Optimal discriminative training for HMMs to recognize noisy speech