We had to do this a few years ago for a major bank. Automated tools like a11y will only get you a certain amount of the way. Your tests will say 100% but your app may still be completely useless to blind people especially if you’re using an SPA as all screen readers perform differently.

Having a partially sighted consultant come in and point out the deficiencies and how they use devices was extremely helpful. A lot of sight impaired people use Apple devices as their voiceover support is excellent, Android is getting better but you will find a lot of differences between their implementations in versions which is extremely frustrating. For desktop on windows JAWS seems to be the most used which again has its own way of parsing.

I don’t know if it’s changed but my advice would be plenty of manual testing, there’s really not much else than actually using the devices with voiceover enabled to test if things actually work properly. And it’s also a case of compromise and picking what devices to support as fixing something on one reader will inevitably break another.

Hope that helps :)