Wavelet Scattering + Spiking Neural Networks for Audio Deepfake Detection

An exploratory study of wavelet scattering transforms paired with spiking neural networks (SNNs) for cross-domain audio deepfake detection. Trained on ASVspoof 2019 LA, evaluated on the In-the-Wild benchmark (Müller et al., 2022).

Headline result: 25.65% In-the-Wild EER with a 136K-parameter wavelet+SNN model, competitive with augmented non-SSL systems in the same parameter class (AASIST-L: 26.86% at 85K params; AASIST: 23.55% at 297K params; Schäfer et al., 2024).

Key findings

The study was designed around three structural predictions, made before running the experiments:

1. SNN backends should benefit from temporally-preserving frontends more than CNN backends do, because SNN dynamics are natively temporal and spectrograms discard the fine temporal structure that SNNs are built to exploit.
2. Wavelet scattering is the right temporally-preserving frontend for this task, because wavelet decompositions capture transient artifacts that spectral averaging smears, the same property that makes them useful for industrial vibration analysis, where synthesis-artifact-like transients are the signal of interest.
3. Kaldi-style augmentation is the activating condition, because cross-domain generalization requires corrupting dataset-specific shortcuts before the wavelet features can be selected for.

The exploratory design (frontend × backend, with and without augmentation) was designed to test these predictions. All three landed:

• Wavelet scattering yields a 24pp improvement over mel with the SNN backend, vs. 17pp with the ResNet backend, the asymmetric interaction predicted by (1).
• Wavelet+SNN achieves a 2.9x smaller generalization gap (ASVspoof eval to ITW) than mel+SNN at matched training conditions, consistent with (2).
• Without Kaldi augmentation, all SNN frontends collapse to ~51% ITW EER (chance); with it, wavelet improves by 21-26pp while mel worsens, consistent with (3).

This is the first SNN evaluation on the In-the-Wild benchmark. The only prior SNN work in audio deepfake detection (SAFE, withdrawn from ICLR 2025) used spectrograms and was not evaluated on cross-domain real-world audio.

Results

Model	Params	ITW EER	Source
RawGAT-ST + aug	437K	18.08%	Schäfer et al. (2024)
WavLM-Large + linear probe	317M + 25K	19.61%	this work
AASIST + aug	297K	23.55%	Schäfer et al. (2024)
Wavelet + SNN + Kaldi	136K	25.65%	this work
AASIST-L + aug	85K	26.86%	Schäfer et al. (2024)
RawNet2, no aug	17.6M	33.94%	Müller et al. (2022)

Parameter counts for AASIST, AASIST-L, and RawGAT-ST are from Jung et al. (2022). Schäfer et al. numbers are from their preparatory tests (Table 1), trained on ASVspoof 2019 LA with Gaussian noise + MP3 compression augmentation. The WavLM-Large reference is a frozen linear probe over layer-weighted hidden states, not a fine-tuned system; fine-tuned SSL approaches currently reach substantially lower ITW EER (<10%) and are the state of the art; the comparison here is between non-SSL architectures in the 85K-437K parameter class.

What this is and isn't

This is a course project (DATASCI 266, UC Berkeley MIDS), not a peer-reviewed publication. Specifically:

• All results use a single random seed. Multi-seed replication is needed to make statistical significance claims about the interaction effects.
• The SNN topology, LIF dynamics, augmentation strategy, and training configuration were chosen from first-principles reasoning, not empirical search. The 25.65% ITW EER should be interpreted as a lower bound on what this architecture family achieves, not a tuned result.
• No neuromorphic hardware deployment. The SNN is simulated on GPU via snnTorch. Energy claims in the paper are theoretical and would require Loihi 2 or Akida deployment to validate.

Setup

pip install -r requirements.txt

Data

Download the datasets and place them under data/raw/:

# Check dataset status and get download instructions
python -m src.data.download_data

# Download MUSAN + RIR for Kaldi-style augmentation
python -m src.data.download_augmentation_data

Expected structure:

data/raw/
  asvspoof2019/LA/...
  in_the_wild/...
  musan/...
  RIRS_NOISES/...

