← Back to brief
ResearchOfficialPreprintarXiv Audio and Speech Processing

Reconstruction-Based Out-of-Distribution Detection for Vocoder Recognition

A new reconstruction-based method for out-of-distribution (OOD) detection in vocoder recognition has been proposed, utilizing an autoencoder to compress and reconstruct acoustic features from a pre-trained WavLM model. The system assigns each vocoder class a dedicated decoder, and samples that cannot be well reconstructed by any decoder are flagged as OOD. The approach achieves a 10% relative improvement over baseline systems on evaluation datasets, with contrastive learning and an auxiliary classifier further enhancing feature distinctiveness.

Why it matters: This method advances the detection of unknown vocoder-generated deepfakes, addressing a critical challenge in anti-spoofing technology.

Full story at: arXiv Audio and Speech Processing