BUILDING SPEECH RECOGNITION APPLICATIONS
<p>

In this tutorial we focus on the impact of grammars on the quality and
success of speech recognition applications.
Grammars used for speech recognition fall apart into 2 large categories:
statistical grammars and finite state
grammars.  The statistical grammars on trained on huge example corpora, as
might be available for dictation applications.
In most command & control system the number of available example phrases is
too small for such approach.
On the other hand, if the application is concise enough, then it may be
possible to construct a suitable semantic grammar
for that particular application.  A danger looming behind the corner, is
that the grammar may not cover all possible
dialogues and hence result in a poor application.  However, the picture is
not as black and white as it looks at first
sight.  A number of 'overspecifying' finite state grammars can be defined in
between 'no grammar' and the cute compact
semantic grammar.  An increased design effort may lead to lower perplexity
grammars, but at the same time have higher
quality requirements in order to be fully functional.

In the morning session we will review the general concepts required.  In the
afternoon session, a practical example
will be worked out.  Different groups will work on different - competitive -
implementations. Finally the different
designs will be informally evaluated against one another for effort,
complexity and completeness.