First Vista remote exploit?

Yesterday I had the idea to use Vista's speech recognition
system for remote exploiting. By embedding commands
into a soundfile offered by an evil website or into
all these Web 2.0 videos, remote attackers might be able
to execute commands on a Vista system while they
are spoken upon viewing.
The idea has little chance to succeed I thought. I posted
to Daily Dave's mailinglist and
George Ou made some samples and it seemed to work.
Read the full thread here:

http://blogs.zdnet.com/Ou/?p=416

There are some constraints but basically it seems that
it is possible to delete files or to start certain
applications remotely.

I could not verify the results by myself since I lack a license.

Update:
It has now been confirmed by various people that the speech recognition
software seems to miss echo cancellation. While it is "easy" to implement
with applications like skype or messenger I do not know how easy it is
within the OS. Remember that the out-path is via the web-browser
(no idea whether the audio-I/O architecture allows access by the recognition
software) which
does not belong to the speech recognition system. From that point of view,
the OS would need to implement a global echo cancellation, a few layers below
the browser/recognition layer. This would make such systems within messenger
or skype obsolete.

Update:
It has been confirmed that this also works with untrained voices. Not as good
as with trained ones, but it works. Downloading and executing of user-level
binaries which do not trigger UAC has also been confirmed to work.