acceptodds
Under review as a conference paper at ICLR 2027

STELA: Stable Temporal Evidence Learning for Audio Forgery Detection and Localization

Abstract

Audio forgery detection and localization aims to identify spoofed frames in manipulated speech and determine the forgery boundaries. Existing methods have primarily improved frame discrimination through stronger feature learning and boundary-aware modeling. However, local acoustic fluctuations within manipulated regions can still compromise prediction, leading to fragmented and discontinuous localization. We propose STELA, a Stable Temporal Evidence Learning framework for Audio forgery detection and localization to address this challenge. Specifically, for stable evidence learning, a segment-oriented acoustic augmentation and Paired State Modeling (PSM) provide complementary constraints against local acoustic variability. The augmentation introduces segment-level perturbations to discourage the detector from conflating incidental acoustic variations with manipulation boundaries. PSM further regularizes the representations within each utterance, encouraging acoustically varied frames sharing the same authenticity state to remain consistent. By mitigating representation drift and enhancing temporal coherence, these complementary constraints provide a more reliable basis for subsequent manipulation localization. Building on this improved evidence, Structured Manipulation Localization (SML) extends threshold-based proposal generation with structured scoring of alternative interval hypotheses, using interior, boundary, and contextual evidence to recover coherent manipulation intervals from fragmented predictions. Extensive experiments on PartialEdit, PartialSpoof, HAD, and LAV-DF demonstrate that STELA achieves state-of-the-art performance in both forgery frame detection and temporal localization, with strong generalization to unseen conditions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Audio forensics, temporal forgery localization, partial forgery detection

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_MFdtOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
3(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents STELA for detecting and localizing partially manipulated speech. It combines segment-oriented acoustic augmentation and boundary-free interior sampling with Paired State Modeling (PSM), an utterance-specific prototype objective that encourages consistent representations for frames with the same authenticity label. Structured Manipulation Localization (SML) then expands thresholded frame predictions into alternative interval hypotheses and estimates their quality using interior, boundary, and contextual evidence. Experiments evaluate frame detection and temporal localization on PartialSpoof, HAD, PartialEdit, and LAV-DF, with additional cross-domain frame-level evaluation including ADD2023.

Critical strengths
  1. The method delivers meaningful improvements in precise temporal localization, which supports the practical value of evaluating complete interval hypotheses rather than relying entirely on thresholded frame predictions.

  2. The ablations provide useful evidence beyond the full-system comparison. The appendix further distinguishes candidate coverage from ranked-output recall and documents failure cases, making the contribution and its limitations easier to assess.

Critical weaknesses
  1. Cross-domain generalization is demonstrated for frame classification, but not for the complete localization pipeline. Table 2 evaluates frame accuracy and F1, whereas the interval-level results in Table 3 use in-domain training and evaluation. The learned SML scorer additionally consumes duration, component count, and posterior statistics, whose relationship to proposal quality may change across manipulation methods and datasets. Consequently, improved frame-level transfer does not establish that the final interval ranking remains reliable under the same shift. Freezing both stages after source-domain training and reporting target-domain localization would directly test an important part of the claimed robustness.

  2. The evidence for recovering difficult fragmentation is more qualified than the overall localization gains suggest. In Table 5, removing proposal merging reduces PartialSpoof mAP only from 83.473% to 83.170%, while boundary adjustment and learned quality estimation account for substantially larger changes. The appendix appropriately explains that difficult fragmentation is uncommon and reports a meaningful Top-5 recall gain on Long-split events. However, for Endpoint-only events, Full SML achieves 81.19% recall when all output proposals are considered but only 0.85% among the Top-5 (Table 9). Thus, generating a correctly aligned interval often does not translate into selecting it. This limits the demonstrated effectiveness on a central motivating failure pattern and warrants further analysis of the scorer's behavior under sparse interior evidence.

Questions for authors
  1. Can the authors report cross-domain localization AP/mAP with both the frame encoder and SML frozen, using the existing source/target setup and selecting all settings on source-domain development data?

  2. What causes the large all-output versus Top-5 recall gap on Endpoint-only events? Does recall improve at moderate output budgets, and are accurately aligned merged candidates systematically penalized for weak interior responses?

4 comments
AuthorsOpenReview ↗

R1: Cross-Domain Generalization of the Complete Localization Pipeline(W1, Q1)

We thank the reviewer for identifying this important gap in our evaluation. To directly assess cross-domain localization, we conduct additional experiments using PE-E1 as the source domain and PS, HAD, PE-E2, and ADD2023 as unseen target domains. Both the frame-level model and SML are kept frozen after source-domain training, with no target-domain adaptation.

Table 1. Cross-domain temporal localization mAP (%) with models trained on PE-E1.

