Investigating different representations for modeling and controlling multiple emotions in DNN-based speech synthesis