Pith. sign in

REVIEW 2 major objections 4 minor 28 references

Source tracing works better when a synthetic voice is treated as architecture plus training data plus residual, not as one monolithic class.

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

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

T0 review · grok-4.5

2026-07-12 04:38 UTC pith:AEERA6KY

load-bearing objection Solid, modest methodological step for compositional open-set source tracing; the residual-subspace story is under-isolated but the empirical gains are real and cleanly reported. the 2 major comments →

arxiv 2607.03134 v1 pith:AEERA6KY submitted 2026-07-03 eess.AS cs.LG

Open-Set Source Tracing as Compositional Factors via Structured Prototypes

Santiago Rubio , Antonio Almud\'evar , Antonio Miguel , Eduardo Lleida , Alfonso Ortega This is my paper
classification eess.AS cs.LG
keywords source tracingaudio deepfake detectionopen-set attributionprototype learningcompositional generalizationsubspace partitioningsynthetic speech
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.

Most source-tracing systems treat each generative model as a single label. This paper argues that a synthetic speech source is better defined as a compositional tuple of architecture, training data, and residual training factors. By forcing the embedding into separate subspaces for architecture and data, and leaving a residual subspace free for everything else, the model can recombine known factors into novel pairings it never saw during training. On the MLAAD benchmark, few-shot open-set attribution improves over standard angular-margin baselines, with the largest gains when at least one generative factor has already been observed. The practical stake is forensic: as new synthesizers appear daily, a system that generalizes by composition can attribute audio without exhaustive retraining on every new model.

Core claim

Redefining a synthetic-speech source as the tuple (architecture, training data, residual training factors) and aligning the embedding to structured orthonormal prototypes in factorized subspaces yields higher few-shot open-set attribution accuracy and a smaller generalization gap than monolithic angular-margin learning, because novel combinations can be assembled from the already-learned factor bases.

What carries the argument

Subspace-partitioned structured orthonormal prototypes: the embedding is split into fixed architecture and data subspaces whose bases are orthonormal prototypes built from metadata labels, plus a residual subspace whose energy is softly constrained so that it absorbs unlabeled interactions without collapsing or dominating the factor subspaces.

Load-bearing premise

The architecture and training-data labels supplied by the dataset are complete and independent enough that fixed prototypes can be aligned to them and everything else can safely live in a residual subspace.

What would settle it

Train the same subspace-partitioned model on a corpus whose architecture and data labels are deliberately corrupted or entangled; if few-shot open-set F1-macro then falls to or below the ArcFace baseline, the compositional claim fails.

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

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

2 major / 4 minor

Summary. The paper redefines synthetic-speech source tracing as attribution of a compositional tuple S=(A,D,H) rather than a monolithic architecture label. It proposes Structured Orthonormal Prototypes and a Subspace Partitioning strategy that splits the embedding into architecture (Z_A), training-data (Z_D) and residual (Z_R) subspaces, with fixed orthonormal bases for the labeled factors and an energy constraint on Z_R (Eqs. 2–4). On MLAAD, under a few-shot open-set protocol, the residual-modeling variant (Strategy 2.B) improves All-vs-All F1-macro over ArcFace (84.52 % vs 80.11 % at K=5) and reduces the seen–unseen generalization gap, with a five-way OOD breakdown (Seen / Unseen Combination / Unseen Dataset / Unseen Arch / Both Unseen) reported in Table III and Figure 3.

Significance. If the claimed compositional mechanism holds, the work supplies a concrete geometric recipe for open-set source tracing that scales more favorably than monolithic prototypes (C_A imes C_D sources from only C_A + C_D bases) and offers a path toward factor-level forensic interpretability. The multi-seed means/stds, explicit closed-set vs few-shot protocol, and five-way OOD decomposition are strengths that make the empirical claims falsifiable. The absolute gains remain modest, so the primary contribution is the problem reformulation and the structured-prototype framework rather than a decisive performance leap.

