You could say the same thing about implementing the graphical portion of the game using system widgets rather than custom graphics code. Or about implementing your roguelike as a terminal application rather than a graphical one (but my personalized font!).
I don’t really see how you make an effective audio UI without direct control over the TTS. Simple, critical, things like stopping speech in progress, synchronizing non-speech indicators with the speech stream, adjusting the mix of music and sound effects while the game is speaking–screen readers can’t support you in that, can they?
If we relax the assumption that most people who would play with speech enabled are super-hackers with screen readers dialed up to 500 WPM, polish and fundamental usability should come way before customizability. There’s a reason audiobooks are recorded by human professionals, and blazing through ebooks with a caffeinated screen-reader is a niche activity.
Even considering the super-hackers, the lesson isn’t necessarily that all speech output is ideally in sixth gear. High-speed speech is useful for certain purposes. Fast navigation, for instance. Skimming. I’m a big fan of cranking up the speed of a video if I feel it’s moving slowly, but I wouldn’t watch something of an artistic nature that way. If your game isn’t fun at its default speech rate, polish it to make it fun, rather than prioritizing the workaround.
Maybe in some cases, accessibility is unapologetically an afterthought. In those cases, it makes sense to lean on the screen reader and just do what it takes to make the game navigable. But if we’re talking about “an accessability-focused text adventure system”, why would compromising on polished audio be acceptable? Get the speech going on day one, don’t target the screen-reader experts (and wind up with a game that’s only playable to screen-reader experts), make a game that even you, with your aural gaming inexperience, can operate.
IMO it would be a very good thing if everybody doing UI design stopped kicking the can down the road to screen readers, and started designing UIs with the understanding that some users would operate them via audio. Maybe not a lot of chance of this, but it’s amazing that we’re just expecting some third party to come along and interpret arbitrary graphical output into speech. Anybody writing a GUI framework could have made it self-narrating, and could have designed its layout widgets to encourage narratable, keyboard-navigable design. Why is it so easy to write inaccessible programs? Just not on the radar though.