Objective-driven discriminative training and adaptation based on an MCE criterion for speech recognition and detection