major comments (2)
  1. The central claim that subspace partitioning enables compositional generalization rests on Z_A and Z_D remaining pure while Z_R absorbs residual H and unlabeled variability (Eqs. 3–4, §II-C). Figure 3 shows the largest relative lift for Strategy 2.B precisely on Unseen Combination and Unseen Architecture. Yet nothing forces the residual energy constraint (µ_ref = 1) to leave the factor subspaces pure: non-linear A×D interactions or unrecorded factors can migrate into Z_R, after which nearest-prototype matching on the full embedding can succeed without true compositional reuse of the marginal prototypes. No ablation freezes or zeros Z_R at test time, nor reports the fraction of cosine similarity for novel (a,d) pairs carried by residual versus factor coordinates. Without that isolation the modest All-vs-All gain is consistent with ordinary prototype regularization plus an extra free subsp
  2. The weakest modeling assumption is that MLAAD’s architecture and training-data metadata are sufficiently complete and independent labels of the true generative factors (Table II, §II-B, §III-A). Sources with undocumented training data receive independent “unknown” labels, and speaker labels are discarded as inconsistent. If those metadata are incomplete, noisy or entangled with unrecorded factors, the fixed bases U and V cannot be cleanly aligned and the residual subspace cannot be guaranteed to absorb only the intended remainder. A sensitivity analysis (e.g., randomly permuting a fraction of factor labels, or treating “unknown” as a single shared class) would quantify how much of the reported OOD lift depends on metadata fidelity.
minor comments (4)
  1. Table III reports means ± std over 4 seeds and 100 few-shot trials, but no statistical significance test (paired t-test or bootstrap CI) is given for the 4.4 pp All-vs-All gap between Strategy 2.B and ArcFace; a short significance statement would strengthen the claim.
  2. Figure 3 caption states “Best-performing configuration selected for each strategy based on validation performance,” yet the main text never lists which λ was chosen for each bar; adding that information would aid reproducibility.
  3. Notation for the residual energy target is written µ_ref = 1/2 (∥u_a∥^{2} + ∥v_d∥^{2}) = 1; the intermediate equality is redundant once unit-norm prototypes are assumed and could be simplified.
  4. The abstract claims the approach “significantly outperforms angular-margin baselines”; given the absolute margins, “consistently outperforms” would be more precise.

Circularity Check

0 steps flagged

No circularity: empirical few-shot F1 gains on held-out MLAAD sources do not reduce by construction to training inputs or self-citations.

full rationale

The paper’s load-bearing claims are empirical comparisons (Table III, Figure 3) of F1-macro under closed-set and few-shot open-set protocols on MLAAD, with architectures/datasets held out of training (Table II). Prototypes for seen factors are fixed orthonormal bases built from training metadata; dynamic prototypes for evaluation sources are averages of K support embeddings computed at test time. The losses (Eqs. 1–4) are ordinary training objectives; none of the reported F1 numbers is algebraically forced by a fitted parameter or by the definition of the subspaces. The self-citation to Almudévar et al. (Interspeech 2024) supplies inspiration for predefined prototypes and is not used as a uniqueness theorem or as the sole justification of the performance claims. Concerns that Z_R may absorb compositional signal rather than isolate it are mechanism/correctness questions, not circular reductions of prediction to input. The derivation chain is therefore self-contained against the external benchmark.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on standard deep-metric-learning machinery plus three paper-specific modeling choices: frozen orthonormal factor prototypes, an orthogonal subspace partition, and a residual energy constraint. Free parameters are the usual training hyper-parameters and the manually chosen subspace dimensionalities and λ values. No new physical entities are postulated; the residual subspace is an engineering construct.

