The reason VocalIQ's success rate was "over 90%" is probably because they were constrained to a specific domain (car control). This is very common approach among NLP community. Of course you will perform better when your algorithm is constrained to a specific domain.

All the previous speech recognition platforms like Siri, Cortana, etc. perform worse since they are for general purposes. On the other hand, the reason Amazon Echo works so well is because it has very limited purpose.

We all know just because AlphaGo beat Lee Sedol doesn't mean AlphaGo is smart enough to take over the world.

Spinning this "90%" vs. "20%" just sounds like a bullshit marketing.