[adrianbg]

Yes, exactly. I do wonder whether a similarly good end-to-end system could be trained by constraining the alignments as I've seen done in some papers.