Reproduce “What and When to Learn: CURriculum Ranking Loss for Large-Scale Speaker Verification” 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
Massa Baali, Sarthak Bisht, Rita Singh, Bhiksha Raj
Curry ranks speaker-verification examples by cosine similarity to a labeled speaker’s best-matching sub-center, then adjusts the influence of Easy, Medium, and Hard tiers across three phases. Its score is a difficulty proxy, not a corruption label, and requires no additional forward pass. On VoxCeleb1-O, EER falls from 2.87% for “Baseline” to 0.38%, but inconsistent naming and absent ablations limit attribution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.