acceptodds
Under review as a conference paper at ICLR 2027

MoLN: Mixture of Layer Normalization Experts for Speech Deepfake Detection

Abstract

Speech deepfake detection (SDD) requires strong generalization ability due to the diverse and heterogeneous factors affecting speech generation, transmission, and recording conditions. Recent SDD systems increasingly leverage large pre-trained speech encoders and parameter-efficient fine-tuning (PEFT) to adapt transferable speech representations, with LoRA-based approaches becoming a common adaptation strategy. In this work, we investigate LayerNorm (LN) tuning as an alternative adaptation interface for generalizable SDD. Different from LoRA, LN tuning performs compact channel-wise feature recalibration, providing a favorable trade-off between parameter efficiency and generalization while better preserving pre-trained representations. We further show that a single shared LN is insufficient to model the heterogeneous and overlapping conditions in SDD. To address this limitation, we propose Mixture of LayerNorm Experts (MoLN), a parameter-efficient framework that extends LN tuning from a single affine parameter to multiple lightweight LN experts. MoLN employs a frame-level soft router to dynamically combine expert-specific scale and shift parameters, enabling adaptive feature recalibration without requiring explicit domain labels or additional task-specific modules. Extensive experiments demonstrate that MoLN consistently improves cross-dataset generalization over state-of-the-art baselines with fewer trainable parameters.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Speech processing, Speech deepfake detection, Large pretrained model, parameter-efficient fine-tuning, Layer normalization, Mixture of experts

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_t1ZFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper propose MoLN, a parameter efficient finetuning method for speech deepfake detection (SDD). They first elaborate the necessity of using PEFT for robust and generalizable SDD, then argue that using LN-only finetuning is more effective than LoRA. They further introduce MoE-alike router with multiple LN expert jointly finetuned for better performance.

Critical strengths

The paper is clearly written and easy to follow.

The empirical evaluation is well-targeted and relevant to the central hypothesis.

Critical weaknesses

Concerns on the soundness of the experiments

The preliminary anlysis shows that the representation CKA before and after LoRA PEFT is smaller than LN. My concern is that is this affected largely by learning rate? Have you searched for the optimal configurations for both methods before comparing them?

Limited Significance and Narrow Scope.

While I acknowledge that the proposed method outperforms LoRA, I find the magnitude of the improvements relatively modest. Moreover, the scope of the work appears somewhat narrow, which further limits its potential impact. Overall, I am not fully convinced that the contributions are substantial enough to meet the expectations of a venue like ICLR.

Questions for authors

Please feel free to address the questions in the above weaknesses part.

Additional questions:

Why did you choose to train on ASV2019 rather than other datasets? Will switching to another corpus for training make any differences?

7 comments
AuthorsOpenReview ↗

Author Response (1/2): W1

Thank you for raising questions about the fairness of the LN–LoRA comparison, the magnitude of the gains, and the choice of training corpus. We clarified the training settings used in the CKA analysis, added a comparison of LoRA training configurations, and expanded the evaluation across benchmarks, architectures, and languages.

W1: Training configurations selection.

A1: As detailed in Appendix D.2, LN tuning uses Adam with a learning rate of 10−410^{-4}. For LoRA, we evaluated Adam and AdamW configurations at ranks 4, 8, and 16. The AdamW configuration in Table 2 uses a cyclic learning rate from 10−610^{-6} to 10−410^{-4}. Prior LoRA studies also use AdamW with a cyclic schedule for XLSR-53–AASIST trained on ASVspoof 2019 LA (Laakkonen et al., 2025, 2026). This configuration achieves a lower ASVspoof 2021 DF EER at all three ranks tested with both optimizers.

LoRA training configurations on ASVspoof 2021 DF

LoRA rank (Q, K, V, out)Trainable params. (M)Adam EER (%) ↓AdamW EER (%) ↓
41.2346.982.82
82.0206.665.02
163.5936.224.52

On the same benchmark, Laakkonen et al. (2025) report 4.37% EER for rank-4 LoRA, and Laakkonen et al. (2026) report 4.35 ± 1.37% for rank-16 LoRA trained with empirical risk minimization. Under the reported training settings, LN tuning achieves higher CKA than LoRA at every rank in Table 2. The difference also holds for rank-4 LoRA, which achieves lower 21DF EER than LN tuning (2.82% vs. 3.75%), while LN tuning retains greater similarity to the pre-trained representation (CKA of 0.1406 vs. 0.0894).

