Acoustic Modeling for Speech Synthesis: from HMM to RNN