Snap Speaker Nulling Deepfake Detection

SNAP: Speaker Nulling for Deepfake Detection

Core Insight: Speaker Entanglement Problem

Audio deepfake detectors trained on self-supervised learning (SSL) representations exploit speaker-identity patterns instead of learning genuine synthesis artifacts. When models like WavLM encode speaker information strongly, detectors learn speaker-specific correlations rather than the acoustic cues that distinguish real from synthetic speech. This causes catastrophic failure when encountering new speakers or unseen TTS models.

Why Non-Obvious: Conventional wisdom assumes SSL representations separate "what is being said" from "who is speaking." In practice, speaker-identity encoding is so dominant that artifact-detection layers become speaker classifiers instead.

Minimal Recipe: Subspace Projection

Decompose feature space as ℋ = 𝒮 + 𝒜 + 𝒞 where:

• 𝒮 = speaker-dependent subspace
• 𝒜 = artifact-related subspace
• 𝒞 = context (utterance, channel) subspace

Three-Stage Pipeline:

1. Feature Extraction: Concatenate WavLM-Large layers 8 and 22, apply mean pooling, L2-normalize
2. Speaker Subspace Identification: Compute PCA on speaker centroids to find the speaker-dependent subspace
3. Nulling via Orthogonal Projection: Project features orthogonally to suppress speaker subspace while preserving artifact and context information
4. Classification: Train logistic regression (only 2,049 parameters) on nullified features

# Speaker subspace nulling via PCA projection
# 1. Compute speaker centroids from enrollment utterances
# 2. Center and PCA-decompose: C = U @ S @ V.T
# 3. For test feature x, compute orthogonal projection:
#    x_nulled = x - U @ U.T @ x
# This removes speaker-dependent variance while preserving other patterns

Results & Cross-Domain Robustness

ASVspoof 2019 LA (benchmark): 0.35% EER—56% improvement over WavLM-ECAPA baseline

In-the-Wild Dataset: 15.39% EER vs. 22.22% for WavLM baseline

Generalization to Unseen TTS:

• CosyVoice2 (not in training set): Maintains strong performance
• F5-TTS (new diffusion-based model): Effective detection despite architectural novelty

Key Finding: Nulling speaker subspace removes speaker-specific decision boundaries, forcing the classifier to rely on genuine synthesis artifacts that generalize across speakers and TTS models.

Deployment Considerations

1. Enrollment Phase: Collect 3-5 clean utterances from each reference speaker to establish speaker subspace
2. PCA Dimensionality: Determine via scree plot or cross-validation; typically 20-50 dimensions
3. Scaling: Method is lightweight—PCA + logistic regression runs on CPU in real-time
4. Generalization: Works best when training set includes diverse speakers; speaker diversity in training improves nulling robustness
5. Failure Mode: Performance degrades if speaker subspace overlaps with artifact subspace (rare but possible for certain TTS models); mitigate via larger enrollment sets

Practical Impact

• Simplicity: Minimal parameters (logistic regression) reduce overfitting to seen speakers/TTS models
• Interpretability: Orthogonal projection provides clear subspace decomposition; decisions are traceable
• Deployment: Fast inference, no need for large models; suitable for edge deployment