References

  • Laakkonen, J., Kukanov, I., & Hautamäki, V. (2025). Mixture of Low-Rank Adapter Experts in Generalizable Audio Deepfake Detection. In Proceedings of the 2025 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), pp. 2211–2216. IEEE.
  • Laakkonen, J., Kukanov, I., & Hautamäki, V. (2026). Generalizable Speech Deepfake Detection via Meta-Learned LoRA. In Proceedings of the 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 19092–19096. IEEE.
AuthorsOpenReview ↗

Author Response (2/2): W2, Q1

W2: How significant are the improvements, and how broadly do they apply?

A2: Voice impersonation is a practical speech security problem. In 2024, an Arup employee in Hong Kong transferred approximately US$25 million after fraudsters impersonated executives using fabricated voices and images on a video call (Magramo, 2024). In 2026, Greek police reported that a fraud ring used AI-cloned voices to impersonate victims’ relatives over the phone (Karagiorgas, 2026). At deployment, in contrast to other fields, shifts in synthesis methods, codecs, and recording conditions can coincide with changes in spoken language (Yamagishi et al., 2021; Marek et al., 2025).

Under this scenario, we ask a question whether a compact adaptation of a pre-trained speech encoder can generalize across these shifts using only source-domain data for adaptation. Prior work on multimodal adaptation motivates LayerNorm as a compact interface for feature recalibration (Zhao et al., 2024). In Section 3.3, different attack and bandwidth conditions favor different local LayerNorm updates, and the top two directions explain 94.6% of the variation in those gradients. This finding motivates MoLN’s small bank of residual affine experts, which are mixed at each frame without domain labels or test-time parameter updates.

Experimental results show that the proposed method is effective. Across three seeds and six benchmarks, MoLN reduces mean EER from 13.62% to 12.90% relative to reproduced LoRA while using 75% fewer trainable parameters (0.90M vs. 3.59M). On In-the-Wild, EER drops from 10.43% to 7.17%. We also evaluate MoLN across encoder–detector architectures and five MLAAD languages, and routing interventions show that the learned frame-level weights contribute to detection performance.

We hope this work draws more attention to speech security and gives the community a parameter-efficient method for studying detection under domain shift.

Q1: Why did you train on ASVspoof 2019 LA, and do MoLN’s gains hold with a more diverse training corpus?

A1: We chose ASVspoof 2019 LA because it is a widely used training source for speech deepfake detection and enables comparisons with prior work. Its training split contains four text-to-speech and two voice-conversion attacks, providing variation in how spoofed speech is generated.

With this training source held fixed, we evaluated transfer across datasets that differ in synthesis methods, recording conditions, and compression. We also added a language-stratified evaluation across five MLAAD languages. MoLN achieves 94.69% macro-balanced accuracy with a threshold selected on the ASVspoof 2019 LA development set and applied unchanged across languages.

To test whether MoLN’s gains persist with a more diverse training corpus, we are retraining MoLN and key baselines on the larger ASVspoof 5 training split using the same set of training seeds. The multi-seed comparison is in progress, and we plan to report the results and analysis in a subsequent revision.

References

  • Magramo, K. (2024, May 17). British engineering giant Arup revealed as $25 million deepfake scam victim. CNN Business.
  • Karagiorgas, I. (2026, September 18). Greece: Police bust gang using AI to clone relatives’ voices for cash. Euronews.
  • Marek, B., Kawa, P., & Syga, P. (2025). Are audio DeepFake detection models polyglots? In Proceedings of the 5th Symposium on Security and Privacy in Speech Communication (SPSC), pp. 48–55.
  • Yamagishi, J., Wang, X., Todisco, M., Sahidullah, M., Patino, J., Nautsch, A., Liu, X., Lee, K. A., Kinnunen, T., Evans, N., & Delgado, H. (2021). ASVspoof 2021: Accelerating progress in spoofed and deepfake speech detection. In Proceedings of the 2021 Edition of the Automatic Speaker Verification and Spoofing Countermeasures Challenge, pp. 47–54.
  • Zhao, B., Tu, H., Wei, C., Mei, J., & Xie, C. (2024). Tuning LayerNorm in Attention: Towards Efficient Multi-Modal LLM Finetuning. In Proceedings of the 12th International Conference on Learning Representations (ICLR).
AuthorsOpenReview ↗

Follow-up to Q1

Follow-up to Q1 — Effect of changing the training corpus

We chose ASVspoof 2019 LA for the main experiments to enable comparison with prior work under a common training protocol. To assess the effect of changing the source corpus, we trained MoLN and four baselines on the same speaker-matched, class-balanced 16,000-utterance subset of ASVspoof 5 and evaluated them on 21DF, In-the-Wild, and Fake-or-Real.

