acceptodds
Under review as a conference paper at ICLR 2027

SILO: Subspace-Preserving Spectral Injection for Generalizable Audio Deepfake Detection

Abstract

Audio deepfake detectors often struggle to generalize across unseen synthesis systems and real-world acoustic conditions. We first analyze the **complementary failure modes** of classical spectral features and self-supervised learning (SSL) representations: spectral cues expose transferable synthesis artifacts but are sensitive to acoustic variation, whereas SSL representations are more robust to such variation yet can miss fine-grained spoofing artifacts. Motivated by this observation, we propose **Spectral Injection via Low-rank Orthogonality (SILO)**, which uses utterance-adaptive spectral evidence to complement the robust SSL representation through a dedicated low-rank spectral-injection branch. SILO further introduces a **geometric constraint** that separates low-rank directions from their signed amplitudes and a **preserved-subspace constraint** that protects selected high-energy pretrained directions. Both constraints are maintained through optimization on restricted Stiefel manifolds. Across cross-dataset evaluations, SILO outperforms state-of-the-art methods, achieving 1.02% EER on ASVspoof 2019, 1.56% on ASVspoof 2021, and 14.38% macro-EER on MLAAD. We further introduce **Voice-Disorder Deepfake (VDD)** and **ForVoice Deepfake (FVD)**, two evaluation datasets targeting pathological speech and long-form forensic recordings, respectively. These datasets extend evaluation to these underexplored conditions and reveal substantial remaining challenges in pathological-speech and long-form deepfake detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Audio Deepfake Detection, Parameter-Efficient Adaptation, Out-of-Distribution Generalization, Unseen Generator Generalization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9wTROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focus on deep fake detection in audio by presenting a LORA-like approach where high frequency constraint is injected in-between the orthonormal matrices of low rank. The authors (like others) assume that high frequencies better capture fake audio and hence a learnable unit within boosted frequencies will better learn those patterns. On tested benchmark this approach is SOTA.

Critical strengths

SOTA results on standard benchmark Interesting approach to combine LORA with spectral constraint

Critical weaknesses

While the author upgrade the conventional LORA with an additional middle layer, and indeed share the motivation from SUV, I still find it a bit unclear if indeed higher frequency are learned. Why would that approach will better or faster learn high frequency is kept un answered. I believe the authors should have at least shown that case on toy problem even if they cannot explain that analytically.

Questions for authors

Can you show the modified LORA learn faster data with high frequency (toy problem) ? Can you prove why?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 9wTR: SILO has no high-frequency constraint; evidence that the spectral branch contributes

We thank the reviewer for the constructive feedback. However, the review describes a mechanism that differs from the one we propose, so we clarify that first and then give the evidence that the spectral branch is used, together with an additional ablation. Table and appendix numbers below refer to the revised PDF.

What SILO does and does not claim. SILO does not place a frequency constraint between the low-rank factors, and we do not claim that the constrained LoRA learns high-frequency content faster or better. The middle factor diag(α)\mathrm{diag}(\alpha) holds the signed amplitudes of the residual, and their absolute values ∣αi∣|\alpha_i| are exactly its singular values (Prop. 1). The geometric constraint controls the strength of adaptation and the preserved-subspace constraint controls which pretrained directions may change (Sec. 3.4, L192-195; Theorem 1). Neither refers to the high frequencies mentioned by the reviewer. Spectral evidence enters through a separate path: a frozen LFCC encoder produces ztz_t, and a second low-rank map, ΔWczt\Delta W_c z_t, scaled by a learned utterance gate, is added to the SSL projection (Eq. 2; Secs. 3.2-3.3; Fig. 3). The LFCC front-end is linear in frequency across the full band (120 coefficients from 140 filters, App. D.4) with no emphasis on high frequencies. Therefore, our claim is complementarity: spectral features expose transferable synthesis artifacts but are fragile to acoustic variation, while SSL features behave the other way round (Sec. 1; Fig. 1).

Moreover, our work does not use any technique related to "SUV". Could the reviewer clarify which technique is meant? If SVD was intended, our three-factor parameterization is indeed inspired by the SVD (Sec. 1, L075-082; Fig. 2), and it involves no frequency constraint.

