Short Speaker Verification Based on Parzen Window Estimation
Yuhong Jin · Journal of Chinese Computer Systems · 2012
In the text-independent speaker verification,those mature and traditionally speaker verification systems with excellent performance are all based on the long duration training and testing data.Such as in the NIST SRE core tasks,both of the training and testing data are around 5 minutes.However,under the real application conditions,it's difficult to achieve or obtain such a long duration speech segments because of the security or the user experience property relate to speaker identity.These adverse conditions make a large performance degradation of the traditional systems.In this paper,considering the short duration property,we propose a direct modeling approach called Parzen Window modeling to estimate the target speaker model distribution which can achieve good generalization abilities.Experiments on NIST SRE2006 10s-10s task show a performance improvement with relative EER=10.76% degradation after combined the scores of GMM-UBM baseline.