SONAR: Spectral-Contrastive Audio Residuals for Generalizable Deepfake Detection

July 21, 2026 | Source: arXiv AI

Tags: deepfake detection, audio AI, SONAR, ASVspoof, contrastive learning, voice cloning

SONAR sets state-of-the-art on ASVspoof 2021 and in-the-wild audio deepfake benchmarks by explicitly disentangling low-frequency content from high-frequency synthesis artifacts using spectral contrastive learning — and converges four times faster than strong baselines while remaining architecture-agnostic.

Details

Most audio deepfake detectors fail to generalize to out-of-distribution synthetic audio because they underexploit high-frequency (HF) spectral artifacts left by speech synthesis models. SONAR (Spectral-cONtrastive Audio Residuals) addresses this by running two parallel signal paths: an XLSR encoder captures dominant low-frequency content, while a cloned path with learnable SRM high-pass filters distills faint HF residuals. Frequency cross-attention then merges the two representations to capture both long- and short-range frequency dependencies. A frequency-aware Jensen-Shannon contrastive loss pulls real audio content-noise pairs together while pushing synthetic embeddings apart, splitting the latent space into two disjoint manifolds — one for genuine audio, one for synthetic — sharpening decision boundaries.\n\nOn ASVspoof 2021 and in-the-wild benchmarks, SONAR achieves state-of-the-art performance and converges four times faster than strong baselines. The frequency-guided contrastive design is architecture-agnostic, meaning the HF residual approach can be integrated into existing detection pipelines without rebuilding them. Given growing use of voice cloning for fraud and disinformation, improved generalizability across unseen TTS systems is the key practical contribution here.