Across three training seeds, MoLN achieves the lowest mean EER on each external benchmark. Its three-benchmark average is 14.31 ± 2.39%, compared with 21.98 ± 2.99% for LN-Tune and 28.30 ± 10.79% for LoRA. On In-the-Wild, MoLN achieves 7.76% mean EER compared with 14.12% for LN-Tune. These results show that the relative gains persist when training on the ASVspoof 5 subset. The new Appendix J reports the full comparison and experimental protocol.

Reviewer_RMX5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a parameter-efficient adaptation of frozen encoders, specifically XLS-R and WavLM, for speech deepfake detection. It first compares LayerNorm tuning with LoRA and argues that LN tuning preserves the pre-trained representation better at a lower parameter cost. It then proposes a new method, Mixture of LayerNorm Experts (MoLN) The idea is simple, easy to adopt, and cheap, but the evaluation is insufficient by community standards (one or two seeds).

Critical strengths
  • Simple idea that is easy to add to a frozen SSL front-end
  • Broad ablations of the method, including expert count, routing, and analysis of learned experts
  • ADD 2023 (Mandarin) is included (for not all configurations), which is a first step towards a cross-lingual evaluation
Critical weaknesses

W1. Table 3 uses only two seeds, all ablations are single runs (Sec. 5.3), and Table 4 reports no standard deviation and I could not find the number of runs. Recent works in the field use three [1,2] or more seeds [3]. Moreover, with two runs, the reported '± std' carries no more information than the difference between the two runs.

W2. It is unclear how the authors selected K and λ. Specifically, 1) the default K also differs between backbones, 2) the authors claim the parameters were selected on the ASVspoof 2019 LA dev subset, which seems to perform very well (Tab.2 ID EER 0.000-0.079%). Therefore, if the EER is this close to zero for all K and λ, it cannot tell them apart, and any choice would be a tie. As shown in Tab.5, on the test sets, the chosen setting (K = 3, λ = 0.1) clearly beats all other tested values. Please report the development-set numbers for every K and λ and the rule used to pick among them.

W3. The motivation for LN tuning over LoRA (Sec. 3.2, contribution 1) is not supported by own results. First of all, Table 2 is underreported; specifically, the paper does not say which sets are ID and OOD or how many runs it reports. Within Table 2, LoRA has a lower ID EER at r∈{2,8,16}r \in \{2, 8, 16\} , and a lower OOD EER at r=4r = 4. The remaining OOD gaps are of the size of the seed spreads shown in Table 3. Furthermore, in Table 3, LoRA beats LN tuning on 21DF, FoR, and ASV5, on average.

W4. EER sets the optimal threshold on each test set using its labels, which a practitioner does not have. For that reason, the MLAAD authors report accuracy. Under dataset shift, a model can keep a low EER while its source-set threshold fails. Additionally, please report balanced accuracy at a threshold fixed on source data (the threshold shouldn't be optimized on the eval set).

W5. MoLEx [4] is cited in Sec. 1 and 2.1 but not compared, and LoRA-MoE [5] is reproduced without its noisy gating, without explaining the reason for this design choice.

W6. I miss some broader analysis of linguistic factors (mentioned in Sec. 1, ~L034-035) as factors influencing DF detection; as shown in [7], they do influence DF detection. (Additionally) I would appreciate a similar, but brief, analysis in cross-lingual settings, e.g., using data from [6]. Such an analysis would provide the community with a broader understanding of PEFT methods for SDD in multilingual settings.

W7. The paper seems to overclaim and oversell the results, such as Consistently outperforms (which is not the case for some datasets), and Extensive means one backbone used for the main results, one training set, and two seeds. Additionally, my concern is that all models are trained on ASVspoof 2019 LA only: six attacks from 2019 or earlier on clean English read speech. This is the common protocol and allows comparison with prior numbers, but it contains little of the heterogeneity that motivated MoLN. Does the gain of MoLN over other baselines hold when training on a more diverse source, e.g., the ASVspoof 5 training set or MLAAD [6]?

