Those in the Blocked And Reported Book Club know me for a couple of things, but one is my complaint about how we’re pawns for Big Podcasting. We’re all caught up in the “Heterodox” sphere, and it’s kind of an incestuous group that’s always promoting other members, even while Jesse Singal tends to say "You can’t see it, but as I say ‘Heterodox’ I’m making the jerk-off motion with my hand”. We end up reading books by people in that sphere, or reading books recommended casually on podcast episodes, or listening to podcasts that have featured Katie or Jesse as guests, etc.
And so, as one of the pawns of Big Podcasting, I have to say that I got suckered and listened to “Sold A Story” at Katie’s endorsement. She devoted her entire segment of a January 2023 Premium episode to covering it (in condensed form)
and then she later interviewed the host of that podcast for a guest episode of Bari Weiss’s Honestly Podcast:
Anyway, I bring that all up for only one tiny reason. In the first episode, around 27:40, she brings up “decodable books”, which I immediately googled and the premise sounds pretty sound. Basically, you are building tutorial levels where you only learn a tiny number of the mechanics at a time, so you don’t get flooded with info and just give up on the game (and declare it a bad game). Seems pretty obvious that you should learn to walk on soft, level ground, before doing gymnastics while dodging lasers, and we’ve identified all kinds of relevant variables to minimize for the easy levels.
The letter A in English makes the ah sound and also the Ay sound, so why not just make a book that only does one instead of two. Also cut all wierd pairs like in “instead” or “pairs”. Also, pretend the letter G doesn’t exist for this book. Stuff like that. Only get practice with a few things for now.
From my perspective, I’ve been around Hatsune Miku and other vocoloids for a while, which synthesize speech and singing. And people always building shitty little IRC chatbots and spamming forums with slightly less shitty chatbots. Also, Text to speech more broadly has been around forever. And now we’re getting to ChatGPT and GPT-3 and we’ve got tacotron and other stuff.
So my question is: why hasn’t the decodable book idea been fully automated yet? It seems like it should be pretty easy to take a week or a month (or at least find a couple of grad students to take a sememester) to just make a big list of every letter, letter pair, and n-letter group (they work on -sion and -tion endings in the podcast), and all pronunciations of them, get them mapped onto a giant dictionary, and then spit out random stories with only the selected tiny set of phonemes and letters and sentence structures.
You should be able to churn out a million or so “decodable texts” pretty easily, it seems like.
The next step, of giving the stories to an AI painter, should also be similarly easy. The hardest part will likely be getting an app that will randomly generate a 10 page story on demand.
Anyway, I got a free copy of Paul Taylor’s Text-to-Speech Synthesis from 2011 just now, so maybe I’ll be the one to do it. Let me know if you are good at doing stuff and want to work together, or if you know of a better textbook to start reading than that.