Reproduce “LISE : Listenable Interpretable Speaker Embeddings” 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

Xiaoliang Wu, Chongxin Gan, Ke Liu, Peter Bell, Jennifer Williams

LISE decomposes pretrained speaker embeddings into graded, non-negative components. In a familiarization-based listening task, participants distinguished speakers associated with high- versus low-weight references with 83.9% accuracy, though 3 of 35 components captured language patterns. On VoxCeleb1-O, reconstructed ECAPA-TDNN embeddings had 2.10% equal error rate versus 1.80% originally, a modest loss in verification performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.