Bibliography:

[1] Müller, Nicolas, et al. "Replay attacks against audio deepfake detection.", Proc. Interspeech 2025, 2245-2249, doi: 10.21437/Interspeech.2025-20

[2] Tak, Hemlata, et al. "Automatic Speaker Verification Spoofing and Deepfake Detection Using Wav2vec 2.0 and Data Augmentation.", Proc. The Speaker and Language Recognition Workshop (Odyssey 2022), 112-119, doi: 10.21437/Odyssey.2022-16

[3] Laakkonen, Janne, et al. "Generalizable speech deepfake detection via meta-learned LoRA." ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026.

[4] Pan, Zihan, Sailor Hardik Bhupendra, and Jinyang Wu. "Molex: Mixture of lora experts in speech self-supervised models for audio deepfake detection." 2025 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2025.

[5] Laakkonen, Janne, Ivan Kukanov, and Ville Hautamäki. "Mixture of low-rank adapter experts in generalizable audio deepfake detection." 2025 Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). IEEE, 2025.

[6] Müller, Nicolas M., et al. "Mlaad: The multi-language audio anti-spoofing dataset." 2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024.

[7] Marek, Bartłomiej, Piotr Kawa, and Piotr Syga. "Are audio DeepFake detection models polyglots?", Proc. 5th Symposium on Security and Privacy in Speech Communication, 48-55, doi: 10.21437/SPSC.2025-8

Questions for authors

Why is ADD 2023 missing from Table 3?

Which ASVspoof 5 partition is used?

The rest is formulated in the Weakness section.

4 comments
AuthorsOpenReview ↗

Author Response (1/3) — W1–W3

Thank you for the constructive feedback and for identifying where the evaluation needed stronger evidence and clearer reporting. We expanded the main and architecture comparisons to three training seeds, clarified hyperparameter selection, and added baselines, benchmarks, and a five-language MLAAD evaluation. We also report balanced accuracy at a source-fixed threshold and revised our claims to reflect performance across individual benchmarks.

W1: Strengthen multi-seed evaluation and uncertainty reporting.

A1: We expanded the main XLSR-AASIST comparison and the experiments across encoder–detector architectures to three independent training seeds (Tables 3 and 4). We report the mean ± sample standard deviation across seeds for each benchmark and compute the six-benchmark average within each seed before aggregating across seeds. MoLN achieves the lowest average EER (12.90 ± 0.99%), compared with LoRA (13.62 ± 0.54%) and LN tuning (16.77 ± 1.46%). The design ablations report single-run results.

W2: Clarify the selection of K and λ given the near-zero development EER.

A2: We select K and λ by minimizing the loss on the ASVspoof 2019 LA development set (Appendix D.3). Development loss distinguishes configurations with near-zero EER.

Expert-count selection

ArchitectureKDevelopment loss
XLSR-AASIST10.000213
XLSR-AASIST30.000126
XLSR-AASIST100.000357
WavLM-AASIST20.000703
WavLM-AASIST30.001235
XLSR-Conformer-TCM20.000368
XLSR-Conformer-TCM30.000229

Regularization-weight selection on XLSR-AASIST (K = 3)

λ_orth = λ_lbDevelopment loss
0.010.000353
0.10.000126
0.20.000132

The same criterion selects K = 3 for both XLS-R configurations and K = 2 for WavLM-AASIST. Each comparison has a unique minimum. We fixed these choices using source development data before external evaluation.

W3: Explain the motivation for LN tuning given that LoRA achieves lower EER in some comparisons.

A3: The comparison in Section 3.2 supports choosing LayerNorm as a compact interface for preserving pre-trained feature structure. The representation analysis in Section 5.3 shows how MoLN builds on this property.

We compare LN tuning with LoRA at ranks 1, 2, 4, 8, and 16 using the same pre-trained XLS-R encoder, AASIST architecture, and ASVspoof 2019 LA training partition. LN tuning updates the LayerNorm scales and shifts, while LoRA adapts the attention projections. As clarified in Appendix A, Table 2 reports EER from one run per configuration on ASVspoof 2019 LA dev (ID) and ASVspoof 2021 DF (OOD). We compute the representation diagnostics on the same 5,096 source-development utterances for all checkpoints.

Linear CKA [1] measures the similarity between the pre-trained and adapted representation geometries after temporal mean pooling. CKA has also been used as a training regularizer in speech deepfake detection [2]. Our relative feature basis preservation (rFBP) metric measures how strongly correlations across aligned frames concentrate on matching feature channels. An rFBP of 1.0 denotes the pre-trained self-reference.

