; Chen, Z.; Wu, Y.; Liu, S.; Chen, Z.; Li, J.; Kanda, N.; Yoshioka, T.; Xiao, X.; et al. 2022. Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 16(6): 1505--1518

    2022

  7. [7]

    Chen, S.; Wu, Y.; Wang, C.; Chen, Z.; Chen, Z.; Liu, S.; Wu, J.; Qian, Y.; Wei, F.; Li, J.; and Yu, X. 2021. Unispeech-Sat: Universal Speech Representation Learning With Speaker Aware Pre-Training. ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 6152--6156

    work page 2021

  8. [8]

    Chetia Phukan , O.; Balaji Buduru , A.; and Sharma, R. 2023. Transforming the Embeddings: A Lightweight Technique for Speech Emotion Recognition Tasks. In Interspeech 2023, 1903--1907

    work page 2023

Show all 38 references
  1. [9]

    B.; and Sharma, R

    Chetia Phukan, O.; Kashyap, G.; Buduru, A. B.; and Sharma, R. 2024. Heterogeneity over Homogeneity: Investigating Multilingual Speech Pre-Trained Models for Detecting Audio Deepfake. In Duh, K.; Gomez, H.; and Bethard, S., eds., Findings of the Association for Computational Li...

    2024

  2. [10]

    Chhibber, M.; Mishra, J.; Shim, H.-J.; and Kinnunen, T. H. 2025. An Explainable Probabilistic Attribute Embedding Approach for Spoofed Speech Characterization. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1--5

    2025

  3. [11]

    ???? High Fidelity Neural Audio Compression

    D \'e fossez, A.; Copet, J.; Synnaeve, G.; and Adi, Y. ???? High Fidelity Neural Audio Compression. Transactions on Machine Learning Research

  4. [12]

    Desplanques, B.; Thienpondt, J.; and Demuynck, K. 2020. ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification. In Interspeech 2020, 3830--3834

    2020

  5. [13]

    Kawa, P.; Plata, M.; Czuba, M.; Szymański, P.; and Syga, P. 2023. Improved DeepFake Detection Using Whisper Features. In Interspeech 2023, 4009--4013

    2023

  6. [14]

    Klein, N.; Chen, T.; Tak, H.; Casal, R.; and Khoury, E. 2024. Source Tracing of Audio Deepfake Systems. In Interspeech 2024, 1100--1104

    2024

  7. [15]

    Li, Y.; Zhang, M.; Ren, M.; Qiao, X.; Ma, M.; Wei, D.; and Yang, H. 2024. Cross-Domain Audio Deepfake Detection: Dataset and Analysis. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 4977--4983

    2024

  8. [16]

    Liu, X.; Wang, X.; Sahidullah, M.; Patino, J.; Delgado, H.; Kinnunen, T.; Todisco, M.; Yamagishi, J.; Evans, N.; Nautsch, A.; et al. 2023. Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild. IEEE/ACM Transactions on Audio, Speech, and Language Processing,...

    2023

  9. [17]

    Lu, Y.; Xie, Y.; Fu, R.; Wen, Z.; Tao, J.; Wang, Z.; Qi, X.; Liu, X.; Li, Y.; Liu, Y.; Wang, X.; and Shi, S. 2024. Codecfake: An Initial Dataset for Detecting LLM-based Deepfake Audio. In Interspeech 2024, 1390--1394

    2024

  10. [18]

    Mishra, J.; Chhibber, M.; Shim, H.-j.; and Kinnunen, T. H. 2025. Towards Explainable Spoofed Speech Attribution and Detection: a Probabilistic Approach for Characterizing Speech Synthesizer Components. arXiv preprint arXiv:2502.04049

    2025 arXiv

  11. [19]

    Müller, N.; Diekmann, F.; and Williams, J. 2022. Attacker Attribution of Audio Deepfakes. In Interspeech 2022, 2788--2792

    2022

  12. [20]

    H.; Vestman, V.; Todisco, M.; Delgado, H.; Sahidullah, M.; Yamagishi, J.; and Lee, K

    Nautsch, A.; Wang, X.; Evans, N.; Kinnunen, T. H.; Vestman, V.; Todisco, M.; Delgado, H.; Sahidullah, M.; Yamagishi, J.; and Lee, K. A. 2021. ASVspoof 2019: spoofing countermeasures for the detection of synthesized, converted and replayed speech. IEEE Transactions on Biometric...

