betaFeedback

WorksEngineering

CLRM: Cross-Layer Relation Modeling for Audio Deepfake Detection

Abstract

Audio deepfake detection (ADD) aims to distinguish bonafide speech from synthetic or manipulated speech. Representations from different layers of self-supervised learning (SSL) speech encoders vary in their ability to distinguish bonafide from spoofed speech. Comparing representations of the same utterance across encoder depths may reveal useful spoofing cues, yet common layer-selection and aggregation methods do not provide these comparisons as detection inputs. We propose Cross-Layer Relation Modeling (\ours{}), which explicitly encodes cross-layer relations alongside layer-specific features, making both directly available for ADD. CLRM reuses the existing detector's SSL encoder and identifies informative layer pairs by combining speaker-relative representation geometry with source-domain detection performance. The resulting relation score is fused with the detector's original score, requiring only the training of compact relation heads and fusion parameters while keeping the existing detector frozen. Across five external datasets, CLRM reduces the mean equal error rate (EER) for all three evaluated detectors, achieving relative reductions of 7.8–16.7%. Controlled comparisons with identical layer summaries and approximately matched auxiliary-head parameter counts also show a lower mean EER for explicit relations than direct concatenation, both before and after score fusion. These findings support explicit cross-layer comparisons as useful detection inputs for SSL-based audio deepfake detectors.

Introduction

Audio deepfake detection (ADD) aims to distinguish bonafide speech from synthetic or manipulated speech. Detecting spoofed speech across different generation methods and recording conditions remains an important challenge [12, 6]. Pretrained self-supervised learning (SSL) speech encoders, such as the multilingual XLS-R model [1], are widely used for ADD. These encoders produce representations at multiple processing depths, and previous analyses have shown that representations from different layers vary in their ability to distinguish bonafide from spoofed speech [4]. To exploit this variation, existing detectors select informative layers, aggregate features from multiple depths, or combine layer-specific predictions [23, 13, 22]. For example, XLSR-SLS employs a Sensitive Layer Selection (SLS) module to aggregate XLS-R representations using learned layer-dependent weights [23].

These approaches demonstrate the value of multilayer features for ADD, motivating us to investigate whether explicit comparisons between layers provide useful detection cues beyond those learned from feature concatenation. Representations at different encoder depths offer distinct views of the same utterance, making their differences and interactions potentially informative for detection. Direct concatenation preserves these representations but leaves the classifier to learn how to compare them. We therefore construct explicit cross-layer comparisons alongside layer-specific features, rather than relying on the classifier to learn these comparisons from direct concatenation.

In this paper, we propose Cross-Layer Relation Modeling (CLRM), a detector-augmentation framework that explicitly models relations between selected SSL layer representations. For each utterance, CLRM reuses the existing detector’s SSL encoder and constructs a joint representation from two selected layers, which retains their individual features while encoding their pairwise discrepancies, interactions, and similarity. Compact relation heads learn an auxiliary detection score from this representation, which is combined with the original detector’s score through score-level fusion. The existing detector, including its SSL encoder, remains frozen, and training is confined to the relation heads and fusion parameters. This design enables existing SSL-based detectors to incorporate relation-based detection evidence without modifying their architectures or retraining their original parameters.

The choice of layer pair determines which cross-layer relationships are modeled by CLRM. We use bonafide speech to summarize speaker-relative organization at different encoder depths, prioritizing candidate pairs that are separated in depth but have similar aggregate speaker statistics. Detection error on held-out source attacks determines the final layer pair and fusion configuration. The selected configuration is then applied unchanged to external datasets.

We evaluate CLRM with three XLS-R-based detector architectures on five external datasets. The five-dataset mean equal error rate (EER) decreases for all three detectors, with a maximum relative reduction of 16.7%. With identical layer summaries and approximately matched auxiliary-head parameter counts, explicit relation features also achieve lower mean EER than direct concatenation before and after fusion.

Our contributions are threefold:

  • Explicit cross-layer relation representation. We introduce a detection input that retains selected SSL layer summaries and directly encodes their discrepancies and interactions.

  • Source-guided detector augmentation. We combine speaker-relative layer-pair candidate construction, source-based configuration selection, and relation-score fusion to augment frozen XLS-R-based detectors with compact trainable heads.

  • Controlled empirical evaluation. We demonstrate improvements across three detector architectures and five external datasets. With identical layer summaries and approximately matched auxiliary-head parameter counts, explicit relations achieve a lower five-dataset mean EER than direct concatenation.

Related Work

Multilayer SSL Representations for ADD

Pretrained SSL speech encoders provide representations at multiple depths that are widely used for ADD. El Kheir et al. [4] analyze the detection

capabilities of different SSL layers, revealing substantial variation across encoder depths. Several methods exploit these representations through layer selection and aggregation. XLSR-SLS (Zhang et al. [23]) learns layer-dependent weights to aggregate XLS-R features, while Pan et al. [13] introduce attentive merging of intermediate WavLM representations. Xiao and Vu [22] instead combine predictions from individual XLS-R layers. These studies demonstrate the effectiveness of leveraging multilayer information through feature aggregation and decision fusion.

Cross-Layer Dependencies and Representation Comparisons

Beyond feature aggregation, recent studies have explored relationships between representations. HierCon (Liang et al. [9]) employs hierarchical attention and contrastive learning to model dependencies across temporal frames and encoder layers. Explicit feature comparisons have also been investigated in representation learning. Sentence-BERT (Reimers and Gurevych, [16]) explores pairwise embedding combinations involving elementwise differences and products for sentence-pair classification. These studies motivate examining explicit comparisons between selected SSL layer representations as detection inputs.

Cross-Dataset ADD

Generalization across synthesis methods and recording conditions remains a major challenge in ADD (Müller et al., [12]). Li et al. [8] investigate cross-domain detection using speech generated by zero-shot text-to-speech systems. SpeechFake (Huang et al., [5] introduces a large-scale multilingual dataset covering diverse speech generation methods. Huang et al. [6] investigate source and generator diversity to improve detection generalization. ICLAD (Chou et al., [2] explores comparison-guided in-context learning with audio language models for detecting unseen deepfakes. These studies address cross-dataset detection through data diversity, evaluation benchmarks, and alternative detection strategies.

Cross-Layer Relation Modeling

Framework Overview

For an input utterance xx, CLRM augments a frozen SSL-based detector with a learned score derived from cross-layer relations. The existing detector produces sbase(x)s^{\mathrm{base}}(x), and its SSL encoder supplies the hidden representations used to construct a relation input from selected layers (p,q)(p, q). Relation heads map this input to sp,qrel(x)s_{p,q}^{\mathrm{rel}}(x), which is fused with the original score to obtain sfinal(x)s^{\mathrm{final}}(x). Source-side geometry guides candidate layer-pair construction, followed by detection-based configuration selection. The detector and encoder remain frozen; only relation heads and fusion parameters are fitted.

