Sorry, and yes you're right. So the question is: how to present a product that will always have some kind of limitation? It will solve a problem but under certain circumstances. But I think the answers here still apply.
Many insightful comments here in this thread, thank you all very much. I'm really afraid we can't solve a "mediocre diction" problem. The problem is that speech to text still is a open subject where noise and diction could ruin many tools, but in the ideal world (no noise and good diction), accuracy is around 80-100%.