MethodPSHADPE-E2ADD2023Avg.
CFPRF-PRN2.8636.4714.090.5313.49
SAL (CC-Unit)1.4110.4311.140.715.92
BAM (CC-Unit)1.022.672.200.591.62
UMMAFormer0.9523.7514.324.1910.80
STELA (CC-Unit)0.6715.7017.240.848.61
STELA + SML2.2647.9140.486.3324.24

*CC-Unit converts contiguous spoof-positive frames into proposals with unit confidence. *

Table 2. Cross-domain localization performance of the complete STELA + SML pipeline (%).

Target[email protected][email protected][email protected][email protected]mAPAR@1AR@5AR@20
PS5.161.720.530.142.263.755.696.15
HAD64.5847.2735.1624.7547.9150.9455.6356.17
PE-E268.4641.6911.091.5540.4850.8356.2356.88
ADD202313.704.852.912.156.3313.5119.0720.52

The results show that SML improves cross-domain localization mAP over direct frame-based localization on all four target datasets, with particularly substantial gains on HAD (15.70% → 47.91%) and PE-E2 (17.24% → 40.48%). The complete pipeline also achieves the highest mAP among the compared methods on three of the four target domains, indicating that the source-trained proposal generation and quality estimation mechanisms retain useful predictive capability under domain shift.

Nevertheless, cross-domain localization remains challenging. Absolute mAP is still low on PS and ADD2023, and CFPRF-PRN outperforms our method on PS. Moreover, the decline in AP at stricter IoU thresholds indicates that precise boundary alignment remains sensitive to domain changes. Thus, while the new experiments support the transferability of the complete SML pipeline, they do not imply that its proposal-quality relationships are fully domain-invariant. We will include these results and clarify this remaining limitation in the revised manuscript.

AuthorsOpenReview ↗

R2: Proposal Ranking under Sparse Interior Evidence (W2, Q2)

Thank you for highlighting this important limitation. We agree that generating accurately aligned proposals does not necessarily ensure that they receive sufficiently high quality scores. To further investigate this issue, we conducted additional ranking and scoring analyses on the difficult fragmentation subsets defined in Appendix A.3, using the original SML scorers.

(1) Recall under different proposal budgets. As shown in Table 1, recall increases substantially as the output budget expands. For Endpoint-only events, [email protected] increases from 0.85% at Top-5 to 40.96% at Top-20 and 75.61% at Top-50, approaching the all-output recall of 81.19%. Among correctly localized Endpoint-only events, the median rank of the first accurate proposal is 20. These results indicate that accurate intervals are often successfully generated but ranked relatively low.

Table 1. [email protected] under different output budgets (%).

SubsetTop-5Top-10Top-20Top-50All
Long-split19.4128.5243.6757.0658.96
Endpoint-only0.854.2540.9675.6181.19
Interior-pulse4.077.6817.4333.9536.41

(2) Analysis of proposal quality estimation. We further compared accurately aligned Merge proposals (IoU > 0.95) with higher-ranked, poorly localized proposals from the same utterances. For Endpoint-only events, accurate Merge proposals receive an average score of only 0.070, compared with 0.774 for higher-ranked Top-5 incorrect proposals.

To reduce the influence of proposal duration, we additionally matched candidates within the same utterance with similar lengths (ratio within [0.5, 2]). As shown in Table 2, accurately aligned Merge proposals still receive lower scores despite having substantially higher true IoU.

Table 2. Length-matched proposal comparison on Endpoint-only events.

ProposalTrue IoUMean posteriorSML scoreDuration (s)
Accurate Merge0.9970.3450.0520.927
Higher-ranked incorrect0.1710.3820.1480.900

Furthermore, the mean interior posterior and predicted quality score exhibit strong positive Spearman correlations among accurately aligned Merge proposals across all three subsets: 0.921 for Long-split, 0.710 for Endpoint-only, and 0.875 for Interior-pulse, averaged over two scorer seeds. For Long-split and Endpoint-only, this positive association persists in within-utterance analyses accounting for proposal duration, component count, and true IoU. Together with the length-matched comparisons, these findings suggest that the quality estimator may rely heavily on interior spoof responses, potentially underestimating well-aligned proposals when interior evidence is sparse.

Overall, these findings explain the substantial gap between All-output and Top-5 recall on Endpoint-only events: accurately aligned proposals are often generated but ranked lower, with their underestimated quality scores closely associated with sparse interior evidence. The substantial recall recovery at moderate proposal budgets further supports this interpretation. We will clarify this distinction between proposal generation and ranking effectiveness in the revised manuscript.

Dear Reviewer MFdt,

We thank you for the valuable feedback, which has encouraged us to further examine the cross-domain behavior of our localization framework and the limitations of proposal quality estimation. We have updated the manuscript accordingly, with all changes highlighted in blue for ease of reference.

