Initial investigation of an encoder-decoder end-to-end TTS framework using marginalization of monotonic hard latent alignments.