Evidence that the spectral branch contributes. Our submission already provides both theoretical and empirical evidence that the spectral branch contributes:

  • Figs. 4 and 11 (Sec. 5.1.2; App. F.4): on generators where the SSL-based detectors fail (MMS-TTS, Tacotron2, OverFlow, SpeechT5, FastPitch), SILO closes much of the gap to the classical spectral detectors.
  • Tab. 19: removing injection raises MLAAD EER from 14.38 to 24.67, and all four classical front-ends (LFCC 14.38, LPS 15.56, CQCC 17.09, MFCC 18.09) recover most of this gain, so the effect is not tied to one feature.
  • Tab. 20: increasing a fixed injection strength from 0 to 1 improves ASVspoof 2019 (6.23 to 1.21) and MLAAD (25.25 to 15.89), while In-the-Wild worsens (1.43 to 7.05), which is the predicted signature of spectral evidence.
  • Props. 3 and 4 (App. C.6) characterize how the gate is learned from the detection loss and how it scales the injected contribution.

In case these results do not fully satisfy the reviewer, we additionally provide an ablation that isolates the contribution of each branch of SILO (EER %, lower is better, all models trained on WaveFake; added to the revised paper as Tab. 6 in Sec. 5.2):

ConfigurationASVS-19ASVS-21ITWMLAAD
Spectral only33.3445.3567.6918.79
SSL + no constraint1.282.182.3726.83
SSL + both constraints1.552.152.8024.67
Both branches, no constraints + no gate2.374.0310.2918.19
Both branches, no constraints + gate2.272.716.1517.01
SILO (full)1.021.564.3514.38

As shown in the table, the SSL-only branch has the highest EER on MLAAD (which covers 116 synthesis systems unseen during training), while the spectral-only branch has the highest EER on ITW (which includes many acoustic variations: background noise, room reverberation, etc.). Full SILO, which combines both branches, obtains the lowest EER on ASVS-19, ASVS-21, and MLAAD, with 4.35 on ITW.

Questions: requested experiment on a toy dataset and proof.

We do not have an analytical result on faster learning of any particular frequency band because the paper does not claim one and does not place a frequency constraint between the low-rank factors. For the same reason, a toy experiment on high-frequency learning would test a property that SILO neither has nor claims. The properties the paper does claim are proven (Prop. 1, Theorem 1, Props. 3 and 4) and tested by the ablation above.

Reviewer_9wTROpenReview ↗

Frequency control

Following the clarification that frequency control is not the mechanism that drives the results, why is the spectral branch doing so? Beside evidence of improved results on the tested data, when will it most likely succeed? when will it fail?

AuthorsOpenReview ↗

Response to Reviewer 9wTR's comment on Why and When does the spectral branch Help

We thank the reviewer for the follow-up. We have added an analysis of the eight recent generators in Fig. 4 to support our explanation of what the spectral encoder detects and when this evidence helps or becomes unreliable (App. F.6, Figs. 13–14). Numbers refer to the revised PDF.

Why the spectral branch helps. The SSL-adaptation branch transforms the existing SSL representation, while the spectral branch supplies an additional representation computed directly from the short-time spectrum. This means that the spectral branch reads a generation artifact that SSL features do not expose. Our new analysis identifies a concrete sensitivity of this spectral encoder:

  • Sensitivity to fine spectral detail. Smoothing the power spectrum of genuine speech along frequency within 4–8 kHz progressively shifts its score toward generated speech. At approximately 300 Hz smoothing width, its mean score reaches the generated-speech mean. Randomizing phase alone leaves the score approximately unchanged (Fig. 13, right). This supports sensitivity to spectral detail removed by smoothing.

  • Evidence within speech. Using speech frames alone gives AUCs of 0.89–1.00 across the eight generators, matching or exceeding full-clip scoring. The most influential frequency band varies across generators (Fig. 13, left and middle).

  • Transfer beyond training generators. The encoder achieves this separation after training only on WaveFake. Together with the reduction in MLAAD EER from 24.67% without injection to 14.38% with injection (Tab. 19), these results support the value of complementary spectral evidence that SSL-only adaptation does not sufficiently exploit.

When it is most likely to help. We expect the largest benefit when an unseen generator leaves spectral differences detectable by the encoder, SSL-only detectors struggle to distinguish its speech, and the recording channel preserves those differences. Figures 4 and 11 support this expectation: SILO improves over the evaluated SSL-based baselines, including single-digit EER on MMS-TTS, Tacotron2, OverFlow, SpeechT5, and FastPitch.

When its benefit weakens or detection fails.

  1. Severe acoustic degradation. Noise masks spectral detail and reverberation blurs it, as illustrated in Fig. 14. Correspondingly, 0 dB noise and 900 ms reverberation increase SILO’s EER by approximately 16–22 and 30–33 percentage points, respectively (Fig. 10).

  2. Genuine speech affected by voice disorders. Pathological phonation can alter spectral structure in ways associated with synthesis (Figs. 7–8). SILO achieves the lowest average EER on VDD, but organic dysphonia remains difficult at 27.9% EER (Tab. 3).

