How does one normalize data such that the Gaussian RBF kernel is not affected?

Question

I\'m training a support vector classifier on some speech data. In pre-processing I extract a bunch of informative aspects from the data and finally run it through a Discrete Cos