Maximum Gaussianality training for deep speaker vector normalization

Yunqi Cai, Lantian Li, Andrew Abel, Xiaoyan Zhu, Dong Wang

Abstract: Highlights•Proposed a novel maximum Gaussian criterion for training deep normalization models.•Our criterion directly enforces the necessary conditions of a Gaussian distribution.•Solves a number of critical issues associated with the conventional ML criterion.•Significant performance improvement was obtained when applying our new criterion.