Hi,
I remember that Aaron once said, that they tested one of commercial solutions and it wasn't accurate enough. But there is no drawback of trying to do something useful in this field....
As you probably know, the heart of voice communications under LMCE is Asterisk - and there exists LumenVox speech recognition engine for it.
You just make it answer certain local extension and then when you want to recognize speech just make a call to that local number (you can do it even from MDs, since they have embedded softphones embedded)...
The well known open source solution for speech recognition is Sphinx family of projects, there is also interesting project Simon listens....
http://www.simon-listens.org/HTH,
regards,
Bulek.