Figure 1. CLRM for ADD. (a) Last-layer detection and (b) multilayer aggregation illustrate conventional uses of SSL representations. (c) CLRM uses selected representations from the frozen XLS-R encoder of the existing detector and combines the resulting relation score with its original score through source-fitted fusion. Snowflakes indicate frozen components.

Explicit Cross-Layer Relation Representation

Our representation is designed to retain information from individual encoder layers while explicitly describing how their features differ and interact. We first summarize the hidden states at each layer into a fixed-dimensional utterance representation and then construct pairwise comparisons between the selected summaries. Let H(l)(x)∈RT(l)(x)×d\mathbf{H}^{(l)}(x) \in\mathbb{R}^{T^{(l)}(x)\times d} denote the output of Transformer block ll, where the 24 blocks are indexed as L1–L24 and T(l)(x)T^{(l)}(x) and dd denote the frame count and hidden dimension. For each layer, we concatenate the temporal mean μl(x)\boldsymbol{\mu}_l(x) and population standard deviation σl(x)\boldsymbol{\sigma}_l(x):

zl(x)=[μl(x);σl(x)]∈R2d.(1)\mathbf{z}_l(x) = [\boldsymbol{\mu}_l(x); \boldsymbol{\sigma}_l(x)] \in\mathbb{R}^{2d}. \tag*{(1)}

Here, [⋅;⋅][\cdot;\cdot] denotes concatenation. Each coordinate of zl(x)\mathbf{z}_l(x) is standardized using statistics estimated for that layer from the official training split of ASVspoof 2019 Logical Access (LA), yielding z~l(x)\tilde{\mathbf{z}}_l(x). The same standardization is applied to development and evaluation utterances.

For the selected layers pp and qq, we construct the relation representation as

rp,q(x)=[z~p;z~q;∣z~p−z~q∣;z~p⊙z~q;cos⁡(z~p,z~q)].(2)\mathbf{r}_{p,q}(x) = [\tilde{\mathbf{z}}_p; \tilde{\mathbf{z}}_q; \lvert\tilde{\mathbf{z}}_p-\tilde{\mathbf{z}}_q\rvert; \tilde{\mathbf{z}}_p \odot\tilde{\mathbf{z}}_q; \cos(\tilde{\mathbf{z}}_p,\tilde{\mathbf{z}}_q)]. \tag*{(2)}

The argument xx is omitted from the layer summaries on the right-hand side for readability. The operators ∣⋅∣\lvert\cdot\rvert, ⊙\odot, and cos⁡(⋅,⋅)\cos(\cdot,\cdot) denote elementwise absolute value, elementwise multiplication, and cosine similarity, respectively. The absolute difference represents coordinate-wise discrepancies, the elementwise product encodes multiplicative interactions, and cosine similarity summarizes the angular relationship between the two summaries. Together with the retained layer summaries, these comparisons form an (8d+1)(8d+1)-dimensional input that makes both layer-specific and relational information available to the classifier. Since the comparisons operate on utterance-level summaries, no frame-wise alignment between the selected layers is required.

Relation Scoring

We learn a detection score from rp,q(x)\mathbf{r}_{p,q}(x) using three independently initialized feed-forward heads. Each head gϕmg_{\phi_m} consists of a linear projection, a Gaussian error linear unit (GELU) activation, dropout, and a linear output layer. The relation score is obtained by averaging the raw outputs of the heads:

sp,qrel(x)=1M∑m=1Mgϕm(rp,q(x)),M=3.(3)s^{\mathrm{rel}}_{p,q}(x) = \frac{1}{M}\sum_{m=1}^{M} g_{\phi_m}(\mathbf{r}_{p,q}(x)), \qquad M = 3. \tag*{(3)}

The heads are fitted independently on the official source training split, and their checkpoints are selected using the official source development split. Score averaging is performed before orientation, normalization, and fusion. For XLSR-SLS, we select the sign of the relation score on Meta-dev, a source partition used for configuration selection. The selection criterion is detailed in Appendix § 9.§ 2.

Score-Level Fusion

We combine the relation score with the score of the existing detector to integrate cross-layer information into its final decision. For a fixed selected pair (p,q)(p, q), we write srel(x)=sp,qrel(x)s^{\mathrm{rel}}(x) = s^{\mathrm{rel}}_{p,q}(x). The original and relation scores are standardized using their respective statistics from the official source development split, yielding s~base(x)\tilde{s}^{\mathrm{base}}(x) and s~rel(x)\tilde{s}^{\mathrm{rel}}(x). We then apply logistic-regression score fusion:

sfinal(x)=as~base(x)+bs~rel(x)+c,(4)s^{\mathrm{final}}(x) = a\tilde{s}^{\mathrm{base}}(x) + b\tilde{s}^{\mathrm{rel}}(x) + c, \tag*{(4)}

where aa, bb, and cc are fusion coefficients fitted on the official source development split. The fused score is the linear decision score of the logistic-regression model, with its regularization strength selected on Meta-dev. This score-level interface retains the existing detector’s feature processing and decision pathway while incorporating the relation-based score. The SSL encoder and original detector remain frozen throughout training; only the relation heads and fusion parameters are optimized.

Geometry-Guided Layer-Pair Selection

The choice of layer pair determines which cross-layer relationships are presented to the relation classifier. We construct a compact candidate set by favoring pairs that are separated in encoder depth but exhibit similar speaker-relative organization on bonafide speech. Depth separation encourages comparisons between different processing stages, while speaker-relative statistics provide a common reference for comparing layers. Detection performance is then used to select the final pair from these candidates.

For each layer ll, we characterize speaker-relative organization using bonafide source utterances with speaker labels. For candidate selection, we obtain

geometry embeddings through temporal mean pooling of H(l)(x)\mathbf{H}^{(l)}(x) followed by ℓ2\ell_2 normalization. These embeddings differ from the standardized mean-and-standard-deviation summaries used to construct the relation representation in § 3.2. Each embedding is compared with its own-speaker centroid and competing-speaker centroids, using a leave-one-out centroid for the correct speaker. We aggregate these comparisons into four layer-level statistics:

gl=[Rl;Al;Mlmax;Mlmean],(5)\mathbf{g}_l = [R_l; A_l; M_l^{\mathrm{max}}; M_l^{\mathrm{mean}}], \tag*{(5)}

where RlR_l is the mean rank of the correct speaker, AlA_l denotes top-1 own-speaker identification accuracy, and MlmaxM_l^{\mathrm{max}} and MlmeanM_l^{\mathrm{mean}} are the mean similarity margins relative to the strongest and average competing speakers, respectively. Their precise definitions are provided in Appendix § 8.§ 1.