Precompute features

Feature extraction (especially wavelet scattering) is slow. Precompute once, then train fast:

# Mel + wavelet features for train/dev
python -m src.data.precompute --config configs/default.yaml \
  --frontends mel wavelet_v3_order1 --splits train dev

# With Kaldi augmentation (5 copies per sample)
python -m src.data.precompute --config configs/default.yaml \
  --frontends wavelet_v3_order1 --splits train dev \
  --augment 5 --augment-mode kaldi \
  --musan-root data/raw/musan --rir-root data/raw/RIRS_NOISES

# Per-channel normalization stats
python -m src.data.precompute --config configs/default.yaml \
  --frontends wavelet_v3_order1 --splits train --compute-stats

Training

# Wavelet + SNN (best config)
python -m src.training.train \
  --frontend wavelet_v3_order1 --model snn \
  --precomputed --augmented --augment-mode kaldi --cross-eval

# Mel + ResNet baseline
python -m src.training.train \
  --frontend mel --model resnet \
  --precomputed --augmented --augment-mode kaldi

Evaluation

# Precompute eval features
python -m src.data.precompute --config configs/default.yaml \
  --frontends wavelet_v3_order1 --splits eval

# Evaluate on ASVspoof eval + In-the-Wild
python -m src.evaluation.evaluate_precomputed \
  --checkpoint results/checkpoints/wavelet_v3_order1_snn_kaldi/best.pt \
  --frontend wavelet_v3_order1 \
  --output-name wavelet_order1_snn_kaldi \
  --raw-datasets in_the_wild

SSL baseline

# Extract frozen WavLM-Large features (~317M params, GPU required)
python -m src.models.ssl_baseline extract --config configs/default.yaml

# Train layer-weighted linear probe
python -m src.models.ssl_baseline train --config configs/default.yaml

# Evaluate
python -m src.models.ssl_baseline eval --config configs/default.yaml

Paper

Full writeup in paper/draft.md (source) and paper/draft.pdf (rendered).

Citation

@misc{collins2026wavelet,
  title={Wavelet Scattering and Spiking Neural Networks for Cross-Domain Audio Deepfake Detection},
  author={Collins, Micah},
  year={2026},
  note={DATASCI 266, UC Berkeley MIDS}
}

References

• Andén, J., & Mallat, S. (2014). Deep scattering spectrum. IEEE Trans. Signal Processing, 62(16), 4114-4128.
• Andreux, M., et al. (2020). Kymatio: Scattering transforms in Python. JMLR, 21(60), 1-6.
• Anonymous (2024). SAFE: Spiking neural network-based audio fidelity evaluation. ICLR 2025 submission (withdrawn). OpenReview: QWDZE2mYIe.
• Eshraghian, J. K., et al. (2023). Training spiking neural networks using lessons from deep learning. Proceedings of the IEEE, 111(9), 1016-1054.
• Jung, J.-W., et al. (2022). AASIST: Audio anti-spoofing using integrated spectro-temporal graph attention networks. Proc. ICASSP, 6367-6371.
• Müller, N. M., Czempin, P., Dieckmann, F., Froghyar, A., & Böttinger, K. (2022). Does audio deepfake detection generalize? Proc. Interspeech.
• Schäfer, K., Neu, M., & Choi, J.-E. (2024). Robust audio deepfake detection. Proc. ASVspoof Workshop.
• Snyder, D., Garcia-Romero, D., Sell, G., Povey, D., & Khudanpur, S. (2018). X-vectors. Proc. ICASSP, 5329-5333.
• Tak, H., et al. (2021). End-to-end anti-spoofing with RawNet2. Proc. ICASSP, 6369-6373.
• Wang, X., et al. (2020). ASVspoof 2019. Computer Speech & Language, 64, 101114.
• Xuan, X., et al. (2026). WST-X series. arXiv:2602.02980.
• Yamagishi, J., et al. (2021). ASVspoof 2021. Proc. ASVspoof Workshop.