    2021

  13. [21]

    Novoselov, S.; Kozlov, A.; Lavrentyeva, G.; Simonchik, K.; and Shchemelinin, V. 2016. STC anti-spoofing systems for the ASVspoof 2015 challenge. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 5475--5479

    2016

  14. [22]

    B.; and Patil, H

    Patel, T. B.; and Patil, H. A. 2015. Combining evidences from mel cepstral, cochlear filter cepstral and instantaneous frequency features for detection of natural vs. spoofed speech. In Interspeech 2015, 2062--2066

    2015

  15. [23]

    Pratap, V.; Tjandra, A.; Shi, B.; Tomasello, P.; Babu, A.; Kundu, S.; Elkahky, A.; Ni, Z.; Vyas, A.; Fazel-Zarandi, M.; et al. 2024. Scaling speech technology to 1,000+ languages. Journal of Machine Learning Research, 25(97): 1--52

    2024

  16. [24]

    W.; Xu, T.; Brockman, G.; McLeavey, C.; and Sutskever, I

    Radford, A.; Kim, J. W.; Xu, T.; Brockman, G.; McLeavey, C.; and Sutskever, I. 2023. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, 28492--28518. PMLR

    2023

  17. [25]

    Rathod, S.; Charola, M.; Vora, A.; Jogi, Y.; and Patil, H. A. 2023. Whisper Features for Dysarthric Severity-Level Classification. In Interspeech 2023, 1523--1527

    2023

  18. [26]

    Snyder, D.; Garcia-Romero, D.; Sell, G.; Povey, D.; and Khudanpur, S. 2018. X-Vectors: Robust DNN Embeddings for Speaker Recognition. 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 5329--5333

    2018

  19. [27]

    Todisco, M.; Delgado, H.; and Evans, N. 2017. Constant Q cepstral coefficients: A spoofing countermeasure for automatic speaker verification. Computer Speech & Language, 45: 516--535

    2017

  20. [28]

    Wang, Z.; Ye, D.; Li, J.; and Deng, J. 2025. Generalize Audio Deepfake Algorithm Recognition via Attribution Enhancement. In ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1--5

    2025

  21. [29]

    Wu, H.; Tseng, Y.; and yi Lee, H. 2024. CodecFake: Enhancing Anti-Spoofing Models Against Deepfake Audios from Codec-Based Speech Synthesis Systems. In Interspeech 2024, 1770--1774

    2024

  22. [30]

    Xie, Y.; Fu, R.; Wen, Z.; Wang, Z.; Wang, X.; Cheng, H.; Ye, L.; and Tao, J. 2024. Generalized Source Tracing: Detecting Novel Audio Deepfake Algorithm with Real Emphasis and Fake Dispersion Strategy. In Interspeech 2024, 4833--4837

    2024

  23. [31]

    Xie, Y.; Lu, Y.; Fu, R.; Wen, Z.; Wang, Z.; Tao, J.; Qi, X.; Wang, X.; Liu, Y.; Cheng, H.; Ye, L.; and Sun, Y. 2025 a . The Codecfake Dataset and Countermeasures for the Universally Detection of Deepfake Audio. IEEE Transactions on Audio, Speech and Language Processing, 33: 386--400

    2025

  24. [32]

    Xie, Y.; Wang, X.; Wang, Z.; Fu, R.; Wen, Z.; Cao, S.; Ma, L.; Li, C.; Cheng, H.; and Ye, L. 2025 b . Neural Codec Source Tracing: Toward Comprehensive Attribution in Open-Set Condition. arXiv preprint arXiv:2501.06514

    2025 arXiv

  25. [33]

    Yan, X.; Yi, J.; Tao, J.; Wang, C.; Ma, H.; Tian, Z.; and Fu, R. 2022 a . System fingerprints detection for deepfake audio: An initial dataset and investigation. CoRR

    2022

  26. [34]

    Yan, X.; Yi, J.; Tao, J.; Wang, C.; Ma, H.; Wang, T.; Wang, S.; and Fu, R. 2022 b . An initial investigation for detecting vocoder fingerprints of fake audio. In Proceedings of the 1st International Workshop on Deepfake Detection for Audio Multimedia, 61--68

    2022

  27. [35]

    Zang, Y.; Shi, J.; Zhang, Y.; Yamamoto, R.; Han, J.; Tang, Y.; Xu, S.; Zhao, W.; Guo, J.; Toda, T.; and Duan, Z. 2024 a . CtrSVDD: A Benchmark Dataset and Baseline Analysis for Controlled Singing Voice Deepfake Detection. In Interspeech 2024, 4783--4787

    2024

  28. [36]

    Zang, Y.; Zhang, Y.; Heydari, M.; and Duan, Z. 2024 b . Singfake: Singing voice deepfake detection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 12156--12160. IEEE

    2024

  29. [37]

    Zeghidour, N.; Luebs, A.; Omran, A.; Skoglund, J.; and Tagliasacchi, M. 2021. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30: 495--507

    2021

  30. [38]

    Zhang, Q.; Wen, S.; and Hu, T. 2024. Audio Deepfake Detection with Self-Supervised XLS-R and SLS Classifier. In Proceedings of the 32nd ACM International Conference on Multimedia, MM '24, 6765–6773. New York, NY, USA: Association for Computing Machinery. ISBN 9798400706868

    2024

Pith tools

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