Figure 2 illustrates how speaker-relative organization varies with encoder depth. In the shallower layers, own-speaker rank is relatively low and max-other margin is relatively high; toward the middle layers, rank increases while the margin decreases, followed by further changes at deeper layers. These non-monotonic patterns motivate comparing layers using their speaker-relative summaries rather than relying on depth alone. We use all four statistics in gl\mathbf{g}_l to construct the candidate ranking.

Figure 2. Speaker-relative statistics across XLS-R layers. Own-speaker rank (lower is better) and max-other margin (higher is better) are two of the four statistics used for candidate construction. Both are computed from bonafide ASVspoof 2019 LA evaluation utterances under the primary source protocol. The figure focuses on L1–L21, as the substantially different scales of L22–L24 would make variations within the candidate range less visible.

After standardizing each component of gl\mathbf{g}_l across candidate layers to obtain g^l\hat{\mathbf{g}}_l, we define the geometry distance and ranking score for layers pp and qq as

DG(p,q)=∥g^p−g^q∥2,D_G(p,q) = \lVert\hat{\mathbf{g}}_p-\hat{\mathbf{g}}_q\rVert_2,
RG(p,q)=∣q−p∣DG(p,q)+ϵ,(6)R_G(p,q) = \frac{\lvert q-p\rvert}{D_G(p,q)+\epsilon}, \tag*{(6)}

where ϵ=10−8\epsilon= 10^{-8}. A higher RG(p,q)R_G(p,q) favors greater separation in encoder depth together with more similar speaker-relative summaries. We consider layers L1–L21 with a minimum separation of eight layers, yielding 91 eligible pairs, and retain the six highest-ranked candidates. The geometry statistics are used for candidate construction and do not enter the utterance-level relation representation in Eq. (2).

For each candidate, the relation heads are trained on the official source training split and fusion coefficients are fitted on the official development split. We select the layer pair and inverse fusion-regularization strength CC on Meta-dev by minimizing fused Macro equal error rate (EER), defined as the unweighted mean of attack-wise EERs. The selected configuration is evaluated on Meta-test, which contains different spoof attacks, and then applied unchanged to the external datasets. Section 4.1 specifies the source partitions and geometry-reference populations. Speaker labels and geometry statistics are not required at inference.

Experimental Setup

Data and Evaluation Protocols

We use ASVspoof 2019 Logical Access (LA) for source-side development [21]. Its official training split is used to fit the relation heads and estimate layer-wise feature-standardization statistics, while the official development split is used for checkpoint selection, score normalization, and fusion-parameter fitting.

We partition the official evaluation split by spoof attack into Meta-dev for configuration selection and Meta-test for assessing the selected configuration. The primary protocol shares bona fide utterances between these partitions and uses them as the geometry reference. We additionally evaluate utterance-disjoint and speaker-disjoint protocols, reconstructing geometry from bona fide training data and repeating configuration selection. Table 1 summarizes these protocols, with further details in Appendix § 9.§ 1.

ProtocolGeometry referenceMeta-dev NNMeta-test NNOverlap U/S
Attack-disjointEval bonafide: 7,355 utt., 67 spk.41,75336,8397,355/67
Utterance-disjointTrain bonafide: 2,580 utt., 20 spk.38,08533,1520/67
Speaker-disjointTrain bonafide: 2,580 utt., 20 spk.20,71818,6050/0

Table 1. Table 1: Source evaluation protocols. NN is the number of trials, and U/S denotes bonafide utterance/speaker overlap between Meta-dev and Meta-test. Meta-dev is used for configuration selection and Meta-test for assessment.

For cross-dataset evaluation, we use ASVspoof 2021 LA and Deepfake (DF) [11], ASVspoof 5 Eval [20], the normalized Fake-or-Real test set [15], and In-the-Wild [12]. The layer pair, relation heads, score orientation, normalization statistics, and fusion parameters are determined from the source data and held fixed during external evaluation. The ASVspoof 2021 results follow this post-challenge source-development protocol.1 Dataset-specific evaluation procedures are described in Appendix § 11.

Evaluated Detectors

We evaluate CLRM with three detector architectures using XLS-R representations: XLSR-SLS [23], XLSR-Conformer-TCM based on Temporal-Channel Modeling (TCM) [19], and XLSR-AASIST, an XLS-R-based variant of Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks (AASIST) [7, 18]. XLSR-SLS is the primary detector for studying the relation representation and selecting the layer pair. The selected layer indices are reused for XLSR-Conformer-TCM and XLSR-AASIST, with relation heads and fusion parameters fitted separately for each detector using source data. For each augmented system, CLRM uses representations from the detector’s existing frozen XLS-R encoder. Each base detector and its augmentation are evaluated on the same external trials using the same dataset-specific evaluator. Additional SSL-based reference results are included in Table 2, with their result sources summarized in Appendix § 11.§ 2.

System21 LA21 DFASVspoof 5In-the-Wild Fake-or-RealAverage
Reference systems
XLSR-Conformer (Rosello et al., [17])1.282.8617.2110.888.45
Nes2Net-X (Liu et al., [10])2.001.7822.066.606.317.75
XLSR-DADGMoE-AASIST (Qin et al., [14])1.532.5418.016.354.426.57
BiCrossMamba-ST (El Kheir et al., [3])3.832.3513.677.946.856.93
Base detectors and CLRM augmentation
XLSR-SLS (Zhang et al., [23])2.951.7718.767.285.037.16
+ CLRM2.561.9116.327.454.086.46
XLSR-Conformer-TCM (Truong et al., [19])2.302.0616.167.276.476.85
+ CLRM2.232.0614.557.135.616.32
XLSR-AASIST (Jung et al., [7])1.273.4615.278.087.777.17
+ CLRM1.172.9214.416.704.665.97

Table 2. Detection performance on five external datasets (EER, %; lower is better). The lower block compares each base detector with its CLRM augmentation on the same evaluation trials and using the same evaluators. 21 LA and 21 DF denote ASVspoof 2021 LA and DF, respectively. Average is the unweighted mean of the five dataset-level EERs (External-5); paired comparisons use unrounded values. Reference-system result records are described in Appendix D.2. Bold indicates the lowest value in each column.