  3. Absence of transferable spectral differences. We expect little benefit when a generator leaves no differences recognizable by the spectral encoder. Uninformative spectral features can still influence predictions, so automatic fallback to SSL-only behavior is not guaranteed.

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer 9wTR,

Thank you for your follow-up question. As the discussion period is closing soon, we would like to kindly check whether our latest response on why the spectral branch helps and when it succeeds or fails (App. F.6, Figs. 13-14) addresses your concern.

Please let us know if anything remains unclear. We would appreciate your reconsideration of your assessment and rating in light of these responses.

Thank you for your time and consideration.

Reviewer_QXrwOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SILO, an audio deepfake detection framework that combines subspace-preserving adaptation of a pretrained SSL backbone with gated spectral-feature injection.

It also introduces VDD and FVD to evaluate detection on pathological speech and long recordings, respectively. The paper addresses a relevant generalization problem and reports encouraging results across several datasets.

However, the current experimental evidence does not sufficiently isolate the contributions of the proposed components, and important details of the training and evaluation protocols remain unclear.

Critical strengths

The motivation of combining pretrained SSL representations with complementary spectral evidence is reasonable.

VDD and FVD address meaningful evaluation settings. Their potential value as benchmarks extends beyond the particular detector proposed in this paper.

The evaluation includes multiple external datasets, and the appendix provides additional analyses of spectral features, gating, and robustness.

Critical weaknesses
  1. The proposed VDD and FVD datasets are meaningful, but their relationship to SILO's core motivation and architectural design is not sufficiently demonstrated.

  2. The experiments do not fully isolate the contributions of the SSL-adaptation branch, spectral-injection branch, and their combination.

  3. Training/testing audio durations, segmentation strategies, and score aggregation methods are insufficiently specified, particularly for the long-form FVD dataset.

  4. The choice of four TTS and three VC systems lacks sufficient discussion of representativeness and coverage of modern generation techniques.

  5. Training configurations, LoRA ranks, parameter budgets, and optimization settings are not sufficiently documented for a reliable comparison.

  6. Additional evaluation on diverse benchmarks from different domain and languages, such as ASVspoof 5 and SpeechFake,SEA-Spoof , would strengthen the conclusions. Moreover, the claim that the learned gate achieves the lowest EER across all datasets is inconsistent with Table 16.

Questions for authors

1.What is the specific relationship between SILO's main research motivation and the newly introduced VDD and FVD datasets? How do these datasets validate the advantages of the proposed architecture beyond general benchmarking?

2.Figure 3 combines the SSL-adaptation and spectral-injection branches. Could the authors provide more controlled ablations to demonstrate the individual contribution of each branch and the benefit of combining them?

3.What audio duration is used for training and inference? Are recordings divided into fixed-length segments (e.g., 4 seconds)? If so, how are segment-level predictions aggregated for FVD, and what specific long-duration detection capabilities does this evaluation demonstrate?

4.What are the selection criteria for the four TTS and three VC models used to construct VDD and FVD? How representative are these systems of current speech generation techniques? Could the authors clarify their architectural differences, model versions, and open-source or closed-source status?

5.Given that cross-domain generalization is a central motivation, could the authors consider additional benchmarks, such as ASVspoof 5, ASVspoof 2021 LA, SpeechFake or SEA-Spoof, to further evaluate robustness across different generators and recording conditions?

6.Were all baseline models trained using consistent data partitions, preprocessing, audio durations, augmentation, and training or tuning budgets? Could the authors provide more complete training configurations to support the fairness and reproducibility of the comparisons?

7.For the LoRA-based baselines in Table 2, what ranks, adapted layers, optimization settings, and trainable parameter budgets were used? Were these configurations comparable to SILO, including the contribution and computational cost of the additional spectral encoder?

8.Although the paper states that all detectors are trained on WaveFake, could the authors clarify the exact training/validation splits, checkpoint-selection criteria, and whether the spectral encoder and full detector use consistent data partitions?

4 comments
AuthorsOpenReview ↗

Response to Reviewer QXrw (1/3): Independent contributions, branch ablation, and audio duration protocol

We thank the reviewer for the detailed feedback. We address each weakness and question point by point below. Line, table, and appendix numbers refer to the revised PDF.

Q1/W1. The contributions of SILO and VDD and FVD are independent. The contribution structure is already explicit in Section 1, lines 101-107: SILO provides a detection method, while VDD and FVD provide additional evaluation conditions. Section 4.1, lines 286-296, explains these conditions: distinguishing pathological speech from synthesis artifacts and evaluating accumulated evidence in long recordings. This means that we do not claim that our method is built on the two datasets, or that the two datasets were built to train or design SILO. Importantly, a method contribution and a dataset contribution are not required to depend on each other, which is standard practice in deepfake detection [1-5]. Therefore, our dataset contribution should likewise be assessed through its stated evaluation purpose and supporting results.

