Reproduce “SoCov: Semi-Orthogonal Parametric Pooling of Covariance Matrix for Speaker Recognition” 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
Rongjin Li, Weibin Zhang, Dongpeng Chen, Jintao Kang, Xiaofen Xing
Researchers from VoiceAI Technologies introduce SoCov, a semi-orthogonal parametric pooling method for speaker recognition that efficiently compresses covariance matrices while preserving cross time-frequency information, achieving a 15.5% relative reduction in Equal Error Rate on the SRE21Eval dataset compared to conventional x-vector systems.
Key insights
- Semi-orthogonal constraints consistently improve performance when vectorizing covariance matrices
- Combining covariance information with self-attentive pooling captures more discriminative speaker characteristics
- The method can be integrated into various neural network architectures like ResNet and ECAPA-TDNN
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.