AT&T uses software trickery to get voice recognition on the iPhone

AT&T has been working on a research project based on the WATSON speech recognition engine. The project, named Speech Mashups, is a web service that is only activated when the iPhone (or any other modern phone) connects to it. So that means that the actual voice recognition is happening on a remote server and not the phone itself; there is no need to install a plug-in or any other software on the phone.

So far the only one working prototype is a mobile version of the YellowPages website. There is a video on Apple Insider showing a demonstration of it working.