LN tuning achieves the highest CKA in Table 2 with fewer trainable parameters than rank-1 LoRA (0.55M vs. 0.64M). Compared with rank-4 LoRA, the strongest LoRA configuration on ASVspoof 2021 DF in Table 2, LN tuning achieves higher CKA (0.1406 vs. 0.0894) and an rFBP closer to the pre-trained reference (1.0469 vs. 0.8576). These results motivate using LN for compact feature recalibration. The condition-specific gradient analysis in Section 3.3 motivates extending shared LN tuning to a small bank of experts with frame-level routing.

In Section 5.3, we fit PCA to the pre-trained representations and project the representations from all checkpoints onto the same axes. MoLN changes the projected distribution while retaining substantial overlap with the pre-trained distribution and a similar spread. Together with MoLN’s rFBP of 1.0172 (Appendix A), this supports its ability to recalibrate features while retaining aspects of their pre-trained structure.

References

[1] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. 2019. Similarity of Neural Network Representations Revisited. ICML, PMLR 97, pp. 3519–3529.

[2] Hoan My Tran, Damien Lolive, Aghilas Sini, Arnaud Delhay, Pierre-François Marteau, and David Guennec. 2025. Multi-level SSL Feature Gating for Audio Deepfake Detection. ACM Multimedia, pp. 11766–11775.

AuthorsOpenReview ↗

Author Response (2/3) — W4–W5

W4: Report balanced accuracy with a threshold fixed before evaluation.

A4: We now report balanced accuracy (BAcc), false acceptance rate (FAR), and false rejection rate (FRR) under two decision rules on MLAAD (Appendix I), complementing EER and following Müller et al. (2024) in using accuracy-based evaluation.

For source-fixed decisions, we select each threshold on the ASVspoof 2019 LA development set and apply it unchanged across all five MLAAD languages (Korshunov and Marcel, 2016). For argmax decisions, we predict the class with the higher output score, equivalent to a 0.5 probability threshold for our binary softmax classifier (Huang et al., 2026). Under both rules, MLAAD labels are used only to compute evaluation metrics.

With the source-fixed threshold, MoLN achieves the highest macro BAcc among PEFT methods (94.69%), ahead of reproduced LoRA (93.71%) and close to full fine-tuning (94.75%) with 0.90M trainable parameters. Switching MoLN from argmax to the source-fixed threshold lowers FRR from 9.47% to 6.23% and raises FAR from 2.01% to 4.40%. Under argmax, LoRA+MoLN achieves the highest macro BAcc (94.85%) and the lowest FRR (6.47%). Section 5.2 reports the language-stratified results, and Appendix I provides additional details.

MLAAD macro results (%). Mean ± sample standard deviation over three training seeds. Macro averages give equal weight to the five languages. Bold marks the best mean in each column.

MethodParams. (M)Source-fixed BAcc ↑Source-fixed FAR ↓Source-fixed FRR ↓Argmax BAcc ↑Argmax FAR ↓Argmax FRR ↓
Full fine-tuning317.8494.75 ± 1.693.83 ± 3.166.68 ± 1.9492.14 ± 2.881.51 ± 1.2014.21 ± 6.58
LN-Tune0.5588.97 ± 3.0710.92 ± 5.5311.13 ± 5.1189.09 ± 3.265.16 ± 5.0316.67 ± 6.93
Dynamic LN1.1694.18 ± 1.035.04 ± 3.016.60 ± 0.9893.15 ± 1.221.48 ± 0.5512.21 ± 2.43
LoRA (reproduced)3.5993.71 ± 1.477.60 ± 3.524.97 ± 1.5693.86 ± 1.441.17 ± 0.2611.11 ± 3.13
LoRA-MoE (reproduced)3.4091.63 ± 1.257.57 ± 5.089.17 ± 4.4691.73 ± 1.942.55 ± 2.0914.00 ± 5.54
MoLEx5.3189.49 ± 2.913.53 ± 0.9917.48 ± 5.0884.85 ± 1.691.36 ± 0.8028.95 ± 2.70
LoRA+MoLN1.4992.84 ± 2.119.95 ± 4.794.37 ± 1.1094.85 ± 0.453.84 ± 1.156.47 ± 2.00
MoLN0.9094.69 ± 1.124.40 ± 0.646.23 ± 1.6394.26 ± 1.202.01 ± 1.969.47 ± 2.42

References

Müller, N. M., Kawa, P., Choong, W. H., Casanova, E., Gölge, E., Müller, T., Syga, P., Sperl, P., & Böttinger, K. (2024). MLAAD: The Multi-Language Audio Anti-Spoofing Dataset. IJCNN, pp. 1–7.

