(protocol) Add/remove grammars and continous recognition

In the recent drafts of the protocol spec, we've moved to a model where
grammars can be added and removed while recognition is ongoing.  Related
to this is the ability to have a continuous stream of
RECOGNITION-COMPLETE events associated with a single request.

 

This is a significant diversion from the MRCP spec from which we were
originally "inspired".  Even with the proposed "continuous" mode, we
were still very much based on discrete dialog states.  In other words,
all grammars were known at the start of a recognition cycle.

 

The new spec mentions dictation with active hotwords as a scenario that
would benefit from the more flexible model.  This seems like a good
example, but I'm curious to know if there are others.  This is obviously
a hard sell for companies based on traditional MRCP.

 

Robert, I'm also curious to know how Microsoft, in particular, is
planning to implement this functionality.  Are you performing this magic
at a low level in the recognizer, or are these add/remove grammar
commands just macros for restarting recognition.  I'm not so much trying
to get at your trade secretes here (although source code would be nice
;) ), but trying to understand how such a thing might eventually impact
a conformance test.

 

Thanks