Reproduce “Interpreting SSL Representations for Spoof Detection: a WavLM Study” 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
Mohamed Mallat, Michele Panariello, Massimiliano Todisco, Nicholas Evans, Anthony Larcher
The authors examine how frozen WavLM Base layers distinguish genuine from spoofed speech and align with acoustic features. They use linear probes across layers and correlate representations with 17 handcrafted measurements, finding that many feature alignments peak mid-layer while jitter, shimmer, and HNR rise later. On ASVspoof 5, a selected five-layer subset reaches 6.1% evaluation EER versus 8.08% for all layers, in this benchmark and setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.