I was a little surprise to read that they're using speech-to-text and text-to-speech rather than an end-to-end speech model. Won't that horrible latency? (I guess the old-person persona disguises it a little...)