Specifically, we added a new subsection, “Cross-Domain Temporal Localization,” to the appendix and referenced it in Section 4.2. This subsection reports AP, mAP, and AR on PartialSpoof, HAD, PE-E2, and ADD2023, with the frame encoder and SML trained on PE-E1 and kept frozen during target-domain evaluation. We also extended the existing “Diagnostic Analysis of Proposal Merging” subsection with additional proposal-budget recall results, a duration-matched comparison of proposal scores, and an analysis of the relationship between interior spoof responses and predicted quality. These analyses provide further insight into why accurately aligned proposals for Endpoint-only events may receive relatively low quality scores and consequently fall outside the top five ranked predictions.

We hope these additional experiments and analyses provide a clearer understanding of our method's effectiveness and limitations. We appreciate your time and consideration and look forward to your further feedback.

Thank you!

Thanks for the rebuttal, which addressed my major concerns. Therefore, I will maintain my inital rating toward weak acceptance.

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

The paper addresses audio forgery detection and temporal localization. It combines local segment training and interior sampling with a WavLM-based temporal model, paired state modeling of genuine and manipulated evidence, and a structured localization stage that generates and scores candidate intervals. The paper evaluates across several datasets and includes ablations of the localization pipeline. The results suggest that the structured stage improves localization beyond simple scoring of the same frame logits, although the claimed stability of evidence within manipulated spans is less directly measured.

Critical strengths
  1. The work tackles both detection and boundary localization, and its experiments cover several manipulation settings. The cross-dataset results provide evidence that the system is more than a single-benchmark optimization.
  2. The ablations isolate a meaningful contribution from structured localization. With fixed frame logits, Table 5 (p. 9) reports higher PS mAP for the full localization stage than for simple context scoring (83.47 versus 77.07), which supports a concrete benefit from this component.
Critical weaknesses
  1. The paper's organizing claim of stable temporal evidence is not yet measured as directly as its detection and localization accuracy. Qualitative examples are suggestive, but aggregate measurements of within-span confidence variation, interior gaps, or fragmentation across forgery lengths would better establish that the proposed training produces the claimed temporal behavior. This matters because stability is central to the title and motivation.
  2. The main method is difficult to parse from Figure 2 (p. 4), which is densely packed with small labels and several interacting stages. The contribution relative to existing local augmentation, prototype-based learning, and proposal refinement techniques also needs sharper separation. Some cross-domain results are mixed, so the paper should identify the settings in which the extra complexity is most useful.
Questions for authors
  1. Can the authors quantify temporal stability directly, for example by reporting within-manipulation confidence variation, interior false-negative gaps, and fragmentation at matched frame-level accuracy?
  2. Can the authors clarify which stage is responsible for the gains on each dataset, and explain the weaker cross-domain cases (including PS relative to SAL and the small DA4 versus DA1 macro-F1 difference)? A simpler Figure 2 or a step-by-step example would also help readers verify the method.
7 comments
AuthorsOpenReview ↗

R1: Quantitative Evaluation of Temporal Stability (W1, Q1)

Thank you for this insightful suggestion. We have conducted additional temporal stability analyses on PS, HAD, and PE-E1, using challenging GT subsets where at least one of SAL, BAM, or STELA failed to achieve IoU > 0.9. We evaluated posterior SD, fragmentation, internal gaps, missed duration, confidence, additional components, and longest-component coverage. Thresholds were selected at approximately matched FNR (5% on PS; 0.5% on HAD and PE-E1). On PS, SAL and STELA also achieve similar Acc (96.93% vs. 97.01%) and FPR (1.85% vs. 1.74%).

Table 1. Temporal stability analysis on common challenging GT intervals.

Dataset / MetricSALBAMSTELASTELA+SML
PS (118,977 GT)
Posterior SD ↓0.10880.12220.03760.0376
Fragmentation (%) ↓8.4136.1235.9510.329
Mean internal gap (ms) ↓231.57332.28296.59115.52
Total missed GT duration (s) ↓4,646.664,732.184,545.041,440.44
Mean confidence ↑0.85540.75340.88320.8832
Additional components / 100 GT ↓14.61814.6157.0870.331
Longest-component coverage (%) ↑90.97196.11090.03098.054
HAD (349 GT)
Posterior SD ↓0.10300.05050.00540.0054
Fragmentation (%) ↓29.5139.7423.4380.000
Mean internal gap (ms) ↓77.01166.2365.14—
Total missed GT duration (s) ↓21.4226.707.000.12
Mean confidence ↑0.93640.90770.99660.9966
Additional components / 100 GT ↓61.03217.4798.0230.000
Longest-component coverage (%) ↑88.92086.90894.35599.961
PE-E1 (1,079 GT)
Posterior SD ↓0.07810.09380.03650.0365
Fragmentation (%) ↓13.7164.9126.1170.834
Mean internal gap (ms) ↓63.9875.06138.4380.00
Total missed GT duration (s) ↓64.7224.5281.9437.46
Mean confidence ↑0.94680.91840.90700.9070
Additional components / 100 GT ↓19.6486.6737.7850.834
Longest-component coverage (%) ↑86.42595.26786.58793.204