[1] Li, X., Li, K., Zheng, Y., Yan, C., Ji, X., & Xu, W. (2024, December). SafeEar: Content privacy-preserving audio deepfake detection. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security (pp. 3585-3599).
[2] Yan, S., Li, O., Cai, J., Hao, Y., Jiang, X., Hu, Y., & Xie, W. (2025, May). A sanity check for AI-generated image detection. In International Conference on Learning Representations (Vol. 2025, pp. 70702-70720).
[3] Wang, Z., Bao, J., Zhou, W., Wang, W., Hu, H., Chen, H., & Li, H. (2023, October). DIRE for diffusion-generated image detection. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 22388-22398). IEEE.
[4] Sun, C., Jia, S., Hou, S., & Lyu, S. (2023, June). AI-synthesized voice detection using neural vocoder artifacts. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (pp. 904-912). IEEE.
[5] Liu, W., She, T., Liu, J., Li, B., Yao, D., Liang, Z., & Wang, R. (2024). Lips are lying: Spotting the temporal inconsistency between audio and visual in lip-syncing deepfakes. Advances in Neural Information Processing Systems, 37, 91131-91155.

Q2/W2. Ablation for SSL-adaptation and spectral-injection branches. We provide the table below to specify each branch and the contribution of the constraints (EER %, lower is better, all trained on WaveFake):

ConfigurationASVS-19ASVS-21ITWMLAAD
Spectral only33.3445.3567.6918.79
SSL + no constraint1.282.182.3726.83
SSL + both constraints1.552.152.8024.67
Both branches, no constraints + no gate2.374.0310.2918.19
Both branches, no constraints + gate2.272.716.1517.01
SILO (full)1.021.564.3514.38

As you can see, the SSL branch alone has the highest EER on MLAAD (which covers 116 new-generation synthesis systems), while the spectral branch alone has the highest EER on ITW (which includes many acoustic variations: background noise, room reverberation, etc.).

We provide this result in the revised submission as Tab. 6 (Sec. 5.2).

Q3/W3. Audio duration, segmentation and aggregation.

All detectors, SILO and every baseline, get the same input, which the dataset configuration sets: 16 kHz mono and 64,600 samples (4.04 s). Shorter clips are repeated to that length. Training takes one random 4.04 s crop per utterance per epoch. On ASVspoof 2019/2021, In-the-Wild, MLAAD and VDD, inference scores the first 4.04 s of each file, which gives one score per trial.

For the FVD duration curve in Figure 6, we take the first WW seconds of each recording, where W∈{1,24,60,120,180,300,420,600}W \in \lbrace 1, 24, 60, 120, 180, 300, 420, 600 \rbrace, and process this input in one forward pass. SILO mean-pools its frame-level features across the input before classification, producing one recording-level score. Independently predicted 4.04-second segment scores are not averaged. The purpose of this is to examine how detectors trained on short clips generalize to longer inputs and how detection performance changes as more audio becomes available.

The caption's phrase "scores are aggregated over increasing amounts of audio" summarizes recording-level scoring as progressively more temporal evidence becomes available. Concretely, SILO aggregates the temporal features used to compute the recording-level score. We refined the caption to make this distinction explicit: "Recording-level scores are computed from progressively longer audio prefixes. SILO aggregates frame-level features through temporal mean pooling before classification. Methods are grouped by their best achieved EER." These rules are now specified in the new Appendix D.1.

AuthorsOpenReview ↗

Response to Reviewer QXrw (2/3): Generation systems, evaluation coverage, and Table 16

Q4/W4. All parts of this question are answered in the submission.