Korshunov, P., & Marcel, S. (2016). Cross-Database Evaluation of Audio-Based Spoofing Detection Systems. Interspeech, pp. 1705–1709.

Huang, W., Mao, Y., & Qian, Y. (2026). A Data-Centric Approach to Generalizable Speech Deepfake Detection. ACL, pp. 17520–17539.

W5: Complete the MoLEx and noisy-gating LoRA-MoE comparisons.

A5: We added MoLEx and included a reproduction of LoRA-MoE with noisy gating in the revised Table 3. We also clarified the baseline implementations in the revised manuscript. Across six benchmarks and three training seeds, MoLN achieves the lowest average EER (12.90 ± 0.99%), compared with reproduced LoRA-MoE (13.85 ± 0.81%) and MoLEx (16.19 ± 0.67%). MoLN uses 0.90M trainable parameters, compared with 5.31M for MoLEx. We also compare both expert-based baselines on MLAAD under the same decision rules.

AuthorsOpenReview ↗

Author Response (3/3) — W6–W7, Q1-Q2

W6: Analyze linguistic factors and cross-lingual generalization on MLAAD.

A6: Thank you for raising this point. In Section 5.2, we evaluate XLSR-AASIST models trained on English ASVspoof 2019 LA across five MLAAD languages. We report per-language BAcc under argmax and macro BAcc under both argmax and a threshold fixed on the source development set.

MLAAD balanced accuracy (%). Mean ± sample standard deviation over three training seeds. Macro averages give equal weight to the five languages. Bold marks the highest mean in each column.

MethodEnglishGermanFrenchSpanishItalianArgmax MacroSource-fixed Macro
Full fine-tuning94.93 ± 1.4788.03 ± 2.2095.00 ± 1.7790.57 ± 5.3892.17 ± 3.9092.14 ± 2.8894.75 ± 1.69
LN-Tune89.30 ± 2.6984.80 ± 3.8992.00 ± 3.0889.47 ± 3.9589.87 ± 3.9389.09 ± 3.2688.97 ± 3.07
LoRA94.60 ± 1.3587.80 ± 1.9196.63 ± 0.6794.87 ± 2.2295.40 ± 1.6193.86 ± 1.4493.71 ± 1.47
LoRA-MoE92.03 ± 1.5087.13 ± 5.3393.87 ± 1.5692.60 ± 1.7893.00 ± 1.3191.73 ± 1.9491.63 ± 1.25
MoLEx90.33 ± 1.1274.67 ± 1.1089.50 ± 3.1283.53 ± 1.0386.20 ± 2.3384.85 ± 1.6989.49 ± 2.91
Dynamic LN96.60 ± 1.2086.20 ± 0.9295.57 ± 0.6593.87 ± 2.1293.53 ± 1.8693.15 ± 1.2294.18 ± 1.03
LoRA+MoLN96.10 ± 0.7589.90 ± 0.9594.80 ± 0.5396.27 ± 1.3697.17 ± 0.8394.85 ± 0.4592.84 ± 2.11
MoLN94.43 ± 0.6489.70 ± 0.6996.20 ± 1.8395.33 ± 1.9695.63 ± 2.4594.26 ± 1.2094.69 ± 1.12

German has the lowest mean argmax BAcc for all methods, and MoLN’s largest gain over LoRA occurs on this subset (87.80% to 89.70%). The gap between the best and worst languages is 6.50 percentage points for MoLN and 8.83 for LoRA. LoRA+MoLN leads in German, Spanish, and Italian under argmax and achieves the highest macro BAcc (94.85%). MoLN achieves the highest source-fixed macro BAcc among PEFT methods (94.69%). Appendix I provides the FAR and FRR breakdown.

W7: Calibrate the performance claims and assess training on a more diverse source.

A7: We revised the manuscript to describe the demonstrated gains more precisely. MoLN achieves the lowest average EER (12.90 ± 0.99%) and the best mean EER on four of six benchmarks, with 0.90M trainable parameters compared with 3.59M for LoRA. The MLAAD results show that MoLN achieves the best source-fixed macro BAcc among PEFT methods and that adding LoRA can improve performance under argmax.

