Reproduce “Stabilizing Short Duration Speaker Verification through Neural Re-scoring with Hybrid Enrollment” 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

Zhiqi Ai, Han Cheng, Shiyi Mu, Zhiyong Chen, Yongjin Zhou, Shugong Xu

The research introduces a hybrid enrollment and neural re-scoring framework to enhance short-duration speaker verification, particularly for user-defined keyword spotting. The method, leveraging both text-independent and text-dependent speech, consistently lowered Equal Error Rates across various speaker backbones, achieving, for instance, a 5.65% EER with ECAPA-TDNN compared to 9.27% for text-dependent enrollment with a verifier.

Key insights

  • The stability and informativeness of speaker representations are highly duration-dependent, with longer Text-Independent (TI) enrollment generally outperforming short Text-Dependent (TD) enrollment unless TI duration is very limited.
  • Frame-level interaction modeling via a neural verifier effectively refines similarity estimation in short-duration scenarios, improving performance even for pure Text-Dependent enrollment.
  • Combining the complementary strengths of TI (rich identity) and TD (lexical consistency) enrollment within a hybrid framework, coupled with neural re-scoring, leads to superior and more robust Short-Duration Speaker Verification.

Summary by alphaXiv.

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