  • Selection criteria, generation mechanisms, and public releases. Sec. 4.2 (L299-308) describes the seven generation systems shared by VDD and FVD. We selected systems that can generate Hungarian speech through diverse mechanisms. The four TTS systems explicitly support Hungarian, and the three VC systems perform zero-shot conversion on Hungarian source recordings. Their mechanisms span autoregressive acoustic-token modeling, flow-matching synthesis, VITS-based synthesis, training-free optimal transport, diffusion-based VC, and discrete-token-conditioned VC. Thus, VDD and FVD go beyond conventional vocoder artifacts and ensure the diversity of generation systems. Lines 303-306 identify and cite XTTS-v2, F5-TTS-Hungarian, MMS-TTS-HUN, Piper, MKL-VC, Seed-VC, and EZ-VC. All seven have publicly available implementations or model releases. In the revised manuscript, we added Appendix E.4, "Generation Systems, Hungarian Support, and Public Releases," with a table (Tab. 18) documenting each system's citation, generation mechanism, Hungarian-language support, and public release link. The table reports software and checkpoint licenses separately, including noncommercial terms where applicable.
  • Purpose of selecting these generation systems. Both source corpora are Hungarian (Section 4.1), so each TTS system was taken in a release that supports Hungarian, and the three VC systems are zero-shot and language-agnostic. The purpose is to ensure the quality of generated audio. We also measure this resulting quality in Tabs. 15, 17, and 13. App. E.2 (L1829-1831) discusses the one lower-fidelity system (A07) and why it is kept.
  • Why not state-of-the-art generation systems. VDD and FVD are not built to evaluate modern generators; they isolate one condition each: pathological speech (VDD) and long-form duration (FVD) (App. E.1, L1774-1775). We want to isolate the evaluation to pathological speech and long-form duration only, which means that the generator panel must be held fixed while the condition varies (Sec. 2, L128-130), so that any change in detector performance is attributable to the condition rather than to the generator. Using the newest generators would mix generator novelty with pathology or duration, and the two effects could no longer be separated. Generalization to modern generators is evaluated separately on MLAAD-v3 (116 systems; Tab. 12; Fig. 4).

Q5/W6. We would like to clarify that we already evaluated on 9 datasets (Sec. 5, L331-334): WaveFake (in-domain), ASVspoof 2019 LA, ASVspoof 2021 DF, In-the-Wild, MLAAD-v3, PolyGlotFake, CVoiceFake, VDD and FVD. Specifically, MLAAD-v3 alone covers 116 new generation systems, 21 architectures and 23 languages (Tab. 12). We additionally report 42 audio transformations in six families on three datasets (App. D.5; Figs. 9-10). Therefore, our conclusions do not rest on a narrow evaluation.

Table 16 (Tab. 20 in the revision). We thank the reviewer for pointing out that the learned gate does not achieve the lowest EER on all three evaluated datasets. The learned gate instead achieves the best balance between cross-generator generalization and robustness to real-world acoustic conditions: it gives the lowest EER on ASVspoof 2019 (1.02) and MLAAD (14.38), and 4.35 on In-the-Wild. This is consistent with the paper's central observation (Sec. 1; Fig. 1): spectral evidence helps cross-generator transfer and hurts under uncontrolled acoustic conditions. We revised the description of this table (L2270-2284) to clarify that "Among the evaluated configurations, the learned gate provides the best overall balance, achieving the lowest EER on ASVspoof 2019 (1.02%) and MLAAD (14.38%) while maintaining a competitive EER of 4.35% on In-the-Wild."

AuthorsOpenReview ↗

Response to Reviewer QXrw (3/3): Baseline protocol, LoRA configurations and cost, and data partitions

Q6/W5. Baseline training protocols and reproducibility.

We have added Appendix D.1, "Data, Preprocessing, and Audio Duration," and Appendix D.2, "Baseline Implementations," to document the shared evaluation protocol and method-specific configurations.

All detectors are trained on WaveFake and receive the same dataset-level preprocessing: 16 kHz mono audio, random 64,600-sample training crops, and repetition of shorter utterances. Appendix D.1 specifies the inference-duration rules, including the recording-prefix protocol for FVD. We use a batch size of 50 for all baselines with early stopping for a fair comparison. We retain the original methods' architectures, training objectives, and method-specific optimization settings, which are now documented in Appendix D.2. Data partitioning and checkpoint selection are addressed separately in our response to Q8.

The full source code will be released after acceptance, which provides per-architecture configuration for each baseline.

Q7/W5. LoRA configurations, parameter budgets, and the spectral encoder.

We have added Appendix D.3, "LoRA-Based Baseline Implementations," with a configuration table (Tab. 9) specifying the adapted projections, ranks, optimizers, learning rates, and trainable parameter budgets. We also added "What is the computational cost of SILO?" to the main paper, with a table (Tab. 7) comparing forward GFLOPs and computational overhead.

All three baselines in Table 2 use a frozen XLS-R 300M backbone and adapt the same 96 projections as SILO: the query, key, value, and output projections across all 24 Transformer layers. Their configurations and measured costs are summarized below.

