Conditional random fields for hierarchical segment selection in text-to-speech synthesis