Kurtosis Normalization after Short-Time Gaussianization for Robust Speaker Verification

Statistical normalization has been popularly used to transform the distribution of the speech feature to decrease the mismatch between the training and test environment in robust speaker verification. This paper presents a method to normalization kurtosis which can measure the non-Gaussianity in quantity. The kurtosis normalization method proposed here is based on the cumulative distribution function (CDF) of a sub-Gaussian function and used after the short-time Gaussianization normalization method in feature space. Thus the distribution of speech parameters becomes more approximate the normal one. Experimental results based on the 2001 NIST SRE database show that when the short-time non-Gaussianity and kurtosis are both normalized, significant improvement could be achieved in not only equal error rate but also minimum detection cost compared to baseline system