Hybrid NN/HMM-Based Speech Recognition with a Discriminant Neural Feature Extraction