Looking for the right Hyperplane Algorithm

Hello,

I have a two sets of speech instances with 39 floating point features. I believe these sets are not linearly separable.

I want to find a hyperplane that does the "best job" possible of separating them. By "best job" I mean that on average the instances are best separated in distance from the plane. I do not care about classification performance.

I am considering a simple Perceptron or SVM. How do these algorithms perform for data that is not separable? Is there some other algorithm I should try?

Thank you
Peter