Self-supervised speech models produce representations that bundle together what is said and who says it. The updated arXiv paper describes an SVD-based factorisation intended to separate these two strands. The method decomposes features into a shared content matrix that tracks temporal variation, leaving speaker-related structure in a complementary subspace.
By interpreting the factorised subspaces, the work aims to make learned speech features more understandable, and potentially more useful for downstream tasks that need to suppress or exploit speaker identity. The abstract does not include quantitative results, so the claims are limited to the proposed decomposition and its interpretability.