Moreover, if you're going to argue the "nobody's actually going to do it" argument, I'd say that "nobody's actually using CSS to generate one HTML page that works in a screen reader, cell phone, rich browser, and REST API" either. For the same value of "nobody", which is to say, not literally nobody, you can find a handful of people, but the idea that there's some mass of developers "correctly" using CSS is pretty silly too.
Semantic HTML on the other hand is accessible by default, you have to specifically go out of your way to make it inaccessible or platform-specific.
Where the author suggested outputting audio when confronted with a "blind" header, I thought more along the lines of returning pure semantic markup. And presumably the abstract code that compiles down to html is purely semantic so it would be trivial.
Just like you could code assembly by hand (and plenty of people still do) you can code HTML by hand. But in reality most html is generated by machine and then 'rendered' into a visual representation, the final binary end-product.
The beauty of correct use of html and css is that the burden of choosing an appropriate rendering is pushed to the client, which in turn means that the developer does not have to know in advance what kind of devices or assistive technologies will be used to render the content.