There's a slightly weird, slightly sad video for the TellMe service that Microsoft bought, merged into its own voice recognition offerings and used for the voice recognition, control and search on Windows Phone. It shows a Siri-style service that fast-forwards a couple of gal pals through planning and celebrating a wedding - only with the kind of really long-term interactions you need to accomplish anything more complex than a Web search or putting an appointment in the diary. In the video, TellMe isn't just remembering who you're talking about two minutes later, it's remembering a week later - and when there's new information it's coming back and telling you about it.
TellMe is pretty good at understanding Americans; before Microsoft bought it, it was driving over half of the 411 calls in the US (the equivalent of 118 phone enquiry services, just without the 'creepy moustachioed twins in tracksuits' advertising). In quiet environments - say, not a press conference - it's pretty good at most UK accents as well. And while Siri is smarter, with its canned responses to predictable geek jokes and ability to grasp what you're talking about, the voice recognition on Windows Phone is cleverer when it comes to the simple tasks.
If you want to search Bing Maps for a restaurant or skip the typing for a Web search, your words get sent up to the cloud to be turned into text. But if all you want is to dial a contact or start an app, add a second person to an existing call or (ironically and recursively) speak the number for an automated voice navigation system instead of putting the phone on speaker while you bring up the keypad on screen), it seems that voice recognition is done on the phone.
At Nokia World, Joe Belfiore explained that "the voice recognition starts locally on the phone, to see if you're doing something like launching an app". With a 'closed domain' where there's a limited number of words you could be saying - names, numbers, the names of just the apps that are on your phone - the phone is powerful enough to do the voice recognition. It's only for the wider range of vocabulary needed to understand something like a Web search that it needs to use the recognition service in the cloud.
Siri sends your words up to the cloud for everything (that uses some data but probably not enough to worry about: Ars Technica's measurements show uploads ranging from 18KB to 60KB). It's more of a problem if you don't have connectivity, because then you can't even do the simple voice tasks.
Good as the voice recognition on Windows Phone is, my main problem is I rarely remember to use it. Siri has the whimsy and the media coverage to be more memorable, and it does something intriguing. But between the impressive Nokia Lumia handsets, the flood of new Windows Phone applications that have arrived since the Mango update - something that seems to be speeding up as Nokia applies its marketing and developer evangelism muscles and the homebrew ChevronWP7 Labs phone unlocking service, Windows Phone has some wind in its sails (and perhaps its sales) at last. There are still standout apps that aren't available anywhere but the iPhone, but Windows Phone is finally feeling like a real contender.