Combining an engine to interpret voice commands with textual keyboard input seems like the best approach here. You get all the useful fuzziness of voice with none of the transcription errors (or revolting workplace noise pollution).

I'm fairly well sold on some sort of "omnibar" interface concept where you just tell the computer what you want with a keyboard. Alfred, Spotlight, Google search, Wolfram Alpha, iCal's smart add (you can just type "2pm next Friday"), the Action palette in IntelliJ, VS Code and Sublime's command palette, and so on. That, just... everywhere. And if you're alone, sure, use voice instead. Just don't deprecate the keyboard: still the most reliable way to get accurate text into a computer.