[microtherion]

1. It's not like Hidden Markov Models (the approach that dominated recognition prior to the deep learning revolution) is any more explainable than deep learning models.

2. You generally don't gain more confidence in the accuracy of a particular word by looking at less context. This is neither how human nor machine recognition works.

[Closi]

I'm not saying that a particular technical approach is better, I'm just saying that from a product perspective the medical industry has specific requirements which are currently satisfied in Dragon and not in the generalised speech libraries Apple / Google have at the moment.

I've not got direct experience in healthcare, but do have experience in industrial voice, and this is an area where Apple/Google generalised libraries perform significantly worse than specialised software (Dragon is also big in this industry, albeit at the SDK level). In industrial voice the main requirements are high levels of background noise, restricted vocabulary (20-30 words) and people speaking very quickly.