Controlled Representation Comparisons

To isolate the effect of explicit cross-layer relations, we compare the proposed Relation input with direct concatenation of the same selected layer summaries. Relation denotes the representation in Eq. (2), whereas CLRM denotes the complete augmentation with relation scoring and fusion. All representation controls use the selected L8–L21 layers and identical layer-wise standardization statistics. The concatenation input is

rp,qconcat(x)=[z~p(x);z~q(x)]∈R4d.(7)\mathbf{r}^{\mathrm{concat}}_{p,q}(x) = [\tilde{\mathbf{z}}_p(x);\tilde{\mathbf{z}}_q(x)] \in\mathbb{R}^{4d}. \tag*{(7)}

We evaluate Concat-256, which matches the hidden width of the relation head, and Concat-512, which approximately matches its trainable parameter count. Each input configuration uses three independently fitted heads with the same source-side training and selection procedure.

We assess the auxiliary scores both independently and after fusion with XLSR-SLS. These comparisons measure the discriminative value of the input representation and its contribution to the existing detector, respectively. Additional source-side controls examine single-layer inputs, individual relation terms, and alternative score-combination rules; their configurations and parameter counts are provided in Appendix § 9.§ 3.

Implementation and Evaluation Metrics

Implementation. We use the frozen XLS-R-300M encoder already employed by each base detector, with hidden dimension d=1024d = 1024 [1]. The relation representation has 8d+1=8,1938d + 1 = 8{,}193 dimensions. Each relation head uses a hidden width of 256, GELU activation, and dropout of 0.20, yielding 2,097,921 trainable parameters per head and approximately 6.29 million across three heads. The pretrained encoder and original detector remain frozen, and only the relation heads and fusion parameters are fitted. Three independently initialized heads are trained for each configuration, and their raw outputs are averaged before score orientation, normalization, and fusion. For the primary representation experiments, we use AdamW with learning rate 2×10−42 \times10^{-4}, weight decay 10−410^{-4}, batch size 2,048, and a maximum of 15 epochs. Training uses weighted binary cross-entropy and early stopping based on official-development EER with patience of four. Audio is converted to mono at 16 kHz and truncated or repeated to 64,600 samples. Appendix § 9.§ 4 provides the preprocessing, optimization, and seed settings.

Evaluation metrics. We report equal error rate (EER, %), with lower values indicating better detection. For source evaluation, Macro EER averages attack-wise EERs, Worst EER is the highest attack-wise EER, and Pooled EER is computed over all trials in a partition. For external evaluation, External-5 is the unweighted mean of the five dataset-level EERs. EER is computed from the corresponding base, auxiliary-only, or fused score stream. For the relation ensemble, raw head scores are averaged before evaluation. ASVspoof 2021 LA and DF use the recorded official-evaluator results, while the remaining evaluation procedures are detailed in Appendix § 11. For the source-side layer-pair diagnostic, we additionally report Δcos=cˉbonafide−cˉspoof\Delta_{\mathrm{cos}} = \bar{c}_{\mathrm{bona fide}} - \bar{c}_{\mathrm{spoof}}, the difference in class-mean cosine similarity between the standardized layer summaries on the Meta-test.

Results and Analysis

Main Results across Detectors

Table 2 shows that CLRM lowers External-5 mean EER for all three detectors, with relative reductions of 9.7%, 7.8%, and 16.7% for XLSR-SLS, XLSR-Conformer-TCM, and XLSR-AASIST, respectively. The largest reduction is obtained with XLSR-AASIST (7.170% to 5.972%), which improves on all five datasets. Together, the paired comparisons demonstrate the benefit of relation scoring across three XLS-R-based detector architectures.

Effectiveness of Explicit Cross-Layer Relations

Controlled representation comparison. With identical L8–L21 summaries, Relation achieves lower source Meta-test and External-5 mean EER than both concatenation controls (Table 3). Against the near-parameter-matched Concat-512, source Macro EER decreases from 0.117% to 0.069%, while External-5 EER decreases from 16.728% to 11.333% before fusion and from 7.330% to 6.462% after fusion. These controlled results show that explicit cross-layer comparisons provide more effective detection inputs than direct concatenation for the selected layers and classifier configurations.

Auxiliary inputParams/head (10610^{6})Source fusedAuxiliary onlyFused with SLS
None (SLS)–0.149–7.158
Concat-2561.0490.11416.9497.422
Concat-5122.0980.11716.7287.330
Relation2.0980.06911.3336.462

Table 3. Explicit relations versus direct concatenation (EER, %). All auxiliary inputs use identical L8–L21 layer summaries and standardization statistics. Concat-256 matches the relation-head hidden width, and Concat-512 approximately matches its per-head parameter count. Source fused is Meta-test Macro EER; the last two columns report External-5 EER before and after fusion with XLSR-SLS.

Cross-dataset comparison. Relation yields the lowest EER among the tested representations on four of the five datasets, both before and after fusion (Table 4). On ASVspoof 5, for example, fused EER decreases from 19.374% for Concat-512 to 16.317% for Relation, with a corresponding improvement in auxiliary-only scoring. The consistent gains across these two scoring conditions strengthen the evidence for relation-based inputs.

DatasetFused with SLS: Concat-256Fused with SLS: Concat-512Fused with SLS: RelationAuxiliary only: Concat-256Auxiliary only: Concat-512Auxiliary only: Relation
2021 LA2.3012.2882.5584.7994.7126.081
2021 DF2.0312.0381.9098.1658.3195.416
ASVspoof 519.90519.37416.31727.86125.93317.598
In-the-Wild8.5808.5487.45032.28932.26621.806
Fake-or-Real4.2944.4024.07811.63212.4085.762
Mean7.4227.3306.46216.94916.72811.333

Table 4. Dataset-level comparison of relation and concatenation inputs (EER, %). Each configuration uses the same L8–L21 summaries; Mean is the unweighted average across the five external datasets. Bold marks the minimum within each dataset and scoring block.

Representation components. The full relation representation achieves the lowest source Meta-test Macro, Worst, and Pooled EER among the broad input controls (Figure 3(a)). Its Macro EER of 0.069% is lower than the near-parameter-matched L21-only control (0.119%) and all single-term-removal variants (0.082–0.114%; Appendix § 9.§ 3). These ablations support retaining both layer summaries and all three comparison terms.

Complementarity with the Base Detector

The XLSR-SLS results provide direct evidence of score complementarity: fusion achieves 6.462% External-5 EER, compared with 7.158% for the original detector and 11.333% for Relation alone (Table 3). For XLSR-Conformer-TCM, fusion likewise improves on the individual scores (6.316% versus 6.849% and 6.852%), while XLSR-AASIST reaches 5.972% after fusion (Appendix § 10). This indicates that relation-based evidence can improve the combined decision even when the auxiliary score is not stronger on its own.

On the source Meta-test, logistic fusion reduces

Macro EER to 0.0690%, compared with 0.0968% for Relation alone and 0.0747% for equal-weight fusion (Figure 3(b)). The external results use the fusion configuration selected on source data.

