New AI method detects out-of-distribution deepfake speech with 10% better accuracy

A team of researchers (Du, Yao, Kong, Cao) tackled the growing threat of deepfake speech impersonation by improving out-of-distribution (OOD) detection in vocoder recognition. Traditional methods rely on probability-score or classified-distance metrics, which struggle with samples near decision thresholds. Their proposed reconstruction-based approach uses an autoencoder architecture: it first extracts acoustic features from a pre-trained WavLM model, then compresses and reconstructs them. Each specific vocoder class has a dedicated decoder that only accurately reconstructs features from its own class. If no decoder reconstructs a feature well, it is flagged as OOD. To sharpen distinctions between classes, they added contrastive learning and an auxiliary classifier to constrain the reconstructions.

Experiments on evaluation datasets showed a 10% relative performance gain over baseline OOD detection methods. Ablation studies confirmed that both the contrastive constraint and the auxiliary classifier contribute significantly to the improvement. The paper was submitted to arXiv in June 2024, but was withdrawn in July 2026 due to an unresolved authorship dispute—some authors withdrew consent. Despite the withdrawal, the technical approach offers a novel direction for detecting unknown deepfake algorithms in audio, which is critical as synthetic speech becomes more convincing.