Nonlinear discriminant analysis for improved speech recognition