I'm mostly documenting this for myself. I've been trying to switch away from eloquence for ages, with little success. One possibility is Espeak-ng, at least with NVDA on Windows, and on Android. Unfortunately, Espeak on IOS has a bunch of serious bugs, and hasn't been updated in forever.

One problem with Espeak, though, is that I loathe the way it reads years. It doesn't do any date processing, and nor should it; I don't want to use a January second cup of flour. But I also don't want 1987 read out as "nineteen hundred eighty seven". In order to fix this, add a dictionary entry in NVDA as follows:

• pattern: \b([1-9][0-9])([1-9]{2})\b
• replacement: \1 \2
• case sensitive: checked
• type: regular expression

This will match on any four digit number that doesn't start with a zero, and where digits three and four are not zeros, and add a space between digits two and three. So "19 87" but not "20 01" or "19 00".

In other news, the pattern I wrote for this years ago was apparently back when I didn't know about \b being the word boundary. It did horrible things like ending with (?:(?!\d)+) to make sure it only modified four digit numbers. I found it while doing something else with NVDA dictionaries, and cleaned it up in horror.