An Euclidean distance measure between covariance matrices of speech cepstra for text-independent speaker recognition

It has been shown that similarity measures between covariance matrices of speech cepstra provide good speaker recognition. We propose a transformation from covariance matrices to vectors in Euclidean space, which allows the use of the Euclidean distance measure. We show how this measure is related to the Gaussian log-likelihood measure between covariance matrices and we compare the performance. The Euclidean measure has comparable performance and allows various manipulations in Euclidean space