Our STELA consistently achieves lower posterior SD across all datasets. It also reduces fragmentation and redundant components on PS, while achieving stronger continuity across multiple metrics on HAD. On PE-E1, STELA reduces fragmentation relative to SAL, but BAM performs better on several continuity metrics. This discrepancy may partly arise from threshold sensitivity: BAM's FNR at the default threshold of 0.5 is already closer to the target FNR, potentially requiring less adjustment and preserving more of its original prediction structure. In contrast, larger threshold adjustments for STELA and SAL may alter internal gaps and fragmentation patterns. Notably, STELA consistently outperforms both baselines on temporal stability metrics at the standard threshold of 0.5.

Table 2. Length-stratified fragmentation on the common challenging PS subset.

GT duration# GTSALBAMSTELASTELA+SML
0.25 s67,5911,777 (2.629%)71 (0.105%)1,250 (1.849%)11 (0.016%)
0.25–0.5 s17,2631,982 (11.481%)549 (3.180%)1,482 (8.585%)38 (0.220%)
0.5–1 s16,6242,662 (16.013%)1,930 (11.610%)1,847 (11.110%)68 (0.409%)
1–2 s12,5922,493 (19.798%)3,034 (24.095%)1,624 (12.897%)121 (0.961%)
≥2 s4,9071,095 (22.315%)1,701 (34.665%)877 (17.872%)153 (3.118%)
All challenging GT118,97710,009 (8.413%)7,285 (6.123%)7,080 (5.951%)391 (0.329%)

The length-stratified analysis shows that STELA's advantage is most evident on longer manipulations. For intervals ≥1 s, its fragmentation rate is 14.29%, compared with 20.50% for SAL and 27.06% for BAM, whereas BAM performs better on shorter intervals. SML further improves interval continuity, reducing fragmentation from 5.951% to 0.329% on PS, 3.438% to 0% on HAD, and 6.117% to 0.834% on PE-E1. On PS, longest-component coverage also increases from 90.030% to 98.054%, demonstrating SML's complementary role in assembling coherent intervals.

Overall, these results quantitatively support STELA's temporal stability motivation, particularly for longer manipulations. We will incorporate this analysis into the revised appendix.

AuthorsOpenReview ↗

R2: Stage-wise Contributions(W2, Q2)

Thank you for highlighting this limitation. Our proposed Segment-Oriented Acoustic Augmentation (denoted as DA4 in the original ablation study) and Paired State Modeling (PSM) belong to the frame-level learning stage. Both are designed to improve the robustness and temporal consistency of frame-level predictions, benefiting both detection accuracy and downstream localization. Structured Manipulation Localization (SML) further improves localization through proposal generation and quality estimation.

To quantify these contributions, we introduce BASE, which replaces DA4 with utterance-level RawBoost augmentation (DA1) and removes PSM, while keeping all other settings identical to STELA. For BASE and STELA, localization proposals are directly formed from consecutive frames predicted as spoofed at a threshold of 0.5, with a uniform confidence score of 1.

Table 3. Stage-wise in-domain frame-level detection and localization performance (%).

DatasetMethodEER↓Acc↑F1↑[email protected]↑[email protected]↑[email protected]↑[email protected]↑mAP↑
PSBASE4.651195.969796.775163.72857.57553.93351.08257.946
STELA3.322997.009397.584869.31063.54960.11057.43063.888
STELA+SML———86.44883.98880.54477.10683.473
HADBASE0.002999.997399.9983100.00099.98399.95899.81799.965
STELA0.001799.998599.9991100.00099.98799.98799.89599.984
STELA+SML———100.00099.98999.98999.91199.987
PE-E1BASE0.188899.843999.900898.56198.08396.73394.33297.655
STELA0.171099.884499.926698.90598.55297.52595.80298.236
STELA+SML———99.74599.53999.00798.03399.389

DA4 and PSM jointly improve frame-level detection and initial localization across all three datasets. On PS, they reduce EER by 1.33 points and improve mAP by 5.94 points. SML further increases mAP by 19.58 points. Similarly, on LAV-DF, SML improves mAP from 74.96% to 92.80%. In contrast, HAD and PE-E1 already exhibit near-saturated initial localization performance, leaving less room for further gains. Thus, SML is particularly beneficial when direct frame-based localization remains less reliable.

