Blind Signal Processing techniques have been receiving an increasing attention by the scientific community also for their benefic impact to speaker recognition based biometric systems. This work deals with the speech separation problem in blind conditions and in presence of more sources than sensors and Post-Nonlinear (PNL) mixing, which likely represent a close-to-reality situation. The addressed method is made of three separate steps: compensation of nonlinearity, mixing matrix recovery and final unknown source estimation. It has been recently proposed and successfully evaluated in the case of synthetic mixtures of real world data (like speech signals). Here, the Extended Gaussianization is employed as first step instead of the common Gaussianization one in order to reduce the approximation error on the linearized mixture pdfs. Computer simulations allowed achieving a significant improvement of separation performances over the previous approach.