Reproduce “Layer-Wise Decision Fusion for Fake Audio Detection Using XLS-R” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yixuan Xiao, Ngoc Thang Vu
Researchers at the University of Stuttgart introduced a layer-wise decision fusion architecture for fake audio detection using XLS-R representations, achieving an Equal Error Rate of 6.90% on the In-the-Wild dataset. This approach leverages per-layer classifiers and decision fusion to improve generalization to unseen spoofing attacks compared to traditional feature fusion methods.
Key insights
- Layer-wise decision fusion mitigates 'feature collapse' by consistently utilizing information across a broader range of deep model layers, unlike feature fusion which tends to sparsely rely on a few layers.
- Training without leading and trailing silence, while degrading in-domain performance, significantly improves cross-dataset generalization for FAD systems, indicating that silence in training data can contain dataset-specific shortcuts.
- Different layers of the XLS-R model extract distinct acoustic and phonetic features, suggesting a hierarchical specialization that is better leveraged by per-layer decision-making.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.