MethodAdaptationOptimizationGFLOPs
XLS-R+LoRARank 16, scaling parameter 16AdamW, lr 10−310^{-3}, weight decay 10−410^{-4}150.05
XLS-R+StelLARank 16 USV⊤USV^\top, with orthonormal U,VU, VAdamW with Stiefel retraction, lr 10−310^{-3}150.06
EffortRank 16 SVD residual with orthogonality and singular-value regularizationAdamW, lr 10−410^{-4}151.94
SILORank 8 SSL residual and rank-32 spectral injection, π=0.80\pi=0.80Restricted-Stiefel momentum, lr 2×10−32\times10^{-3}; AdamW, lr 10−310^{-3}153.88

SILO additionally uses a 56K-parameter spectral encoder pretrained on WaveFake and frozen during detector training. Spectral features are computed once per input and reused across the adapted layers. The encoder's forward computation is included in the reported cost, while its frozen parameters are excluded from the trainable count reported in Table 1.

We measure forward computation for one 4.04-second clip using PyTorch's FLOP counter. SILO requires 3.83 additional GFLOPs relative to rank-16 LoRA, a 2.55% increase, demonstrating that spectral injection adds modest forward computational overhead.

Q8. Data partitions, checkpoint selection, and spectral-encoder pretraining.

We have added Appendix D.1, "Data, Preprocessing, and Audio Duration," to explicitly document the data partitions and checkpoint-selection criterion, complementing the pretraining procedure described in Appendix D.4, "SILO Implementation."

We split WaveFake into 70% training, 10% validation, and 20% testing. All detectors and spectral-encoder pretraining use the same partitions. Hyperparameters are selected using the validation partition, and the checkpoint with the lowest validation EER is retained for evaluation.

The spectral encoder is pretrained on the shared WaveFake training partition as a standalone detector. Its classification head is then discarded, and its encoder weights are frozen during SILO training. The held-out WaveFake test partition and external datasets, including VDD and FVD, are used for evaluation.

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer QXrw,

As the discussion period is closing soon, we would like to kindly follow up on our response. We have added the branch-level ablation, clarified the duration and FVD scoring protocol, and documented the baseline, LoRA and data-partition details in the revised appendices.

Please let us know if any concern remains. We would appreciate your reconsideration of your assessment and rating in light of these responses.

Thank you for your time and consideration.

Reviewer_ioitOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SILO for generalizable audio deepfake detection. It combines pretrained SSL representations with spectral features through a gated low-rank injection mechanism. Experiments show improved cross-dataset generalization, and two new datasets are introduced for pathological-speech and long-form deepfake detection.

Critical strengths
  1. The paper addresses an important problem of cross-domain generalization in audio deepfake detection.
  2. The combination of spectral injection and subspace-preserving adaptation is interesting, with some theoretical analysis.
  3. The experiments cover multiple datasets and strong baselines, showing promising improvements over existing methods.
Critical weaknesses
  1. The method still shows performance degradation under strong noise and reverberation, which raises concerns about its robustness in real-world environments.
  2. The new VDD and FVD datasets only cover Hungarian speech and seven generation methods, limiting the generality.
Questions for authors

1.How does the spectral gate behave under different noise levels and acoustic conditions? Does it actually reduce spectral injection when the spectral features become unreliable? 2.Can the authors provide more analysis of possible generation or reconstruction artifacts in the newly constructed datasets?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ioit (1/2): Robustness to noise and reverberation; scope of VDD and FVD

We thank the reviewer for the time spent on our paper and for recognizing SILO's motivation, theoretical analysis, and cross-dataset results. Line, table, and appendix numbers below refer to the revised PDF.

W1. Performance degradation under strong noise and reverberation.

We explicitly simulate, evaluate and discuss this sensitivity in App. F.3 and Figs. 9-10: 42 transformations in six families (Tab. 11), including six noise and reverberation settings and three echo settings, applied identically to all 13 detectors on three datasets. The largest effects occur at the most severe settings (RT60=900\mathrm{RT60} = 900 ms, 0 dB SNR), and we attribute them to the spectral evidence SILO relies on (L2091-2098). This is consistent with our design motivation.

The same analysis shows SILO's strength. In four of the six families (codec, filtering, dynamics, quantization), SILO remains among the most robust across all three datasets, with 7.94M trainable parameters, while the spectral and raw-waveform baselines significantly lose performance. Under noise and reverberation, the strongest baselines also degrade, including the state-of-the-art SONAR (Fig. 9).

Therefore, this section should be read as a systematic analysis of robustness: it identifies the hard conditions, explains their link to spectral evidence, and compares SILO with every competing detector under one protocol. Reporting a limitation with its mechanism is part of the contribution, and the assessment should weigh it together with the full comparative results.

W2. Hungarian only, seven generation methods.

(1) We explain the purpose of constructing the two datasets in Sec. 1, L090-098; Sec. 2, L122-130; Sec. 4, L273-278; and App. E.1, L1774-1786 and L1801-1807. Specifically, VDD and FVD are evaluation datasets that isolate two conditions missing from existing benchmarks: pathological speech (VDD) and long-form recordings (FVD). These two datasets are not designed to test cross-language or cross-generator generalization. Using one language and one attack panel keeps the condition under test as the only factor that varies. The source corpora were also obtained under data-use agreements (Ethics Statement). These agreements are tied to a specific research project, which we cannot describe without breaking ICLR's double-blind policy; we will state it in the camera-ready version.

