[justapassenger]

I'm very well aware that low accuracy voice transcription is possible.

But it's naive thinking that having an algorithm equals solving a technical problem. That's not even the problem. Problem is how to deploy it, at scale, without anyone leaking it (both employees and vendors). And hiding it so well, that none of the security researchers will be able to find it. And doing all of that in a way, that they can use it and get value out of it.

And then compare risks of doing with risks and ROI of, for example, improving search accuracy, so people will just come and tell you more about stuff they want.