free parameters (4)
  • MSE regularization strength λ
    Grid-searched over {0.0, 0.05, 0.1, 1.0}; best open-set performance reported at λ=0.05 for Strategy 2.B. Directly affects the claimed ranking of methods.
  • Subspace dimensionalities (M_A, M_D, M_R)
    Hand-chosen (64/64 for 2.A; 48/48/32 for 2.B) with total M=128 fixed. Controls capacity allocated to each factor and therefore the measured compositional gains.
  • ArcFace scale s and margin m
    Standard but unreported exact values for the baseline; affect the absolute numbers against which gains are claimed.
  • Residual reference magnitude μ_ref
    Fixed to 1 by construction (½(‖u_a‖²+‖v_d‖²)). Controls residual energy; not learned from data but chosen to match unit prototypes.
axioms (4)
  • domain assumption A generative source can be adequately represented as the compositional tuple (A, D, H) whose labeled factors are recoverable from MLAAD metadata.
    Stated in Section II-B and used to justify the entire factorization; if metadata are incomplete the subspaces cannot be correctly supervised.
  • ad hoc to paper Frozen orthonormal bases for architecture and data factors minimize class overlap while permitting shared structure for sources that reuse a factor.
    Core design choice of Strategies 1 and 2; not derived from a uniqueness theorem but imposed geometrically.
  • domain assumption Inference-time variables (speaker, prompt, phonetic content) and unlabeled training stochasticity can be absorbed into a residual subspace without destroying factor discriminability.
    Section II-B/C; the residual energy constraint is introduced precisely to make this absorption stable.
  • standard math Cosine similarity to averaged support embeddings is a valid few-shot open-set decision rule.
    Standard prototype / nearest-centroid evaluation used throughout metric-learning literature.
invented entities (2)
  • Residual subspace Z_R with energy constraint no independent evidence
    purpose: Absorb unlabeled training residual H, non-linear A×D interactions, and inference-time nuisance so that the labeled factor subspaces remain clean.
    Introduced in Strategy 2.B; no independent physical existence claimed; purely a modeling device whose utility is measured by the open-set F1 gains.
  • Composite factorized prototype p_s = u_a ⊕ v_d no independent evidence
    purpose: Encode shared architecture or data structure while still distinguishing full sources.
    Constructed in Strategy 2.A from the orthonormal bases; existence is definitional once the bases are chosen.

pith-pipeline@v1.1.0-grok45 · 15665 in / 3219 out tokens · 28373 ms · 2026-07-12T04:38:40.255693+00:00 · methodology

0 comments
read the original abstract

Recent research expands beyond binary anti-spoofing with the emergence of Source Tracing, the task of identifying the specific generative origins of synthetic speech. However, current research often equates a "source" with its generative architecture. We propose redefining a source as a compositional tuple of Architecture, Training Data, and other training factors affecting the generated speech. We propose a framework using Structured Orthonormal Prototypes to minimize class overlap and intra-class variance. Our Subspace Partitioning strategy splits the embedding into architecture and data subspaces, while a residual subspace captures stochastic variability, enabling "compositional generalization" for novel factor combinations. This approach improves performance for partially seen sources and maintains robustness in fully open-set scenarios. MLAAD evaluations for Few-Shot open-set Identification show our approach significantly outperforms angular-margin baselines.

Figures

Figures reproduced from arXiv: 2607.03134 by Alfonso Ortega, Antonio Almud\'evar, Antonio Miguel, Eduardo Lleida, Santiago Rubio.

Figure 1
Figure 1. Figure 1: Variability in synthetic speech attribution. A source [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Proposed Source Tracing Framework. Top (Strategy 1): Structured Orthonormal Prototypes constrained by [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: F1-macro scores for few-shot source attribution with K= 5 support [PITH_FULL_IMAGE:figures/full_fig_p006_3.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

28 extracted references · 4 linked inside Pith

  1. [1]

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

    Z. Wu, T. Kinnunen, N. Evans, J. Yamagishi, C. Hanilc ¸i, M. Sahidullah, and A. Sizov, “Asvspoof 2015: the first automatic speaker verification spoofing and countermeasures challenge,” inInterspeech 2015, 2015, pp. 2037–2041

    2015

  2. [2]

    The asvspoof 2017 challenge: Assessing the limits of replay spoofing attack detection,

    T. Kinnunen, M. Sahidullah, H. Delgado, M. Todisco, N. Evans, J. Ya- magishi, and K. A. Lee, “The asvspoof 2017 challenge: Assessing the limits of replay spoofing attack detection,” inInterspeech 2017, 2017, pp. 2–6

    2017

  3. [3]

    Asvspoof 2019: Future horizons in spoofed and fake audio detection,

    M. Todisco, X. Wang, V . Vestman, M. Sahidullah, H. Delgado, A. Nautsch, J. Yamagishi, N. Evans, T. H. Kinnunen, and K. A. Lee, “Asvspoof 2019: Future horizons in spoofed and fake audio detection,” inInterspeech 2019, 2019, pp. 1008–1012

    2019

  4. [4]

    Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild,

    X. Liu, X. Wang, M. Sahidullah, J. Patino, H. Delgado, T. Kinnunen, M. Todisco, J. Yamagishi, N. Evans, A. Nautsch, and K. A. Lee, “Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 2507–2522, 2023

    2021

  5. [5]

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

    X. Wang, H. Delgado, H. Tak, J. weon Jung, H. jin Shim, M. Todisco, I. Kukanov, X. Liu, M. Sahidullah, T. Kinnunen, N. Evans, K. A. Lee, and J. Yamagishi, “Asvspoof 5: Crowdsourced speech data, deepfakes, and adversarial attacks at scale,” 2024. [Online]. Available: https://arxiv.org/abs/2408.08739

    Pith/arXiv arXiv 2024

  6. [6]

    Add 2022: the first audio deep synthesis detection challenge,

    J. Yi, R. Fu, J. Tao, S. Nie, H. Ma, C. Wang, T. Wang, Z. Tian, Y . Bai, C. Fan, S. Liang, S. Wang, S. Zhang, X. Yan, L. Xu, Z. Wen, H. Li, Z. Lian, and B. Liu, “Add 2022: the first audio deep synthesis detection challenge,” inICASSP 2022 – 2022 IEEE International Conference on Acoustics, Speech and Signal Processing, 2022, pp. 9216–9220

    2022

  7. [7]

    Add 2023: the second audio deepfake detection challenge,

    J. Yi, J. Tao, R. Fu, X. Yan, C. Wang, T. Wang, C. Y . Zhang, X. Zhang, Y . Zhao, Y . Ren, L. Xu, J. Zhou, H. Gu, Z. Wen, S. Liang, Z. Lian, S. Nie, and H. Li, “Add 2023: the second audio deepfake detection challenge,” inDADA@IJCAI 2023 (Second Audio Deepfake Detection Challenge Workshop), 2023, pp. 125–130

    2023

  8. [8]

    Source Tracing of Audio Deepfake Systems,

    N. Klein, T. Chen, H. Tak, R. Casal, and E. Khoury, “Source Tracing of Audio Deepfake Systems,” inInterspeech 2024, 2024, pp. 1100–1104

    2024

  9. [9]

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

    N. M. M ¨uller, P. Kawa, W. H. Choong, E. Casanova, E. G¨olge, T. M¨uller, P. Syga, P. Sperl, and K. B ¨ottinger, “Mlaad: The multi-language audio anti-spoofing dataset,” in2024 International Joint Conference on Neural Networks (IJCNN), 2024, pp. 1–7

    2024

  10. [10]

    STOPA: A Dataset of Systematic VariaTion Of DeePfake Audio for Open-Set Source Tracing and Attribution,

    A. Firc, M. Chhibber, J. Mishra, V . Pratap Singh, T. Kinnunen, and K. Malinka, “STOPA: A Dataset of Systematic VariaTion Of DeePfake Audio for Open-Set Source Tracing and Attribution,” inInterspeech 2025, 2025, pp. 1553–1557

    2025

  11. [11]

    Codecfake+: A large-scale neural audio codec-based deepfake speech dataset,

    X. Chen, J. Du, H. Wu, L. Zhang, I. Lin, I. Chiu, W. Ren, Y . Tseng, Y . Tsao, J.-S. R. Janget al., “Codecfake+: A large-scale neural audio codec-based deepfake speech dataset,”arXiv preprint arXiv:2501.08238, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Audio Deepfake Source Tracing using Multi-Attribute Open-Set Identification and Verification,

    P. Falez, T. Marteau, D. Lolive, and A. Delhay, “Audio Deepfake Source Tracing using Multi-Attribute Open-Set Identification and Verification,” inInterspeech 2025, 2025, pp. 1528–1532

    2025

  13. [13]

    Codec-Based Deepfake Source Tracing via Neural Audio Codec Taxonomy,

    X. Chen, I.-M. Lin, L. Zhang, J. Du, H. Wu, H. yi Lee, and J.-S. R. Jang, “Codec-Based Deepfake Source Tracing via Neural Audio Codec Taxonomy,” inInterspeech 2025, 2025, pp. 1538–1542

    2025

  14. [14]

