Discriminative and adaptive training for robust speech recognition and understanding