Voice recognition is now possible using the Call Flow Designer (CFD). However, building a CFD application for each IVR would probably be ... burdensome ... to build and modify (when needed). The CFD will submit the request (audio) to Google and Google will respond with it's best guess of what the caller wanted (true of any voice recognition system).
I've pondered whether we could make a generic CFD that takes configuration information and "builds" the needed IVR, but it sounds like a fair amount of work with no real sense of demand for such a tool. Building an alternative company directory, however, seems like a logical first effort in that direction.