A Hybrid HMM-Based Speech Recognizer Using Kernel-Based Discriminants as Acoustic Models