| |
More
Questions on VoiceXML
(Continued
from Part 1)
Q:
When doing recognition in a VoiceXML app, how can I
access the recognizer's n-best list?
A:
VoiceXML 2.0 provides support for examining what's called
the "n-best" list. Voice recognition engines
work by comparing what the caller said to the list of
active grammars, and deciding how confident it is that
the caller said a particular matching value. The recognizer
actually computes this confidence interval for several
of the most likely possibilities, and then selects the
one with the highest confidence value as a match (if
it's more confident than the current setting of the
confidence property).
For
richer programmatic control, you can examine the read-only
application.lastresult$ array--- it is always available
at application scope. Each element of the array contains
information about one of the highest-confidence possible
matches for the last attempted recognition. If no recognition
has happened yet in the application or when inside of
an application root document, its value is ECMAScript
undefined.
Each
element of the array contains a set of properties and
values, which are as follows:
|
Variable
|
Description
|
| application.lastresult$ |
Application-scope
array of elements containing information about
the last recognition to occur in the current application.
Array contains one element for as all possible
interpretations with confidence levels greater
than the current setting of the confidence property.
If
no recognition has happened yet in the application
or when inside of an application root document,
its value is ECMAScript undefined.
Each
element of the array application.lastresult$[i]
is an ECMAScript object with the set of properties
described below. After each recognition attempt,
the array is reinitialized and contains (1 <=
i <= maxnbest) elements, where
i is the number of elements and maxnbest
is the current value of the maxnbest property.
The elements are sorted in confidence order from
greatest to least.
application.lastresult$.confidence
et al. can also be referenced directly as shorthand
for application.lastresult$[0].confidence,
etc.
|
application.lastresult$[i].confidence
|
Float
specifying the recognizer's confidence that the
caller actually said this particular match, expressed
on a scale from 0.0 (minimum) to 1.0 (maximum). |
| application.lastresult$[i].utterance |
String
of words actually said by the caller. The exact
spelling is platform-specific (e.g. "five hundred
fifty" vs. "5 hundred 30" vs. "530"). |
| application.lastresult$[i].inputmode |
String
either "dtmf" or "voice", indicating
whether the grammar was matched via touch-tone or
spoken input. |
| application.lastresult$[i].interpretation |
ECMAScript
representation of the recognizer's interpretation
of the utterance as matched by the relevant grammar,
including slot information.
At
this point, the exact method for representing
interpretation results in VoiceXML remains a matter
of annotated discussion within the various specifications
under the auspices of the W3C Voice Browser Working
Group. For detailed information on how to use
interpretation results within VoiceXML, please
see http://www.w3.org/voice
and your vendor's system documentation.
|
Note:
This content is excerpted in part from VoiceXML:
Strategies and Techniques for Effective Voice Application
Development with VoiceXML 2.0 by Jeff Kunins and
Chetan Sharma (Wiley & Sons).
back
to the top
Copyright
© 2001-2002 VoiceXML Forum. All rights reserved.
The VoiceXML Forum is a program of the
IEEE
Industry Standards and Technology Organization (IEEE-ISTO).
|
