Kernel Approximation Methods for Speech Recognition