Table 4. Cross-domain frame-level performance (%), trained on PE-E1.

TargetBASE Acc / F1STELA Acc / F1
ADD202367.60 / 79.0377.69 / 86.68
PE-E285.81 / 90.6387.95 / 92.23
HAD82.81 / 89.0389.69 / 93.82
PS71.89 / 80.2169.53 / 79.92

Compared with BASE, STELA achieves higher Acc and F1 on three of the four cross-domain datasets, with particularly notable improvements on ADD2023 and HAD. Despite a modest performance decline on PS, the overall results demonstrate the positive contribution of our proposed DA4 and PSM to improving cross-domain generalization.

Overall, these results clarify the stage-wise contributions of STELA. DA4 and PSM jointly improve frame-level detection performance, as reflected by better EER, Acc, and F1 across in-domain and most cross-domain evaluations. Beyond frame-level classification, they also enhance initial localization quality, as demonstrated by improvements in AP and mAP even without SML. Building upon these more reliable frame-level predictions, SML further optimizes interval localization through structured proposal generation and quality estimation, delivering substantial gains on PS and LAV-DF, while providing smaller improvements on HAD and PE-E1, where initial localization is already near saturation.

AuthorsOpenReview ↗

R3: Technical Distinctions and Beneficial Settings (W2)

Thank you for this comment. We would reclarify the distinctions between our designs and existing techniques:

Segment-Oriented Acoustic Augmentation (DA4) and Paired State Modeling (PSM) jointly improve temporal evidence learning through coordinated data- and representation-level strategies. Unlike segment-mixing methods such as SAL's CSM, which recombines speech segments to diversify manipulation patterns, DA4 introduces label-preserving local acoustic perturbations with cross-fading and boundary-free interior sampling. This exposes the model to diverse acoustic conditions within the same authenticity state. PSM complements this process by encouraging frames with different acoustic characteristics but the same authenticity state to maintain consistent representations, using dynamically estimated utterance-specific prototypes and permutation-invariant supervision rather than global class prototypes. More importantly, we recognized the potential synergy between DA4 and PSM and deliberately integrated them into a coordinated learning framework, where DA4 introduces label-preserving acoustic variability and PSM encourages consistent authenticity-state representations despite such variability. Structured Manipulation Localization (SML) goes beyond conventional proposal refinement by constructing complete-interval hypotheses across fragmented predictions and evaluating alternatives using interior, boundary, and contextual evidence.

To clarify their individual contributions, we provide additional component ablations. BASE uses utterance-level RawBoost augmentation without PSM, with all other settings unchanged. Initial localization uses contiguous spoof-positive frames as proposals with uniform confidence, without SML.

Table 5. Component-wise detection and initial localization performance (%).

DatasetMethodEER↓Acc↑F1↑mAP↑
PSBASE4.651195.969796.775157.9456
BASE w/ PSM4.773496.097796.895961.4092
BASE w/ DA44.259496.414697.140563.4794
STELA3.322997.009397.584863.8882
HADBASE0.002999.997399.998399.9647
BASE w/ PSM0.010199.996299.997799.9361
BASE w/ DA40.002199.998499.999199.9749
STELA0.001799.998599.999199.9845
PE-E1BASE0.188899.843999.900897.6551
BASE w/ PSM0.217599.848599.903797.9897
BASE w/ DA40.164099.872299.918898.0914
STELA0.171099.884499.926698.2362

Both DA4 and PSM independently improve initial localization on PS and PE-E1, even when frame-level EER does not consistently improve. On HAD, individual gains are limited by near-saturated performance, while full STELA achieves the best overall results.

Regarding beneficial settings, the results in Appendix.4 also show that SML provides substantial additional localization gains on PS (+19.58 mAP) and LAV-DF (+17.83 mAP), where direct frame-based localization remains less reliable, but smaller gains on HAD and PE-E1.

Overall, DA4 and PSM provide complementary benefits to frame-level evidence learning. DA4 encourages the model to distinguish genuine authenticity changes from label-preserving acoustic perturbations, benefiting both frame-level detection and initial localization. PSM further promotes consistent representations within each authenticity state, improving localization even when frame-level EER slightly deteriorates. Their combination therefore improves both frame-level detection and downstream localization, while SML further refines interval localization when the initial predictions remain fragmented or misaligned.

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

This paper works on partial-spoof audio detection and localization. The starting observation is that frame-level predictions tend to break up inside manipulated regions, which hurts localization. To deal with this, the authors put together several pieces: segment-level acoustic augmentation that also draws samples from inside manipulated regions, a frozen SSL backbone equipped with adapters and bidirectional state-space layers, a consistency loss built on two prototypes per utterance, and a localization stage that grows thresholded segments into endpoint and merged candidates before a learned scorer ranks them. They test on four partial-spoof benchmarks, reporting both in-domain and cross-domain results for detection and localization under strict IoU thresholds, ablate each component, and include an appendix analysis of the localization stage.

