A Fully Time-domain Neural Model for Subband-based Speech Synthesizer