Initial investigation of encoder-decoder end-to-end TTS using marginalization of monotonic hard alignments