Critical strengths
  1. The empirical work is broad and careful: retrained baselines, in-domain and cross-domain evaluation, strict-IoU localization metrics, released code, and an honest diagnostic of where the localization stage helps. This is what makes the large localization gain credible.
  2. Scoring alternative interval hypotheses instead of inheriting a fixed segmentation is a clean formulation, and the gains land at strict overlap thresholds, where precise alignment matters most.
Critical weaknesses
  1. The frame-level gain is hard to attribute. The margin over the best baseline is small, the backbone adaptation alone moves results by more, and there is no architecture-only control with the proposed pieces removed.
  2. No variance is reported. The frame encoder is trained with a single seed, and the differences involved are small enough that this matters.
  3. The motivation and the results do not line up. Fragmented long manipulations turn out to be rare, merging barely changes aggregate numbers, and the localization gain comes from the learned proposal scoring instead.
  4. The remaining components are incremental, and the claim that acoustic fluctuations cause prediction interruptions is shown only through example plots.
Questions for authors
  1. Which SSL front end and adaptation does each retrained baseline use, and can a baseline be reported with the same backbone as STELA?
  2. Can you report multi-seed means and standard deviations for the main detection and ablation tables, and an architecture-only baseline without the proposed augmentation or consistency loss?
  3. How comparable are the Split and Conquer numbers taken from its paper to your retrained setting?
6 comments
AuthorsOpenReview ↗

R1: Multi-seed Evaluation, Architecture Ablations, and SSL Backbone Fairness (W1, W2, Q1, Q2)

Thank you for raising these concerns. To assess result robustness and disentangle backbone adaptation from the proposed components, we report three-seed results (mean ± SD) for key baselines and ablations on PS, HAD, and PE-E1. For LAV-DF, repeating multiple model configurations across several random seeds is considerably more computationally demanding due to the dataset's large scale. We initially attempted full-dataset multi-seed retraining but found that the required training time would extend beyond the rebuttal deadline. Rather than omitting this evaluation entirely, we therefore initiated multi-method, multi-seed experiments on a smaller subset of LAV-DF. These experiments are currently ongoing, and we will make every effort to report the corresponding results before the discussion period closes.

(1) Baseline comparisons and SSL backbone fairness. In our original experiments, SAL and BAM employed fully fine-tuned WavLM, whereas CFPRF used fully fine-tuned XLS-R. For fairer comparisons, we additionally reproduced CFPRF with WavLM and SAL with the same WavLM adapter-based adaptation strategy as STELA, keeping their remaining settings unchanged. As shown in Table 1, STELA consistently achieves the lowest mean EER and highest mean Acc and F1 across all three datasets, although the improvement margin on PS is relatively small. These results further support the effectiveness of our proposed framework under more closely matched backbone and adaptation settings.

(2) Architecture controls and component ablations. To isolate the contributions of the proposed components, we introduce two additional controls: FT BASE, which fully fine-tunes WavLM without adapters, and BASE, which retains STELA's architecture and adapter-based adaptation strategy but excludes the proposed training components. For the previously reported BASE w/ PSM, BASE w/ DA4, and STELA, we conducted two additional training runs to obtain three-seed results. As shown in Table 2, the ablations reveal dataset-dependent effects. On HAD and PE-E1, DA4 improves EER over BASE, while PSM alone does not consistently improve this metric. On PS, although the individual variants yield higher EER than BASE, their combination in Full STELA achieves the lowest EER, suggesting complementary effects. Overall, Full STELA improves frame-level detection over the architecture-matched BASE across all three datasets, providing evidence of gains beyond backbone adaptation alone.

Table 1. Multi-seed comparison with existing methods. Results are reported as mean ± SD (%) over three training seeds.

DatasetMethodEER ↓Acc ↑F1 ↑
PSSAL3.7841 ± 0.262596.4623 ± 0.282297.1638 ± 0.2196
SAL w/ Adapter4.1986 ± 0.218296.0350 ± 0.193596.8313 ± 0.1484
CFPRF-WavLM8.3471 ± 0.217492.8938 ± 0.197794.4698 ± 0.1337
STELA3.6489 ± 0.282596.7546 ± 0.224497.3840 ± 0.1777
HADSAL0.3843 ± 0.080299.6819 ± 0.061499.8052 ± 0.0375
SAL w/ Adapter0.2772 ± 0.030099.7485 ± 0.070399.8463 ± 0.0429
CFPRF-WavLM0.1439 ± 0.004399.9253 ± 0.004299.9544 ± 0.0026
STELA0.0017 ± 0.000299.9985 ± 0.000299.9991 ± 0.0001
PE-E1SAL0.3553 ± 0.011199.7803 ± 0.002699.8606 ± 0.0016
SAL w/ Adapter0.3530 ± 0.007499.7659 ± 0.006699.8514 ± 0.0042
CFPRF-WavLM0.5306 ± 0.015099.6337 ± 0.004699.7671 ± 0.0029
STELA0.1643 ± 0.010699.8799 ± 0.011699.9237 ± 0.0073

Table 2. Architecture and component ablations. Results are reported as mean ± SD (%) over three training seeds.