(2) Cross-language generalization is evaluated elsewhere in the paper, on three multilingual datasets: CVoiceFake (5 languages), PolyGlotFake (5 languages) and MLAAD-v3 (23 languages, of which 5 are reported language-wise) (App. F.5, L2178-2196; Fig. 12). Cross-generator generalization is evaluated on the 116 systems of MLAAD-v3 (Tab. 1; Figs. 4 and 11).

AuthorsOpenReview ↗

Response to Reviewer ioit (2/2): Learned gate behaviour; generation and reconstruction artifacts

Line, table, and appendix numbers below refer to the revised PDF.

Q1. Learned gate behaviour under noise.

The gate learns from the detection objective. Proposition 3 (App. C.6) shows that its local training signal favors stronger spectral injection when this reduces classification loss and weaker injection when it increases loss (L1409-1417):

∂ℓx∂a=(1−p) g(1−g) [ρ(f^θ(x,g))−y] ∂f^θ(x,β)∂β∣β=g\frac{\partial \ell_x}{\partial a} = (1-p)\,g(1-g)\,\big[\rho(\hat f_\theta(x,g)) - y\big]\,\frac{\partial \hat f_\theta(x,\beta)}{\partial \beta}\Big|_{\beta=g}

Proposition 4 shows that lowering the gate proportionally reduces the injected contribution for fixed parameters and spectral features, ∥Iβ(z)∥2≤β∥αc∥∞∥z∥2\lVert I_{\beta}(z)\rVert_{2} \le \beta \lVert \alpha_{c}\rVert_{\infty} \lVert z\rVert_{2} (L1477-1482).

We evaluate this spectral gate in two settings.

  1. Under graded synthetic degradation. Figs. 9-10 (App. F.3) report SILO with the learned gate at three severity levels each of additive noise (SNR 20, 10, 0 dB), reverberation (RT60 300, 600, 900 ms) and echo (50, 150, 300 ms) on three datasets (Tab. 11); degradation is concentrated at the most severe level of each (L2091-2095).
  2. Under real-world acoustic conditions. Tab. 20 (App. F.6) compares the learned gate with fixed injection strengths. On In-the-Wild, the learned gate lowers EER from 7.05 under always-on injection (β=1\beta = 1) to 4.35, while keeping the lowest EER on ASVspoof 2019 (1.02) and MLAAD (14.38), consistent with the gate reducing injection where spectral evidence is less reliable. Q2. Generation and reconstruction artifacts.

The submission already describes controls for potential construction cues.

  • VDD applies the same final resampling and normalization to genuine and generated recordings (Sec. 4.2, L309-314).
  • FVD preserves the original timeline and duration, matches segment and recording loudness, and fills inter-segment gaps with background noise matched to the source recording. Short generated chunks are padded without time-stretching (App. E.3, L1873-1885).
  • Tabs. 15 and 17 report word error rates and speaker similarities to characterize linguistic fidelity and speaker consistency.
  • App. F.2 and Figs. 7-8 additionally analyze spectral and harmonic patterns in VDD, illustrating how synthesis can regularize pathological speech and how genuine pathological characteristics can resemble synthesis artifacts. Together, these sections provide construction controls, generation-quality diagnostics, and analysis of the resulting acoustic characteristics.

If these analyses are insufficient, could you specify which additional artifacts should be examined and what further analysis you consider necessary?

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer ioit,

As the discussion period is closing soon, we would like to kindly follow up on our response regarding robustness under noise and reverberation, the scope of VDD and FVD, and the behaviour of the learned gate.

Please let us know if any concern remains. We would appreciate your reconsideration of your assessment and rating in light of these responses.

Thank you for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Contribution, Rebuttal, and Revisions (to the AC and all Reviewers)

We sincerely thank the Area Chair and all three reviewers for their time and their detailed feedback. This comment states our contribution and how each concern was addressed. All numbers refer to the revised PDF.

1. Contribution

SILO shows that spectral and SSL detectors fail in complementary ways and resolves this with an utterance-gated low-rank spectral-injection branch, adapted on restricted Stiefel manifolds so that the selected pretrained singular directions are provably left exactly unchanged (Theorem 1). We also introduce VDD and FVD, two evaluation datasets for pathological speech and multi-minute forensic recordings, which expose failure modes that existing benchmarks do not measure.

