CONTEXTUAL ADDITIVE STRUCTURES IN HMM-BASED SPEECH SYNTHESIS