We chose ASVspoof 2019 LA to enable comparison with prior work using the same training source. The cross-dataset and multilingual evaluations measure transfer from this source. To test whether MoLN’s gains persist with a more diverse training corpus, we are retraining MoLN and key baselines on the larger ASVspoof 5 training split using the same training seeds. We plan to report the multi-seed comparison in a subsequent revision once it is complete.

Q1: Why was ADD 2023 missing from Table 3?

A1: ADD 2023 was previously included only in the additional architecture evaluations. We now report ADD 2023 R1 and R2 in the main XLSR-AASIST comparison, and the revised average EER covers all six benchmarks.

Q2: Which ASVspoof 5 partition is used?

A2: We use the official ASVspoof 5 Track 1 evaluation partition for standalone speech deepfake detection. We clarified this in the experimental setup.

AuthorsOpenReview ↗

Follow-up to W7

Follow-up to W7 — Training on a different source corpus

To answer the question and show the heterogeneity efficiently, we trained MoLN and four XLSR-AASIST baselines on the same speaker-matched, class-balanced subset of the official ASVspoof 5 training partition. The subset contains 16,000 utterances and covers all eight training attack IDs. We used the full official development partition for checkpoint selection and evaluated transfer to 21DF, In-the-Wild, and Fake-or-Real.

Method21DFITWFoRAvg3
Full fine-tuning19.50 ± 3.8419.00 ± 6.8133.51 ± 3.9324.00 ± 4.62
LN-Tune22.52 ± 1.5414.12 ± 1.6229.30 ± 6.3821.98 ± 2.99
LoRA20.38 ± 7.1217.84 ± 8.1146.68 ± 17.4628.30 ± 10.79
MoLEx20.81 ± 2.0715.93 ± 0.9534.13 ± 10.2923.63 ± 4.13
MoLN14.81 ± 1.647.76 ± 0.7420.36 ± 5.2014.31 ± 2.39

Results are mean ± sample standard deviation over three training seeds, with EER reported in percent. MoLN achieves the lowest mean EER on all three external benchmarks. It reduces average EER by 7.67 percentage points relative to LN-Tune, the strongest baseline on average, with 0.90M trainable parameters. This experiment shows that MoLN’s cross-dataset advantage extends to training on the ASVspoof 5 subset. The new Appendix J section provides the sampling and training details.

Reviewer_G1qYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes to tune layer normalization to adapt a model for speech deep fake detection tasks. There are two main insights: i) a single set of layer norm stats for all samples is suboptimal and therefore one needs to learn and store a set of statistics and apply them frame by frame. and ii) One can effectively mix these layer norm stats for each frame by a linear router module. The experiments have been done on suitable datasets.

Critical strengths
  • The paper is generally well written and easy to follow.
  • The aspect of layer normalization to be varied across samples / frames is interesting
Critical weaknesses
  • The method seems to be close to the Dynamic LM mentioned in Table 1, it'll be good to have this in the comparison table as well.
  • It is unclear if the routers got the mixture accurately. Fig 4 shows that the entropy is very high -- does a random expert mixture perform as well? It is unclear whether the gain in accuracy is due to increased parameter space (in the router).
  • It will be useful to contrast the method against the dynamic LM in the method section (4.1)?
Questions for authors

As above

2 comments
AuthorsOpenReview ↗

Author Response (1/1): W1–W3

Thank you for your careful review and constructive feedback. We appreciate you highlighting the connection to Dynamic LN and asking whether the learned routing contributes to detection performance. We added Dynamic LN to the three-seed main comparison and clarified the methodological distinction in Section 4.1. We also revised the routing visualization to explain the high aggregate entropy and added routing interventions and a trained static-mixture control to assess the role of input-dependent mixing.

W1: Add Dynamic LN as a baseline.

A1: We added Dynamic LN to the revised main comparison (Table 3). Originally proposed for automatic speech recognition, Dynamic LN generates utterance-specific affine parameters to adapt to varying acoustic conditions. MoLN uses frame-level routing to adapt to shifts in speech deepfake detection. We implemented Dynamic LN in the same XLSR-AASIST backbone used for MoLN and evaluated both methods on six benchmarks. The table reports EER (%) as the mean ± sample standard deviation over three training seeds. Lower is better.

MethodTrainable params. (M)21DFITWFoRASV5ADD R1ADD R2Avg.
Dynamic LN1.164.07 ± 0.3111.00 ± 1.256.24 ± 0.3916.53 ± 1.3130.51 ± 0.9031.10 ± 3.4116.58 ± 0.73
MoLN0.902.59 ± 0.367.17 ± 0.845.78 ± 2.6417.53 ± 1.2523.09 ± 0.2121.23 ± 0.8312.90 ± 0.99

Avg. is computed by averaging the six benchmark EERs within each seed before aggregating across seeds. MoLN achieves lower mean EER on five of the six benchmarks and reduces average EER from 16.58% to 12.90% with fewer trainable parameters. Dynamic LN achieves a lower mean EER on ASV5.

W2: High routing entropy. It is unclear if the routers got the mixture accurately.

A2: Thank you for raising this concern. We agree that high aggregate usage entropy alone does not establish whether routing is input-dependent or useful.

We have added Figure 5 to the main text to clarify this distinction. The example heatmaps show within-utterance variation in learned routing probabilities. Across all 50 LN sites, mean frame entropy is lower than aggregate usage entropy, showing that balanced overall expert usage coexists with non-uniform, input-dependent routing.

To examine its functional role, we replaced the routing weights in trained MoLN checkpoints while keeping the encoder, expert responses, and detector fixed. Appendix Table 14 reports results on the complete ITW and FoR test sets.

Table 1: Routing interventions on a trained MoLN checkpoint. EER on the full ITW and FoR test sets with all model parameters held fixed.

RoutingITW EER (%) ↓FoR EER (%) ↓
Learned6.713.67
Uniform35.6131.68
Source mean38.0435.43
Random37.2634.78
Shuffled35.744.88

Random routing substantially increases EER on both benchmarks with the model parameters held fixed, showing that the trained model relies on its learned routing weights. In a complementary single-seed comparison, MoLN achieves lower EER on evaluated benchmarks than an independently trained Static Multi-LN model with the same number of experts and input-independent mixing weights. Together, these results support the value of input-dependent mixing.

Table 2: Independently trained static and input-dependent mixtures with K=3K=3. EER (%) ↓. Bold indicates the lower EER in each column.

Method21DFITWFoRASV5
Static Multi-LN3.1510.7811.2217.08
MoLN2.446.713.6716.49

W3: Explain the distinction from Dynamic LN in Sec. 4.1.

A3: Thank you for the suggestion. We clarified this distinction in the first paragraph of Section 4.1:

In contrast to Dynamic LN, which generates utterance-specific affine parameters, MoLN uses frame-level soft routing to combine shared residual affine experts on top of frozen pre-trained LN parameters, allowing the adaptation to vary within an utterance.

Reviewer_G1qYOpenReview ↗

Response to authors

Thanks for taking time to produce more results and respond. My concerns are broadly addressed, though considering the discussion in this board I'll keep my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Summary of Revisions and Thanks

Dear Area Chair and Reviewers,

Thank you for the time and care you have put into reviewing our paper and discussing the results with us. Your feedback helped us strengthen the evaluation and explain the contribution more precisely. As the discussion comes to a close, we would like to summarize the main changes.

  • We expanded the main and cross-architecture comparisons to three training seeds, added Dynamic LN and MoLEx baselines, and included both ADD 2023 rounds in the main comparison. We revised the abstract and contributions to emphasize that MoLN achieves performance comparable to reproduced LoRA with approximately 75% fewer trainable parameters.
  • We clarified the distinction from Dynamic LN and added routing visualizations, inference-time interventions, and an independently trained static-mixture control. These experiments provide complementary evidence for the contribution of learned input-dependent routing.
  • We added a five-language MLAAD evaluation under both argmax and source-fixed thresholds, together with FAR and FRR. We also clarified the source-development criterion used to select hyperparameters and the order of the selection and sensitivity experiments.
  • We tested a different training source using the same balanced 16,000-utterance ASVspoof 5 subset for all compared methods. Across three training seeds, MoLN achieves the lowest mean EER on each of three external benchmarks, averaging 14.31 ± 2.39%, compared with 21.98 ± 2.99% for LN-Tune, the strongest baseline on average.

We hope the revised manuscript and additional experiments clarify the value of frame-level LayerNorm adaptation for speech deepfake detection. We appreciate your consideration of the updated evidence and thank you again for helping us improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.