2. The revision and rebuttal addressed the reviewers' concerns

We responded to every weakness and question from the three reviewers.

New requested experiments. Beyond the comparison with 13 baseline detectors and 3 LoRA-based methods on 9 datasets and the ablations of the constraints, preserved energy, injection rank, spectral front-end and gate in our submission, we added 2 experiments requested by the reviewers:

  • Branch ablation (QXrw W2/Q2, 9wTR; Sec. 5.2, Tab. 6). The spectral branch alone fails on In-the-Wild (67.69% EER) and the SSL branch alone fails on MLAAD (26.83%). Full SILO has the lowest EER on ASVspoof 2019, ASVspoof 2021 and MLAAD (1.02 / 1.56 / 14.38%), with 4.35% on In-the-Wild.
  • Computational cost (QXrw W5/Q7; Sec. 5.2, Tab. 7). SILO adds only 2.55% forward computation over XLS-R+LoRA (153.88 vs. 150.05 GFLOPs), and its spectral encoder has 56K frozen parameters.

New requested documentation. Beyond the implementation details in our submission, we added 4 appendix sections requested by the reviewers:

  • Data, preprocessing and audio duration (QXrw W3/Q3, Q8; App. D.1; Fig. 6). All detectors share one 70/10/20 WaveFake split, the same 16 kHz, 4.04 s input and the same checkpoint rule. FVD scores one prefix of up to 600 s in a single forward pass with temporal mean pooling; the caption of Fig. 6 now states this.
  • Baseline implementations (QXrw W5/Q6; App. D.2). Architecture, objective and optimization settings of every baseline.
  • LoRA-based baselines (QXrw W5/Q7; App. D.3, Tab. 9). Ranks, optimizers and parameter budgets. All LoRA baselines adapt the same 96 projections as SILO.
  • Generation systems of VDD and FVD (QXrw W4/Q4; App. E.4, Tab. 18). Our submission already names, cites and describes all 7 systems (Sec. 4.2). At the reviewer's request, we consolidated this in one table listing the mechanism, Hungarian support, public release and license of each system.

Corrected statement. Reviewer QXrw was right that the learned gate is not the lowest on every dataset (W6; App. F.6, Tab. 20). The text now states that it gives the best overall balance between cross-generator generalization and robustness to acoustic variation.

Concerns beyond the claims and scope of the submission. We answered these with evidence as well:

  • High-frequency learning, toy problem and proof (9wTR). SILO places no frequency constraint between the low-rank factors and makes no claim about learning high frequencies: diag(α)\mathrm{diag}(\alpha) holds signed singular values (Prop. 1) and the LFCC front-end is linear in frequency. The contribution of the spectral branch is shown in Tabs. 6, 19, 20 and Figs. 4, 11.
  • Relation of VDD/FVD to SILO (QXrw W1/Q1). The method and the datasets are independent contributions (Sec. 1, Sec. 4.1).
  • Hungarian only, 7 generators (ioit W2, QXrw W4). VDD and FVD hold language and generators fixed to isolate pathology and duration. Cross-language and cross-generator generalization are evaluated on CVoiceFake, PolyGlotFake and MLAAD-v3 (App. F.5).
  • Additional benchmarks (QXrw W6/Q5). The submission already evaluates 9 datasets, including the 116 systems and 23 languages of MLAAD-v3, plus 42 audio transformations.

Clarifications supported by results in the submission.

  • Noise and reverberation (ioit W1; App. F.3, Figs. 9, 10). Degradation is confined to the most severe settings (RT60 = 900 ms, 0 dB SNR), and SILO is among the most robust of the 13 detectors in four of six transformation families.
  • Gate behaviour under noise (ioit Q1; App. C.6, Props. 3, 4; Tab. 20). Letting the gate choose the injection strength for each utterance gives a lower EER than always injecting spectral evidence at full strength, on all three datasets (ASVspoof 2019: 1.02 vs. 1.21%; In-the-Wild: 4.35 vs. 7.05%; MLAAD: 14.38 vs. 15.89%).
  • Generation and reconstruction artifacts (ioit Q2). Construction controls (App. E.3), WER and speaker similarity (Tabs. 15, 17), voice-quality measures (Tab. 13) and spectral analysis (App. F.2).
AuthorsOpenReview ↗

Kind Reminder to reviewers before the discussion deadline

Dear Reviewers,

Thank you again for the time and care you have put into reviewing our submission.

With the discussion period closing soon, we would like to kindly draw your attention to the detailed responses we have posted to each review, together with the revised manuscript. We would greatly appreciate it if you could read them and update your assessment and scores accordingly.

If anything remains unclear, please let us know, and we will respond immediately within the remaining time.

Thank you for being so considerate.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.