Current speech deepfake detectors treat each utterance as a flat sequence of frames, limiting their ability to capture long-range spectral inconsistencies that span non-adjacent regions of an audio signal. We propose GraphSeal, which constructs a temporal-spectral graph over short speech segments and propagates artifact cues through graph attention layers before a global readout produces the bonafide-vs-spoof decision. Each node in the graph corresponds to a fixed-length segment; edges encode temporal adjacency and spectral similarity, letting the network attend to distant frames that share suspicious spectral patterns. A multi-head graph attention mechanism weights neighboring nodes according to both learned features and edge attributes, and a hierarchical readout aggregates node representations into a single utterance-level embedding for binary classification. GraphSeal is front-end agnostic: it operates on top of any self-supervised or handcrafted feature extractor. On ASVspoof 2019LA, ASVspoof 2021LA, ASVspoof~5, and the In-the-Wild dataset, GraphSeal achieves equal error rates of 0.52%, 2.41%, 8.17%, and 5.24%, respectively, matching or outperforming prior methods that rely on fixed-topology graph networks or purely sequential architectures.
Figures & tables
Figure 1: GraphSeal overview. A waveform is encoded by an SSL front end and segmented into overlapping windows. A temporal-spectral graph is constructed with edge attributes encoding adjacency and spectral similarity. L graph attention layers propagate artifact cues, and a hierarchical readout produces the utterance-level decision.
Method
19 LA
21 LA
ASV5
ItW
Non-SSL front ends
RawNet2 [ 54 ]
5.13
9.26
22.74
32.15
LCNN [ 78 ]
5.81
15.62
28.39
36.72
AASIST [ 24 ]
0.83
9.57
18.91
12.47
RawGAT-ST [ 52 ]
1.06
8.31
17.53
11.83
SAMO [ 11 ]
0.68
7.83
16.27
10.52
Table 1: EER (%) on four deepfake detection benchmarks. Lower is better. Bold and underline mark the best and second best. SSL-based methods use WavLM-Large as the front end unless noted.
Figure 2: Per-attack EER (%) on ASVspoof 2019 LA evaluation set. GraphSeal reduces EER on 12 of 13 attack types compared to WavLM-AASIST, with the largest gains on waveform-concatenation (A17) and neural vocoder (A10, A16) attacks.
Front end
21 LA
ItW
Whisper-Medium [ 44 ]
3.18
6.79
wav2vec 2.0-Large [ 3 ]
2.93
6.12
WavLM-Large [ 6 ]
2.41
5.24
Table 2: EER (%) with different SSL front ends on ASVspoof 2021 LA and In-the-Wild. All use the GraphSeal back end.
Variant
21 LA
ItW
GraphSeal (full)
2.41
5.24
No spectral edges ( ks=0 )
3.12
6.53
No temporal edges ( kt=0 )
3.97
7.41
No edge attributes ( eij=1 )
2.78
5.89
No paralinguistic features
2.63
5.57
Mean-pool readout only
2.71
5.81
Table 3: Ablation study on ASVspoof 2021 LA and In-the-Wild (EER %). All variants use WavLM-Large.
w
ks
21 LA
ItW
16
5
2.73
5.58
32
5
2.41
5.24
64
5
2.56
5.47
128
5
2.81
5.72
32
1
2.87
5.91
32
3
2.58
5.49
Table 4: Sensitivity to segment window w and spectral neighbors ks on ASVspoof 2021 LA and In-the-Wild (EER %). Default setting ( w=32 , ks=5 ) is highlighted.
Figure 3: t-SNE visualization of utterance embeddings on ASVspoof 2021 LA. GraphSeal (right) produces tighter, better-separated clusters compared to WavLM-AASIST (left), reflecting its stronger discriminative power.
Back end
Params
Latency
EER
Attn. pool [ 33 ]
0.1M
2 ms
3.42
AASIST [ 24 ]
0.3M
8 ms
2.85
GraphSeal
2.1M
12 ms
2.41
Table 5: Back-end efficiency comparison. Parameters and latency are for the trainable back end only (excluding the frozen SSL encoder). EER is on ASVspoof 2021 LA.
Previous speech deepfake detectors can adaptively capture spectro-temporal dependencies through graph attention, yet they largely overlook the co-variation between spectral and temporal representations. To address this gap, we construct a normalized affinity graph from their joint covariance and apply polynomial graph filtering to capture higher-order covariance-induced dependencies. We first develop Cov-ST to isolate the contribution of covariance-based relational modeling. Although it improves detection performance, its sensitivity to the polynomial order suggests limited robustness when covariance relations are modeled alone. We therefore propose DuRe-ST, which jointly exploits covariance-induced and graph-attention-induced relations to capture complementary second-order co-variation and adaptive spectro-temporal dependencies. Experiments show that DuRe-ST achieves an average relative EER reduction of 25.9% over XLSR-AASIST on the ASVspoof benchmarks and 28.4% across four cross-dataset benchmarks with only 4-8k additional trainable back-end parameters. It further outperforms the strongest publicly available comparison models by 2.2-13.2% in relative EER on four benchmarks, while remaining smaller than the publicly available models considered.
Shaole Li, Siqing Qin, Youzhi Tu +1
Department of Electrical and Electronic Engineering The Hong Kong Polytechnic University, Hong Kong SAR
The acoustic front-end determines which forensic cues a speech deepfake detector can exploit. The wavelet scattering transform (WST) provides stable multiscale coefficients with explicit coordinates, yet direct flattening obscures the parent relation between paths. We introduce WST-Graph, reconstructing these paths as a sparse modulation-carrier grid for an AASIST graph backend. Modulation-level normalization and length-aware adaptive local attention pooling produce fixed relative-time representations while retaining the acoustic axes before learned adaptation. This yields a waveform-to-graph interface with a fixed, parameter-free WST. Our configurations remain competitive with AASIST while using approximately 60% fewer trainable parameters and show clear gains on selected out-of-domain benchmarks. These results underscore the value of preserving parent-child relations within the carrier-modulation topology when constructing a compact, physically grounded interface for graph-based speech deepfake detection. Code will be released at https://github.com/saki-ciallo/wst-graph.
Kwok-Ho Ng, Tingting Song, Bingwen Feng +1
College of Cyber Security, Jinan University, Guangzhou, China
Audio deepfakes generated by neural text-to-speech and voice-cloning systems threaten speaker verification and public discourse at scale. The core challenge is cross-dataset generalization: detectors trained on one synthesis pipeline collapse on unseen forgeries. We argue that this failure is primarily because of structural synthetic speech artifacts which are multi-timescale trajectory anomalies. Though every existing detector aggregates a fixed-window frame statistics, this misaligns the architecture with the signal. We propose FlowFake, a Liquid Time-Constant (LTC) architecture whose hidden state evolves via a learned ODE, with per-neuron adaptive time constants simultaneously resolving spectral (10ms) and prosodic (2s) cues. At only 34K parameters FlowFake achieves formal BIBO stability and O(dt^4) integration error. On a four-dataset cross domain benchmark (ASVspoof2019-LA, FakeOrReal, InTheWild, MLAAD), FlowFake reaches 75.29% on ASVspoof2019 trained only on FakeOrReal and 79.97% trained only on MLAAD. It outperforms RawGAT-ST and Whisper-DF on every evaluated pair and matching SSL Wav2vec2 (300x larger) at 0.01% of its parameter count. The source code is available on : https://github.com/GhostRider2023/FlowFake