Reproduce “Attention Back-end for Automatic Speaker Verification with Multiple Enrollment Utterances” 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

Chang Zeng, Xin Wang, Erica Cooper, Xiaoxiao Miao, Junichi Yamagishi

Researchers at the National Institute of Informatics developed an attention-based back-end for Automatic Speaker Verification, which adaptively aggregates information from multiple enrollment utterances to create a more robust speaker representation. This model consistently outperformed traditional PLDA and cosine similarity back-ends, achieving an EER of 8.93% on the CNCeleb dataset when combined with an ECAPA-TDNN encoder.

Key insights

  • Utterance-level attention mechanisms effectively aggregate information from multiple speaker embeddings, creating a more robust speaker representation by adaptively weighing diverse enrollment samples.
  • The proposed attention back-end demonstrates complementarity with frame-level attention modules already present in advanced speaker encoders (e.g., ECAPA-TDNN), leading to enhanced overall performance.
  • A detailed ablation study reveals that both Scaled-Dot Self-Attention and logistic regression for score calibration are critical components for the model's superior performance, and the combined BCE and GE2E losses optimize across different ASV metrics.

Summary by alphaXiv.

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