Linear and nonlinear prediction for speech recognition with hidden Markov models