Geometry-Guided Layer-Pair Selection

Geometry-guided ranking yields six candidates, and detection-based selection chooses L8–L21, which achieves the lowest Meta-dev Macro EER (0.0888%; Table 5). With six candidates, the selected configuration achieves 0.069% Meta-test Macro EER, compared with 0.078% for depth-gap-only ranking and 0.103% for geometry-only ranking. Random shortlists average 0.076%, and exhaustive selection achieves 0.071% (Appendix A.3). These results support a compact geometry-guided shortlist followed by detection-based selection.

PairGapDGD_GRGR_GMacro EER (%)
L9–L1780.4517.660.1188
L9–L21120.6917.410.1062
L8–L21130.8515.260.0888
L7–L18110.7215.220.0892
L10–L21110.7714.270.0965
L7–L19120.8514.200.1062

Table 5. Geometry-guided layer-pair candidates. Gap denotes encoder-depth separation, DGD_G speaker-relative geometry distance, and RGR_G the candidate-ranking score. Macro EER is measured on source Meta-dev and determines the selected pair.

Robustness to Source Partitioning and Initialization

CLRM improves source Meta-test Macro EER under both utterance-disjoint (0.170% to 0.101%) and speaker-disjoint (0.107% to 0.045%) protocols (Table 6). Geometry and configuration selection are repeated for each partition, supporting the improvements under both stricter source protocols.

ProtocolPairCCSLS + CLRM
Utterance-disjointL7–L180.0010.170 0.101
Speaker-disjointL8–L210.0010.107 0.045

Table 6. Source Meta-test results under stricter partitions (Macro EER, %). Each pair and fusion regularization setting is selected independently on Meta-dev.

Across three runs with L8–L21 fixed, Macro EER averages 0.070% (sample SD: 0.007 percentage points), compared with 0.149% for XLSR-SLS. Repeated layer-pair selection chooses L8–L21 in two runs and L9–L17 in the third, where Macro EER reaches 0.092%. These findings support the consistency of the source-side improvements across the evaluated partitions and initializations.

Conclusion

We introduced CLRM, which augments frozen SSL-based detectors by jointly modeling selected layer summaries and explicit cross-layer comparisons. With identical layer inputs and approximately matched auxiliary-head parameters, relation-based inputs achieve lower five-dataset mean EER than direct concatenation, both independently and after fusion with XLSR-SLS. The augmentation improves mean EER for all three evaluated XLS-R-based detectors, with a maximum relative reduction of 16.7%. These results support explicit cross-layer comparisons as effective inputs for audio deepfake detector augmentation.

Limitations

The study uses XLS-R-300M representations, ASVspoof 2019 LA for source-side development, and five external evaluation datasets. Speaker labels are needed for geometry-based candidate construction, but not for inference. The primary Meta-dev/Meta-test protocol shares bonafide utterances, while additional utterance- and speaker-disjoint controls assess source-side performance with separately constructed partitions. Reported external EERs are point estimates; initialization and component ablations are evaluated on source data. The relation-based system adds three prediction heads to the original detector, and its latency and memory requirements remain to be measured.

Speaker-Relative Geometry and Candidate Search

We describe the speaker-relative statistics used for candidate construction and report alternative layer-pair searches and a post-hoc cosine analysis.

Speaker-Relative Statistics

Let Bk\mathcal{B}_k be the set of bona fide reference utterances from speaker kk, with KK speakers and B=⋃k=1KBk\mathcal{B} = \bigcup_{k=1}^{K}\mathcal{B}_k. For utterance xix_i, we construct a geometry embedding at layer ll by temporal mean pooling followed by ℓ2\ell_2 normalization:

hˉi(l)=1Ti(l)∑t=1Ti(l)Hi,t(l),\bar{\mathbf{h}}_i^{(l)} = \frac{1}{T_i^{(l)}} \sum_{t=1}^{T_i^{(l)}} \mathbf{H}_{i,t}^{(l)},
hi(l)=hˉi(l)∥hˉi(l)∥2.(8)\mathbf{h}_i^{(l)} = \frac{\bar{\mathbf{h}}_i^{(l)}}{\left\|\bar{\mathbf{h}}_i^{(l)}\right\|_2}. \tag*{(8)}

For speaker kk, the centroid and the leave-one-out centroid for an utterance i∈Bki \in\mathcal{B}_k are

ck(l)=1∣Bk∣∑j∈Bkhj(l),\mathbf{c}_k^{(l)} = \frac{1}{|\mathcal{B}_k|}\sum_{j\in\mathcal{B}_k}\mathbf{h}_j^{(l)},
ck,−i(l)=1∣Bk∣−1∑j∈Bkj≠ihj(l).(9)\mathbf{c}_{k,-i}^{(l)} = \frac{1}{|\mathcal{B}_k|-1}\sum_{\substack{j\in\mathcal{B}_k\\j\ne i}}\mathbf{h}_j^{(l)}. \tag*{(9)}

We compute the cosine similarity to the correct speaker and each competing speaker k′≠kk' \ne k as

si,k(l)=cos⁡(hi(l),ck,−i(l)),s_{i,k}^{(l)} = \cos\left(\mathbf{h}_i^{(l)},\mathbf{c}_{k,-i}^{(l)}\right),
si,k′(l)=cos⁡(hi(l),ck′(l)).(10)s_{i,k'}^{(l)} = \cos\left(\mathbf{h}_i^{(l)},\mathbf{c}_{k'}^{(l)}\right). \tag*{(10)}

Let ri(l)r_i^{(l)} be the rank of the correct speaker among the KK scores in descending order, and let ai(l)=I[ri(l)=1]a_i^{(l)} = \mathbb{I}[r_i^{(l)} = 1]. The margins relative to the strongest and average competing speakers are

mi,max⁡(l)=si,k(l)−max⁡k′≠ksi,k′(l),m_{i,\max}^{(l)} = s_{i,k}^{(l)} - \max_{k'\ne k}s_{i,k'}^{(l)},
mi,mean(l)=si,k(l)−1K−1∑k′≠ksi,k′(l).(11)m_{i,\mathrm{mean}}^{(l)} = s_{i,k}^{(l)} - \frac{1}{K-1}\sum_{k'\ne k}s_{i,k'}^{(l)}. \tag*{(11)}

Averaging over the reference utterances gives four layer-level statistics:

Rl=1∣B∣∑i∈Bri(l),R_l = \frac{1}{|\mathcal{B}|}\sum_{i\in\mathcal{B}} r_i^{(l)},
Al=1∣B∣∑i∈Bai(l),A_l = \frac{1}{|\mathcal{B}|}\sum_{i\in\mathcal{B}} a_i^{(l)},
Mlmax⁡=1∣B∣∑i∈Bmi,max⁡(l),M_l^{\max} = \frac{1}{|\mathcal{B}|}\sum_{i\in\mathcal{B}}m_{i,\max}^{(l)},
Mlmean=1∣B∣∑i∈Bmi,mean(l).(12)M_l^{\mathrm{mean}} = \frac{1}{|\mathcal{B}|}\sum_{i\in\mathcal{B}}m_{i,\mathrm{mean}}^{(l)}. \tag*{(12)}

The four statistics form gl=[Rl;Al;Mlmax⁡;Mlmean]\mathbf{g}_l = [R_l; A_l; M_l^{\max}; M_l^{\mathrm{mean}}]. These mean-pooled, ℓ2\ell_2-normalized geometry embeddings are distinct from the mean-and-standard-deviation summaries used to construct relation inputs in Eq. (2).

