Source-linked AI summary
WST-Graph: Topology-Preserving Wavelet Scattering Front-End for Speech Deepfake Detection
Kwok-Ho Ng, Tingting Song, Bingwen Feng, Zhihua Xia
TL;DR
Speech deepfake detectors can degrade under unseen generators and recording conditions, while existing front-ends may lose explicit acoustic coordinates. WST-Graph preserves scattering paths in a sparse carrier–modulation grid for AASIST graph processing, reporting competitive performance with fewer parameters and gains on selected out-of-domain benchmarks.
Problem
Detectors can degrade under unseen generators and recording conditions, while SSL front-ends use substantially more parameters and lack explicit acoustic coordinates for artifact analysis.
Source: §1 · p. 1Method
WST-Graph organizes WST paths into a sparse carrier–modulation grid and uses modulation-level normalization and length-aware pooling before AASIST-style graph processing.
Source: §1 · p. 1, §2.1 · p. 2, §2.2 · p. 3Results
On 7 of 13 OOD sets, WST-Graph-Q81 has the lowest EER among four systems; WST-Graph-Q82 scores 2.92% EER on ASVspoof2019 LA with 59.6% fewer parameters than AASIST.
Source: §3.5 · p. 4Takeaways & Limitations
The study identifies representation geometry and information aggregation as jointly shaping detection performance.
Source: §4 · p. 4
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
Speech deepfake detectors face increasingly natural synthetic speech and can degrade under unseen generators and recording conditions. WST-Graph addresses the loss of explicit acoustic coordinates by preserving scattering paths' carrier–modulation structure for graph-based processing.
1. INTRODUCTION
- Detectors can degrade under unseen generators and recording conditions, while SSL front-ends use substantially more parameters and lack explicit acoustic coordinates for artifact analysis.These issues motivate a compact front-end that retains interpretable acoustic structure.Source: §1 · p. 1
- AASIST's pooling and residual convolutions mix neighboring frequency-time responses, so its channels lose explicit carrier and modulation-frequency coordinates before graph construction.The paper seeks to delay fusion across distinct spatiotemporal pathways.Source: §1 · p. 1
- WST produces stable multiscale coefficients, with second-order paths explicitly linked to their parent carrier bands.First-order coefficients capture averaged wavelet-modulus responses, while second-order coefficients characterize temporal modulations of those envelopes.Source: §1 · p. 1
- Unlike WST-X1's global pooling and WST-X2's spatial flattening, WST-Graph retains 1-D scattering topology as a structured interface for graph backends.It groups second-order paths by modulation band, aligns them to parent carriers, and uses length-aware pooling before learned adaptation.Source: §1 · p. 1
2. PROPOSED METHODS
The methods transform first- and second-order WST paths into a sparse carrier–modulation grid, pool valid frames into fixed relative-time bins, and adapt the resulting channels for AASIST-style graph classification.
2.1. Scattering Paths and Path Normalization
- First- and second-order WST coefficients are stacked as flat paths; PathNorm log-compresses their amplitudes and standardizes paths by modulation-level groups while excluding padded frames from training statistics.The log modulation policy groups first-order paths together and second-order paths by modulation band, preserving carrier-dependent variation.Source: §2.1 · pp. 1–2
2.2. Topology Recovery via OrderGrid
- OrderGrid maps normalized paths into a sparse modulation–carrier tensor, retaining valid coefficients and marking structurally absent path combinations with a fixed binary mask.The mask distinguishes topology-defined vacancies from measured zero amplitudes and is used to exclude absent cells from pooling.Source: §2.2 · pp. 2–3
2.3. Length-aware Channel-wise Temporal Pooling
- Length-aware channel-wise ALAP pools each valid trajectory within dynamically partitioned valid-frame intervals, contracting variable duration to a fixed K-bin grid.The pooling operation excludes structurally absent cells and yields a fixed-size representation.Source: §2.3 · p. 3
2.4. Channel Adaptation and Graph Classification
- A pointwise adapter first mixes scattering channels at each carrier–time coordinate; latent blocks preserve grid dimensions before GraphAASIST forms spectral and temporal nodes for AASIST graph attention and classification.GraphAASIST uses the AASIST spectro-temporal graph architecture to model interactions between carrier and relative-time nodes.Source: Figure 1 · p. 2, §2.4 · p. 3
3. EXPERIMENTS AND RESULTS
WST-Graph’s experiments evaluate temporal reduction, normalization, model capacity, and scattering resolution before comparing the final configurations with AASIST. The final results report a smaller model on ASVspoof2019 LA and competitive performance across out-of-domain datasets.
3.1. Implementation Details
- Models were trained on ASVspoof 2019 LA at 16 kHz for 12 epochs across three seeds, using 4-second training and development crops and onset-aligned evaluation.Cross-dataset generalization was evaluated on 14 Speech DF Arena sets, including 13 OOD sets.Source: §3.1 · p. 3
3.2. Temporal Reduction and Local Context
- At K = 64, ALAP delivers 19.14% average EER and 14.91% overall best EER; within the depthwise family, D3 (N = 3) performs optimally, with 4.98% average EER and 100k parameters.ALAP introduces only 49 parameters.Source: §3.2 · p. 3
3.3. Normalization and Graph-Node Construction
- Modulation-level mean centering and scaling achieve 3.80% average EER, while attentive-statistics pooling achieves 3.25%, versus 3.87% for absmax.Attentive statistics adds 19k parameters and is retained.Source: Table 3 · p. 4, §3.3 · p. 4, Table 4 · p. 4
3.4. Model Capacity and Scattering Resolution
- Increasing relative-time bins K is more consistently beneficial than expanding feature width C; K = C = 64 gives the lowest EER, while K = 64, C = 32 yields 3.39% EER with 83k parameters.The experiments also vary averaging scale J and first-order filterbank resolution Q1, and compare second-order resolutions.Source: §3.4 · p. 4, Table 5 · p. 4
3.5. Baseline Comparison and OOD Evaluation
- 2.92% EER on ASVspoof2019 LA with 120k parameters (59.6% fewer than AASIST); across 13 OOD datasets, WST-Graph-Q81 has the lowest EER on seven.Its most pronounced advantage is on DFADD; across the remaining domains, it remains generally competitive without consistently large margins.Source: Table 6 · p. 4, §3.5 · p. 4
4. CONCLUSION
WST-Graph is a topology-preserving scattering front-end for graph-based speech deepfake detection. The study finds that detection performance is sensitive to representation and aggregation choices, and proposes using preserved carrier-modulation coordinates in future attribution and source-tracing work.
4. CONCLUSION
- Detection performance is sensitive to temporal granularity, local-context extent, path calibration level, graph-node statistics, and modulation resolution.The sequential study highlights how representation geometry and information aggregation jointly shape detection.Source: §4 · p. 4
- Future work will use the preserved, physically interpretable carrier-modulation coordinates for decision attribution and deepfake source tracing.Source: §4 · p. 4
- The paper introduces WST-Graph as a topology-preserving scattering front-end for graph-based speech deepfake detection.Source: §4 · p. 4
5. COMPLIANCE WITH ETHICAL STANDARDS
The study uses publicly available datasets and did not collect new data from human participants; ethical approval was not required.