    Advancing Zero-Shot Open-Set Speech Deepfake Source Tracing,

    M. Chhibber, J. Mishra, and T. H. Kinnunen, “Advancing Zero-Shot Open-Set Speech Deepfake Source Tracing,” inOdyssey 2026, 2026, pp. 185–190

    2026

  15. [15]

    Source Verification for Speech Deepfakes ,

    V . Negroni, D. Salvi, P. Bestagini, and S. Tubaro, “ Source Verification for Speech Deepfakes ,” inInterspeech 2025, 2025, pp. 1548–1552

    2025

  16. [16]

    Open-Set Source Tracing of Audio Deepfake Systems,

    N. Klein, H. Tak, and E. Khoury, “Open-Set Source Tracing of Audio Deepfake Systems,” inInterspeech 2025, 2025, pp. 1578–1582

    2025

  17. [17]

    How people make their own environments: A theory of genotype→environment effects,

    S. Scarr and K. McCartney, “How people make their own environments: A theory of genotype→environment effects,”Child Development, vol. 54, no. 2, pp. 424–435, 1983. [Online]. Available: http://www.jstor.org/stable/1129703

    arXiv 1983

  18. [18]

    Unveiling Audio Deepfake Origins: A Deep Metric learning And Conformer Network Approach With Ensemble Fusion,

    A. Kulkarni, S. Dowerah, T. Alum ¨ae, and M. M. Doss, “Unveiling Audio Deepfake Origins: A Deep Metric learning And Conformer Network Approach With Ensemble Fusion,” inInterspeech 2025, 2025, pp. 1533– 1537

    2025

  19. [19]

    Syn- thetic Speech Source Tracing using Metric Learning,

    D. Koutsianos, S. Zacharopoulos, Y . Panagakis, and T. Stafylakis, “Syn- thetic Speech Source Tracing using Metric Learning,” inInterspeech 2025, 2025, pp. 1558–1562

    2025

  20. [20]

    Predefined Prototypes for Intra-Class Separation and Disentanglement,

    A. Almud ´evar, T. Mariotte, A. Ortega, M. Tahon, L. Vicente, A. Miguel, and E. Lleida, “Predefined Prototypes for Intra-Class Separation and Disentanglement,” inInterspeech 2024, 2024, pp. 3809–3813

    2024

  21. [21]

    An interpretable deep learning model for automatic sound classification,

    P. Zinemanas, M. Rocamora, M. Miron, F. Font, and X. Serra, “An interpretable deep learning model for automatic sound classification,”Electronics, vol. 10, no. 7, 2021. [Online]. Available: https://www.mdpi.com/2079-9292/10/7/850

    2021

  22. [22]

    Assessing the impact of speaker identity in speech spoofing detection,

    A.-T. Daoet al., “Assessing the impact of speaker identity in speech spoofing detection,” inICASSP 2026 - IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026

    2026

  23. [23]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, J. Yang, N. Xue, I. Kotsia, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 10, p. 5962–5979, Oct. 2022. [Online]. Available: http://dx.doi.org/10.1109/TPAMI.2021.3087709

    doi:10.1109/tpami.2021.3087709 2022

  24. [24]

    Extensions of lipschitz mappings into a hilbert space,

    W. B. Johnson, J. Lindenstrausset al., “Extensions of lipschitz mappings into a hilbert space,”Contemporary mathematics, vol. 26, no. 189-206, p. 1, 1984

    1984

  25. [25]

    Ring loss: Convex feature nor- malization for face recognition,

    Y . Zheng, D. K. Pal, and M. Savvides, “Ring loss: Convex feature nor- malization for face recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    2018

  26. [26]

    Vicreg: Variance-invariance- covariance regularization for self-supervised learning,

    A. Bardes, J. Ponce, and Y . LeCun, “Vicreg: Variance-invariance- covariance regularization for self-supervised learning,” inInternational Conference on Learning Representations (ICLR), 2022

    2022

  27. [27]

    Mls: A large-scale multilingual dataset for speech research,

    V . Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “Mls: A large-scale multilingual dataset for speech research,”ArXiv, vol. abs/2012.03411, 2020

    Pith/arXiv arXiv 2012

  28. [28]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” 2015. [Online]. Available: https://arxiv.org/abs/1512.03385

    Pith/arXiv arXiv 2015