Geometry Normalization and Search Domain

Each coordinate of gl\mathbf{g}_l is standardized across candidate layers using its population mean and standard deviation. We consider layers L1–L21 with a minimum separation of eight layers, yielding 91 eligible pairs. The ranking score RG(p,q)R_G(p,q) in Eq. (6) combines depth separation with similarity of the standardized geometry vectors. The six highest-ranked pairs form the shortlist for detection-based selection on Meta-dev. The geometry statistics determine the candidate shortlist, while the relation classifier uses the representation in Eq. (2).

Alternative Candidate-Selection Strategies

We compare the geometry-guided six-pair shortlist with ten random six-pair shortlists and an exhaustive search over all 91 eligible pairs. All strategies select the layer pair and fusion regularization on Meta-dev by fused Macro EER and evaluate the selection on Meta-test. Random shortlists are sampled without replacement using NumPy’s default random generator (seed 20261001), with eligible pairs ordered by their lower and upper layer indices.

The random shortlists yield a mean Meta-test Macro EER of 0.076% (sample standard deviation: 0.008 percentage points). The geometry-guided configuration achieves 0.069%, while the configuration selected by exhaustive search achieves 0.071% (Table 7).

SearchSelected pairMeta-devMeta-test
Random 00L2–L160.0680.085
Random 01L3–L110.0640.071
Random 02L2–L110.0690.064
Random 03L3–L170.0690.064
Random 04L2–L170.0750.085
Random 05L10–L210.0970.075
Random 06L2–L210.0720.078
Random 07L2–L120.0650.081
Random 08L2–L160.0680.085
Random 09L2–L190.0910.076
Exhaustive (91)L3–L110.0640.071

Table 7. Alternative candidate searches (source Macro EER, %; lower is better). Each random shortlist contains six pairs; exhaustive search considers all 91 eligible pairs. The pair and fusion regularization are selected on Meta-dev and evaluated on Meta-test.

Attack-Wise Cross-Layer Cosine Analysis

For the selected L8–L21 pair, we measure Δcos⁡=cˉbona fide−cˉspoof\Delta_{\cos} = \bar{c}_{\mathrm{bona\ fide}} - \bar{c}_{\mathrm{spoof}}, where each term is the corresponding class-mean cosine similarity between standardized layer summaries on source Meta-test. Table 8 reports the attack-wise results.

AttackA08A10A12A14A16A18
Δcos⁡\Delta_{\cos}6.863.681.814.74−9.45-9.455.00

Table 8. Attack-wise Δcos⁡\Delta_{\cos} for L8–L21 on source Meta-test (10−210^{-2}).

This statistic is used only for post-hoc analysis and does not enter layer-pair selection or score fusion.

Source Protocol and Implementation Details

Construction of Disjoint Source Protocols

The primary source protocol assigns attacks A07, A09, A11, A13, A15, A17, and A19 to Meta-dev, and A08, A10, A12, A14, A16, and A18 to Meta-test. Both partitions contain the same 7,355 bona fide utterances from 67 speakers, which also provide the primary geometry reference. The utterance-disjoint and speaker-disjoint controls retain the attack assignments while separating bona fide utterances or speakers between Meta-dev and Meta-test. In both controls, the geometry reference consists of 2,580 bona fide training utterances from 20 speakers, and candidate construction and configuration selection are repeated independently.

Both disjoint protocols use seed 20261001. For the utterance-disjoint protocol, utterance IDs are sorted within each speaker by the first 16 hexadecimal digits of a Secure Hash Algorithm 256 (SHA-256) hash of the seed, a null separator, and the utterance ID. Let nkn_k denote the number of bona fide utterances available for speaker kk. The first ⌈nk/2⌉\lceil n_k/2\rceil utterances are assigned to Meta-dev and the remainder to Meta-test, giving 3,687 and 3,668 bona fide utterances, respectively. For the speaker-disjoint protocol, speakers are sorted by decreasing trial count with seed-dependent hash tie-breaking, then greedily assigned to the partition with fewer trials. This gives 30 Meta-dev speakers (3,330 bona fide trials) and 37 Meta-test speakers (4,025 bona fide trials). Table 1 provides the trial counts and overlap statistics.

Relation-Head Fitting and Score Fusion

For each candidate pair, three independently initialized relation heads are trained on the official source training split, with checkpoints selected on the official source development split. Their raw scores are averaged before score orientation, normalization, and fusion. Score orientation refers to the choice of score sign. For the primary XLSR-SLS experiments, this sign is selected on Meta-dev by minimizing 0.50 Pooled + 0.35 Macro + 0.15 Worst EER. Worst and Pooled EER are successive tie-breakers, and the positive orientation is retained if all criteria tie.

Each oriented score stream is standardized using the official source development mean μ\mu and population standard deviation σ\sigma:

s~(x)=s(x)−μmax⁡(σ,10−8).(13)\tilde{s}(x) = \frac{s(x)-\mu}{\max(\sigma,10^{-8})}. \tag*{(13)}

The normalized detector and relation scores are fused using logistic regression with an intercept, ℓ2\ell_2 regularization, and no class reweighting. We use the limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) solver with at most 5,000 iterations and tolerance 10−1010^{-10}. The inverse regularization strength is selected jointly with the layer pair on Meta-dev from C∈{10−4,10−3,10−2,0.1,1,10}C \in\{10^{-4},10^{-3},10^{-2},0.1,1,10\} using fused Macro EER. The selected orientation, normalization statistics, layer pair, and fusion configuration are fixed for Meta-test and external evaluation.

