Discriminative training of hierarchical acoustic models for large vocabulary continuous speech recognition