DatasetMethodEER ↓Acc ↑F1 ↑
PSFT BASE4.3261 ± 0.142796.3080 ± 0.183897.0368 ± 0.1413
BASE4.4659 ± 0.447796.1242 ± 0.305096.8929 ± 0.2394
BASE w/ PSM4.4326 ± 0.309096.1560 ± 0.153596.9331 ± 0.1191
BASE w/ DA45.3102 ± 0.910095.5826 ± 0.724196.5028 ± 0.5549
STELA3.6489 ± 0.282596.7546 ± 0.224497.3840 ± 0.1777
HADFT BASE0.0108 ± 0.002699.9842 ± 0.004299.9904 ± 0.0026
BASE0.0025 ± 0.000499.9979 ± 0.000599.9987 ± 0.0003
BASE w/ PSM0.0074 ± 0.003699.9965 ± 0.001499.9979 ± 0.0008
BASE w/ DA40.0024 ± 0.000899.9979 ± 0.001199.9987 ± 0.0007
STELA0.0017 ± 0.000299.9985 ± 0.000299.9991 ± 0.0001
PE-E1FT BASE0.2654 ± 0.005399.7865 ± 0.010099.8644 ± 0.0064
BASE0.2053 ± 0.021199.8301 ± 0.022099.8920 ± 0.0140
BASE w/ PSM0.2166 ± 0.006999.8391 ± 0.008599.8978 ± 0.0054
BASE w/ DA40.1637 ± 0.002999.8711 ± 0.003599.9181 ± 0.0023
STELA0.1643 ± 0.010699.8799 ± 0.011699.9237 ± 0.0073
AuthorsOpenReview ↗

R2: Explanation for Limited Frame-level Improvements and Motivation for PSM/DA4(W1, W3, W4)

We acknowledge that STELA's frame-level detection improvements over the strongest baseline are relatively modest. However, the primary motivation behind PSM and DA4 is not merely to minimize EER, but to address the discrepancy between frame-level detection and interval localization.

This motivation arose from our baseline reproductions. For instance, SAL achieves over 96% frame-level accuracy on PS, yet only approximately 58% mAP when contiguous spoof-positive frames are directly converted into proposals with a fixed confidence of 1. Moreover, CFPRF introduces PRN for proposal confidence estimation and boundary regression, but reports only 53.03% mAP on PS. These observations motivated us to focus on the quality of frame-level predictions for accurate interval localization, rather than solely optimizing detection metrics.

We agree that our original experiments did not sufficiently highlight this motivation. We therefore provide additional localization results for BASE, BASE w/ PSM, BASE w/ DA4, and Full STELA, using the same proposal construction without SML. As shown in Table 3, both component variants improve mean localization mAP over BASE on PS and PE-E1, even when their frame-level EER does not consistently improve. Full STELA further increases mAP from 58.12% to 63.99% on PS and from 97.48% to 98.34% on PE-E1, with improvements particularly evident at strict IoU thresholds.

These results demonstrate that the proposed components improve initial localization quality independently of SML, while Full STELA also achieves better frame-level detection than the architecture-matched BASE. Regarding Weakness4, we emphasize that the contribution of PSM and DA4 lies in their localization-oriented design, which targets the temporal consistency and interval quality of frame-level predictions rather than merely optimizing classification accuracy. We will better clarify this motivation and the specific roles of these components in the revised manuscript.

Table 3. Localization performance of frame-level predictions without SML. Results are reported as mean ± SD (%) over three seeds unless otherwise specified. Contiguous spoof-positive frames are converted into proposals with confidence 1. mAP is averaged over IoU thresholds from 0.50 to 0.95.

DatasetMethod[email protected] ↑[email protected] ↑[email protected] ↑[email protected] ↑mAP ↑
PSBASE63.9989 ± 0.236057.7960 ± 0.287953.9483 ± 0.376050.9626 ± 0.448858.1185 ± 0.2679
BASE w/ PSM64.7762 ± 1.930058.6866 ± 2.168654.8601 ± 2.276051.8473 ± 2.404658.9589 ± 2.1503
BASE w/ DA467.0736 ± 1.400761.6857 ± 1.363758.3594 ± 1.274455.8306 ± 1.208461.9460 ± 1.3444
STELA69.3483 ± 0.316263.7059 ± 0.337660.1932 ± 0.315257.4500 ± 0.244963.9915 ± 0.3208
HADBASE99.9970 ± 0.005399.9912 ± 0.008699.9740 ± 0.015099.8364 ± 0.027599.9730 ± 0.0093
BASE w/ PSM99.9853 ± 0.013199.9686 ± 0.030599.9370 ± 0.045499.7586 ± 0.072799.9462 ± 0.0304
BASE w/ DA499.9943 ± 0.005599.9902 ± 0.012099.9727 ± 0.019799.8649 ± 0.038799.9758 ± 0.0131
STELA99.9985 ± 0.002699.9901 ± 0.004799.9804 ± 0.012199.8839 ± 0.010999.9823 ± 0.0022
PE-E1BASE98.4047 ± 0.212897.8966 ± 0.229696.5630 ± 0.242294.0767 ± 0.238297.4758 ± 0.2279
BASE w/ PSM98.8420 ± 0.286598.3075 ± 0.303896.8790 ± 0.446294.1845 ± 0.572997.8636 ± 0.3291
BASE w/ DA498.6278 ± 0.243098.1392 ± 0.236797.0584 ± 0.211295.0994 ± 0.131397.8382 ± 0.2271
STELA99.0217 ± 0.212598.6526 ± 0.275497.6694 ± 0.255995.7909 ± 0.337798.3402 ± 0.2691
AuthorsOpenReview ↗

R3: Explaination for Prediction Fragmentation and the Role of SML (W3, W4)

Thank you for this insightful observation. We agree that our original presentation overemphasized prediction fragmentation without clearly distinguishing it from the broader localization problem. As discussed in R2, fragmentation is one manifestation of the mismatch between frame-level detection and interval localization, rather than its sole cause.

(1) Quantitative analysis of prediction discontinuities. Regarding W4, we acknowledge that our original visualizations provided only qualitative examples. To provide broader evidence, we analyze GT intervals overlapping with at least two frame-level proposals across the full test sets. As shown in Table 4, STELA consistently achieves the lowest multi-proposal GT ratio across all three datasets. On PS, this ratio decreases from 5.893% (SAL) and 9.590% (BAM) to 3.013% (STELA). These results provide dataset-level evidence of reduced prediction discontinuities beyond isolated examples, supporting our motivation to improve temporal consistency for localization. We also refer you to our response to Reviewer jWov (W1, Q1), where we provide a more detailed analysis of prediction fragmentation and temporal consistency.

Table 4. Multi-proposal GT analysis.

MethodPSHADPE-E1
SAL13,627 (5.893%)194 (2.138%)117 (0.433%)
BAM22,177 (9.590%)27 (0.298%)138 (0.510%)
STELA6,967 (3.013%)0 (0%)63 (0.233%)

Each entry reports the number and percentage of GT intervals overlapping at least two frame-level proposals.

(2) Contribution of merging and learned proposal scoring. Importantly, SML does not solely aim to recover fragmented intervals. Merging generates potentially more complete proposals, while learned scoring prioritizes accurately localized candidates over incomplete or inaccurate ones, followed by NMS to suppress redundant predictions.

To further examine the contribution of proposal merging, we conduct additional experiments on PS using both BASE and STELA as frontends (Table 5). We define a useful merged proposal as a merged candidate that achieves an IoU ≥ 0.9 with a ground-truth interval and exceeds the best IoU attainable by any non-merged candidate for that same interval. This metric measures whether merging introduces high-quality interval hypotheses that cannot be obtained from the original and boundary-adjusted proposals. We also compare localization mAP with and without merging to quantify its impact on final predictions.

Table 5. Contribution of proposal merging on PS.

FrontendUseful Merged ProposalsmAP w/o MergemAP w/ Merge
BASE8,376 (2.799%)80.28381.053
STELA5,683 (1.999%)83.17083.473

For BASE, merging introduces 8,376 useful proposals and improves mAP by 0.770 points. For STELA, it introduces 5,683 useful proposals and improves mAP by 0.303 points. The smaller gain with STELA is consistent with its stronger initial localization quality, which leaves fewer fragmented intervals requiring recovery. Nevertheless, these results confirm that merging provides additional interval hypotheses, while the larger improvements of SML primarily come from learned proposal scoring and selection.

Overall, we will revise the description of SML as merely recovering coherent regions from fragmented predictions. More precisely, SML improves localization by generating alternative interval hypotheses and learning to prioritize accurately localized proposals, with merging providing complementary candidates when needed. We will also clarify that local variations in learned representations are a possible explanation for prediction interruptions, rather than an established causal mechanism.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.