Auxiliary Inputs and Operator Controls

Table 9 summarizes the input dimensions, hidden widths, and trainable parameters of the auxiliary heads. Parameter counts refer to one auxiliary head. and exclude the frozen SSL encoder, original detection head, and fusion coefficients. Each configuration uses three independently trained heads. Concat-256 matches the hidden width of Relation, while Concat-512 approximately matches its per-head parameter count. Single-layer and relation-term removal controls are trained independently using the same source-side fitting procedure.

InputDim.WidthParams/head
Relation8,1932562,097,921
Concat-2564,0962561,049,089
Concat-5124,0965122,098,177
L21-2562,048256524,801
L21-10232,0481,0232,097,151
w/o abs. difference6,1452561,573,633
w/o product6,1452561,573,633
w/o cosine8,1922562,097,665

Table 9. Auxiliary input dimensions and trainable parameters per head. Each configuration uses three independently trained heads.

Auxiliary inputSelected CCMacro EER
L21-2560.10.121
L21-1023100.119
w/o absolute difference10−410^{-4}0.114
w/o product10.110
w/o cosine10−310^{-3}0.082

Table 10. Additional representation controls on source Meta-test (fused Macro EER, %; lower is better). CC is independently selected on Meta-dev.

Table 10 reports fused Macro EER on source Meta-test for the single-layer and relation-component controls. Each configuration uses independently fitted heads and Meta-dev-selected fusion regularization; the full Relation input is reported in Table 3.

The distance-statistics input in Figure 3 comprises eight scalar features: cosine similarity, mean squared difference, mean absolute difference, maximum absolute difference, the 90th percentile of absolute difference, log ratio of vector norms, and mean absolute difference for the temporal-mean and temporal-standard-deviation components separately. Cosine similarity and norm ratios use ϵ=10−8\epsilon= 10^{-8} for numerical stability. This control uses a head with hidden width 32 and is independently trained from the capacity-controlled concatenation configurations.

Optimization, Preprocessing, and Ensemble Accounting

Audio is converted to mono at 16 kHz and represented by 64,600 samples; longer utterances are truncated and shorter ones are repeated to this length. Temporal standard deviations are floored at 10−610^{-6}, and coordinate-wise feature standardization uses a denominator floor of 10−510^{-5}. The XLS-R-300M encoder used by the existing detector has hidden dimension d=1024d = 1024 and remains frozen.

Each relation head follows a Linear–GELU–Dropout–Linear architecture with hidden width 256 and dropout 0.20. Training uses AdamW with learning rate 2×10−42 \times10^{-4}, weight decay 10−410^{-4}, batch size 2,048, gradient clipping at 5.0, and at most 15 epochs. Early stopping uses official-development EER with patience four. The primary representation experiments use weighted binary cross-entropy with spoof encoded as class 1 and positive-class weight n0/max⁡(n1,1)n_0/\max(n_1,1), where n0n_0 and n1n_1 denote training-set class counts.

The nominal seeds are 1111, 1337, and 2026, with a recorded method offset of +300, giving effective seeds 1411, 1637, and 2326. For each configuration, the three raw head outputs are averaged before score orientation, normalization, and fusion. Reported ensemble EERs are computed from the averaged score stream.

Initialization Sensitivity

With L8–L21 and C=10−4C = 10^{-4} fixed, the recorded initialization check gives source Meta-test Macro EERs of 0.063%, 0.072%, and 0.075%. The recorded summary reports a mean of 0.070% and a sample standard deviation of 0.007 percentage points, compared with 0.149% for the XLSR-SLS baseline. When layer-pair selection is repeated, L8–L21 is selected in two runs and L9–L17 in the third; the latter obtains 0.092% Macro EER on Meta-test. These are source-side evaluations, separate from the reported external-dataset EERs.

Additional Detector Fusion Results

Table 11 reports relation-only scoring, equal-weight fusion of source-normalized scores, and source-selected logistic fusion for XLSR-Conformer-TCM and XLSR-AASIST. Both detectors use the layer indices selected with XLSR-SLS and detector-specific relation heads and fusion coefficients. The main results use the logistic-fusion configuration selected on source data.

DetectorBaseRelationEqualLogistic
TCM6.8496.8526.2956.316
AASIST7.1705.9805.8155.972

Table 11. Additional detector fusion results (External-5 mean EER, %; lower is better). Logistic denotes the source-selected configuration used in the main results.

External Evaluation and Reproducibility

External Evaluation Populations

Table 12 reports the aligned score-record populations for the primary XLSR-SLS representation comparisons. External-5 is the unweighted mean of the five dataset-level EERs, rather than an EER computed by pooling their trials.

External test setUtterances
ASVspoof 2021 LA181,566
ASVspoof 2021 DF611,829
ASVspoof 5 Eval680,774
Normalized Fake-or-Real test4,634
In-the-Wild31,212

Table 12. Aligned external score-record populations for the XLSR-SLS relation comparison.

For source and generic internal evaluation, we select the operating point that minimizes the absolute difference between the false-positive rate (FPR) and false-negative rate (FNR). At this point, EER is reported as 100(FPR+FNR)/2100(\mathrm{FPR}+\mathrm{FNR})/2. For ASVspoof 2021 LA and DF, we retain the EER values from the official evaluators.

Result Sources and Artifact Coverage

Table 13 summarizes the evaluation records and supporting artifacts in the assembled result archive. Reference systems are represented by EER summaries, while the XLSR-SLS relation comparison includes aligned score streams and source-side fusion records. The assembled archive contains source configurations and per-dataset evaluations for XLSR-Conformer-TCM and XLSR-AASIST, but not their auxiliary checkpoint-to-score mappings.

SystemAvailable records
XLSR-ConformerEER summary
Nes2Net-XEER summary
XLSR-DADGMoE-AASISTEER summary
BiCrossMamba-STEER summary
XLSR-SLS + CLRMAligned scores; fusion settings; source controls
XLSR-Conformer-TCM + CLRMDataset results; source configurations
XLSR-AASIST + CLRMDataset results; source configurations

Table 13. Evaluation records and supporting materials. Coverage refers to the assembled result archive.

Score Replay and Verification Scope

The XLSR-SLS L8–L21 records include aligned base, relation, and fused scores for all five external test sets. Reapplying the saved normalization and fusion parameters reproduces the archived relation-based scores with a maximum absolute difference

of 6.4×10−76.4 \times10^{-7}, without refitting. Relation and Concat use the same L8–L21 feature caches and standardization statistics, with independently trained heads and source-fitted fusion coefficients. Saved source score streams and fusion configurations also support replay of the corresponding source-side control metrics. For XLSR-Conformer-TCM and XLSR-AASIST, the retained per-dataset evaluation values allow recomputation of External-5 from unrounded results.

