Sorry, I might be totally off the base here but why not use IBM Watson for that? It's free for a thousand minutes in a month and it's speech-to-text looks quite good on the first glance.
I was just looking around for a custom solution to have something listening, send it off somewhere to get the text out of the voice command and then respond to keywords/contexts with appropriate custom-made actions. All you need then is to process that and read out the response.