A novel strategy for speaker verification based on SVM classification of pairs of speech sequences

We introduce a novel strategy for speaker verification based on the conception of a classifier which is independent of the target speaker, as opposed to traditional systems where the classifier is always target dependent. The basic principle is to build a system that decides whether two sequences were pronounced by the same speaker. In our view, this system is aimed to complement traditional ones. We borrow techniques from the speaker segmentation area, namely the Bayesian Information Criterion (BIC), to conceive a kernel between pairs of sequences. We then use this kernel to implement our new system in an SVM scheme. We present experiments on NIST SRE data using the Biosecure project protocol. The individual performance of the new system is poor as compared to the baseline UBM-GMM and the GLDS-SVM. However, as expected, the fusion leads to better performances.