References

  1. [1]Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. 2022. XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale. In Interspeech 2022, pages 2278–2282.arxiv.org/abs/2111.09296
  2. [2]Benjamin Shiue-Hal Chou, Yi Zhu, and Surya Kopisetti. 2026. ICLAD: In-context learning with comparison-guidance for audio deepfake detection. In Findings of the Association for Computational Linguistics: ACL 2026, pages 9242–9256, San Diego, California, United States. Association for Computational Linguistics.
  3. [3]Yassine El Kheir, Tim Polzehl, and Sebastian Möller. 2025. BiCrossMamba-ST: Speech deepfake detection with bidirectional Mamba spectro-temporal cross-attention. In Interspeech 2025, pages 2235–2239.arxiv.org/abs/2505.13930
  4. [4]Yassine El Kheir, Younes Samih, Suraj Maharjan, Tim Polzehl, and Sebastian Möller. 2025. Comprehensive layer-wise analysis of ssl models for audio deepfake detection. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 4070–4082.
  5. [5]Wen Huang, Yanmei Gu, Zhiming Wang, Huijia Zhu, and Yanmin Qian. 2025. SpeechFake: A large-scale multilingual speech deepfake dataset incorporating cutting-edge generation methods. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9985–9998, Vienna, Austria. Association for Computational Linguistics.arxiv.org/abs/2507.21463
  6. [6]Wen Huang, Yuchen Mao, and Yanmin Qian. 2026. A data-centric approach to generalizable speech deepfake detection. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 17520–17539, San Diego, California, United States. Association for Computational Linguistics.
  7. [7]Jee-weon Jung, Hee-Soo Heo, Hemlata Tak, Hye-jin Shim, Joon Son Chung, Bong-Jin Lee, Ha-Jin Yu, and Nicholas Evans. 2022. AASIST: Audio anti-spoofing using integrated spectro-temporal graph attention networks. In 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6367–6371. IEEE.arxiv.org/abs/2110.01200
  8. [8]Yuang Li, Min Zhang, Mengxin Ren, Xiaosong Qiao, Miaomiao Ma, Daimeng Wei, and Hao Yang. 2024. Cross-domain audio deepfake detection: Dataset and analysis. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4977–4983, Miami, Florida, USA. Association for Computational Linguistics.arxiv.org/abs/2404.04904
  9. [9]Zhili Nicholas Liang, Soyeon Caren Han, Qizhou Wang, and Christopher Leckie. 2026. Hiercon: Hierarchical contrastive attention for audio deepfake detection. In Proceedings of the ACM Web Conference 2026, pages 8413–8416.arxiv.org/abs/2602.01032
  10. [10]Tianchi Liu, Duc-Tuan Truong, Rohan Kumar Das, Kong Aik Lee, and Haizhou Li. 2025. Nes2net: A lightweight nested architecture for foundation model driven speech anti-spoofing. IEEE Transactions on Information Forensics and Security, 20:120005–12018.arxiv.org/abs/2504.05657
  11. [11]Xuechen Liu, Xin Wang, Md Sahidullah, Jose Patino, Héctor Delgado, Tomi Kinnunen, Massimiliano Todisco, Junichi Yamagishi, Nicholas Evans, Andreas Nautsch, and Kong Aik Lee. 2023. Asvspoof 2021: Towards spoofed and deepfake speech detection in the wild. IEEE/ACM Trans. Audio, Speech and Lang. Proc., 31:2507–2522.DOI
  12. [12]Nicolas Müller, Pavel Czempin, Franziska Diekmann, Adam Frohgyar, and Konstantin Böttinger. 2022. Does Audio Deepfake Detection Generalize? In Interspeech 2022, pages 2783–2787.arxiv.org/abs/2203.16263
  13. [13]Zihan Pan, Tianchi Liu, Hardik B. Sailor, and Qiongqiong Wang. 2024. Attentive Merging of Hidden Embeddings from Pre-trained Speech Model for Anti-spoofing Detection. In Interspeech 2024, pages 2090–2094.arxiv.org/abs/2406.10283
  14. [14]Siqing Qin, Zhe Li, Kong Aik Lee, and Man-Wai Mak. 2026. Domain-adaptive dual-gating mixture of experts for generalizable speech deepfake detection. In Proc. Interspeech.arxiv.org/abs/2609.33709
  15. [15]Ricardo Reimao and Vassilios Tzerpos. 2019. For: A dataset for synthetic speech detection. In SpeD, pages 1–10.DOI
  16. [16]Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992. Association for Computational Linguistics.arxiv.org/abs/1908.10084
  17. [17]Eros Rosello, Alejandro Gomez-Alanis, Angel M Gomez, and Antonio Peinado. 2023. A conformer-based classifier for variable-length utterance processing in anti-spoofing. In Proc. Interspeech 2023, pages 5281–5285.DOI
  18. [18]Hemlata Tak, Massimiliano Todisco, Xin Wang, Jee weon Jung, Junichi Yamagishi, and Nicholas Evans. 2022. Automatic Speaker Verification Spoofing and Deepfake Detection Using Wav2vec 2.0 and Data Augmentation. In The Speaker and Language Recognition Workshop (Odyssey 2022), pages 112–119.arxiv.org/abs/2202.12233
  19. [19]Duc-Tuan Truong, Ruijie Tao, Tuan Nguyen, Hieu-Thi Luong, Kong Aik Lee, and Eng Siong Chng. 2024. Temporal-Channel Modeling in Multi-head Self-Attention for Synthetic Speech Detection. In Interspeech 2024, pages 537–541.DOI
  20. [20]Xin Wang, Héctor Delgado, Hemlata Tak, Jee weon Jung, Hye jin Shim, Massimiliano Todisco, Ivan Kukanov, Xuechen Liu, Md Sahidullah, Tomi H. Kinnunen, Nicholas Evans, Kong Aik Lee, and Junichi Yamagishi. 2024. ASVspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale. In The Automatic Speaker Verification Spoofing Countermeasures Workshop (ASVspoof 2024), pages 1–8.arxiv.org/abs/2408.08739
  21. [21]Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Héctor Delgado, Andreas Nautsch, Nicholas Evans, Md Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, et al. 2020. Asvspoof 2019: A large-scale public database of synthesized, converted and replayed speech. Computer Speech & Language, 64:101114.arxiv.org/abs/1911.01601
  22. [22]Yixuan Xiao and Ngoc Thang Vu. 2025. Layer-Wise Decision Fusion for Fake Audio Detection Using XLS-R. In Interspeech 2025, pages 5618–5622.DOI
  23. [23]Qishan Zhang, Shuangbing Wen, and Tao Hu. 2024. Audio deepfake detection with self-supervised xls-r and sls classifier. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6765–6773.

Paper details

vlnc:W_6xj65vby
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents