I wonder if sub-vocal will happen anytime soon, and if it does, I am still not sure how it will solve voice being sub-optimal for some types of text. One example is plain old programming, and needs a fairly elaborate setup. The best example I've been able to find was Tavis Rudd's presentation [1], which seems to be a decent solution.

I have tried to more or less replicate this setup, but found it less satisfactory. Aside from it demanding near-perfect audio quality to work remotely well in general (as well as a heavy time investment, initial and ongoing), you frequently run into edge cases where it just doesn't want to understand something you're saying. After the 5th or 6th mis-transcription of some specific phrase, you tend to just give up and reach for the keyboard, RSI be damned. (And this as a native English-speaker, with a relatively neutral accent.)

I can definitely imagine better fleshed out solutions to either voice recognition for programming, or programming languages whose syntax lends themselves to easier and more productive voice recognition. I could see some definite productivity enhancements if it were achieved.

[1] https://www.youtube.com/watch?v=8SkdfdXWYaI