Audio deepfake detector uses speech trajectory dynamics, hits 0.75% EER on ASVspoof
A causal LSTM learns real speech dynamics and beats supervised baselines on deepfake benchmarks.
A new paper from researcher Tomás Andrade Weber proposes a self-supervised approach to audio deepfake detection that leverages the temporal dynamics of human speech. The method trains a causal Long Short-Term Memory (LSTM) next-frame predictor on bonafide speech only, using the deepfake-specialized SSL backbone Wav2Vec2-Large-AntiDeepfake. By comparing against a static global-average-pooling baseline with identical features, the study isolates the contribution of temporal modeling. A second supervised stage uses a Multi-Layer Perceptron on frozen LSTM states to characterize the role of spoof supervision.
The system achieves state-of-the-art performance across six benchmarks: ASVspoof 2019/2021, Codecfake, In-the-Wild, MLAAD-EN, and Deepfake-Eval-2024. Notably, it reaches 0.75% EER on ASVspoof 2021 — the best published result. Even more striking, the unsupervised Stage 1, trained only on real speech, outperforms the supervised baseline from the same backbone on Deepfake-Eval-2024 (30.35% EER). While static and dynamic approaches perform similarly on near-domain benchmarks, trajectory dynamics deliver substantial gains on hard cross-corpus tests with diverse synthesis methods, confirming that temporal physiological constraints can expose synthetic speech beyond what utterance-level statistics capture.
- Best published EER of 0.75% on ASVspoof 2021, using trajectory dynamics from a causal LSTM next-frame predictor
- Unsupervised Stage 1 (bonafide-only training) beats a supervised baseline on Deepfake-Eval-2024 with 30.35% EER
- Delivers major gains on cross-corpus benchmarks with diverse synthesis methods, unlike static pooling approaches
Why It Matters
This shows unsupervised learning of speech physiology can outperform supervised deepfake detection, making robust audio forensics more scalable and harder to fool.