In our last entry to the tech-tips blog, we detailed the challenges inherent in capturing alphabetical, or alpha-numeric entries from our callers, and detailed several paths for minimizing the chance of mis-recognition when implementing input fields based on these two categories of voice recognition. The long and short of this posting was that IVR developers should refrain from attempting this wherever possible, and to instead try these alternatives:
- Pre-compiled Statistical Language Model grammars
- Leveraging TargusInfo services for advanced recognition accuracy
However, the IVR project requirements dictate what we can, and can’t do as developers, so in some cases, we have to try and whip out a user grammar that takes alpha, or alpha-numeric input. As mentioned in our last blog entry, there are a few things we can do to stack the deck to try and squeeze more accuracy out of these grammars so that we don’t end up with frustrated callers, but the plain truth is that we will never, ever be able to write a grammar that accepts alphabetical characters to be 100% accurate using todays recognition technology. What we will do today is twofold:
(1) Craft an SRGS+SISR subgrammar for alphabetical, and numeric characters
(2) Plug this grammar into a mixed-initiative form dialog that will minimize (but not fully address!), the possibility for mis-recognitions.
Those developers who have the need for such a grammar and dialog within their production-grade applications are advised to take this basic framework as a starting point, and then expand on it by:
(a) Test carefully with a broad range of users, and to fully flesh out alternate utterance values for alphabetic characters
(b) Apply item weighting to specific characters based on the probability of a given character versus another like-sounding character – this will depend greatly on the specific usage of the grammar
(c) Track results by using w3c-compliant utterance recording, and logging all shadow variables, so that these results can be used to further tune and tweak our grammar for maximum accuracy
(d) Consider using n-best post-processing as an additional confirmation step to ensure that the results we receive are indeed accurate
For today’s entry, lets assume that we need to track a three digit zip code, which are prevalent in Canadian locales. Our predefined format for utterance values are “Alpha Digit Alpha”, and luckily, not all alpha characters are applicable: Instead of trying to recognize 26 letters accurately, we only need to recognize 16, which helps a lot!
We won’t dig into the specifics of a mixed-initiative form dialog, as we have already done so in our mixed-initiatve tutorial, but the gist is that this feature of VoiceXML allows us to fill multiple fields with a single utterance, and breaking up each alpha and numeric character into it’s own recognition field greatly cuts down on disambiguation problems that can occur.
For the purposes of brevity, what we have below is a stripped-down version of our fully fleshed-out grammar, but you may download the full grammar, and the mixed-initiative dialog right here, which contains lots more inline notations.
<?xml version= "1.0"?><grammar xmlns="http://www.w3.org/2001/06/grammar" xml:lang="en-US">
<rule id="canadianZip" scope="public">
<one-of>
<!-- ALL THREE FIELDS FILLED -->
<item>
<item>
<ruleref uri="#alphaRule1"/>
<tag>out.alphaSlot1=rules.alphaRule1.alphaSlot1;</tag>
</item>
<item>
<ruleref uri="#numRule"/>
<tag>out.numSlot=rules.numRule.numSlot;</tag>
</item>
<item>
<ruleref uri="#alphaRule2"/>
<tag>out.alphaSlot2=rules.alphaRule2.alphaSlot2;</tag>
</item>
</item><!-- ONLY TWO FIELDS FILLED -->
<item>
<item>
<ruleref uri="#alphaRule1"/>
<tag>out.alphaSlot1=rules.alphaRule1.alphaSlot1;</tag>
</item>
<item>
<ruleref uri="#numRule"/>
<tag>out.numSlot=rules.numRule.numSlot;</tag>
</item>
</item>
<item>
<item>
<ruleref uri="#numRule"/>
<tag>out.numSlot=rules.numRule.numSlot;</tag>
</item>
<item>
<ruleref uri="#alphaRule2"/>
<tag>out.alphaSlot2=rules.alphaRule2.alphaSlot2;</tag>
</item>
</item>
<item>
<item>
<ruleref uri="#alphaRule1"/>
<tag>out.alphaSlot1=rules.alphaRule1.alphaSlot1;</tag>
</item>
<item>
<ruleref uri="#alphaRule2"/>
<tag>out.alphaSlot2=rules.alphaRule2.alphaSlot2;</tag>
</item>
</item>
<!-- ONLY ONE FIELD FILLED -->
<item>
<ruleref uri="#alphaRule1"/>
<tag></tag>
</item>
<item>
<ruleref uri="#numRule"/>
<tag>out.numSlot=rules.numRule.numSlot;</tag>
</item>
</one-of>
</rule>
<rule id="alphaRule1" scope="public">
<one-of>
<item weight="1.0">
<one-of>
<item> ex</item>
<item> ax</item>
<item> x </item>
</one-of>
<tag>out.alphaSlot1="X"; </tag>
</item>
</one-of>
</rule>
<rule id="numRule" scope="public">
<one-of>
<item> one <tag>out.numSlot="1"; </tag> </item>
</one-of>
</rule>
<rule id="alphaRule2" scope="public">
<one-of>
<item weight="1.0">
<one-of>
<item> ay</item>
</one-of>
<tag>out.alphaSlot2="A"; </tag>
</item>
</one-of>
</rule>
</grammar>
In brief, our top-level rule assumes that we can have any of the following entries:
"X1A""X"
"X1"
"XA"
"1"
"1A"
And in the event that we get one or two characters matched in our utterance, the VoiceXML mixed-initiative logic will then take over, and prompt the caller to fill in any “blanks” remaining.
A few things of note about the grammar defined below is that in the event that we receive only a single alpha utterance, we will assume that it is the first character, not the last. Additionally, when we construct a grammar that contains multiple slot returns, it is required that we explicitly define the slot values all the way up the chain: if we didn’t define the “out.[slotname]=rules.[rulename].[subslot]” within the context of the top-level rule, the last slot value would overwrite all others, meaning that we would only get a value for “alphaSlot2″ within the VoiceXML dialog. To illustrate even further, the below snippet for a top-level return would make this a reality:
<item>
<ruleref uri="#alphaRule1"/>
<ruleref uri="#numRule"/>
<ruleref uri="#alphaRule2"/>
</item>
You’ll also see that each possibility for character recognition is specified within the top-level rule, so in the event that we get 1, 2 or 3 character strings, we can pipe the return value back to the VoiceXML, and let the mixed-initiative dialogs then access the sub-rules (alphaRule1/2 and numRule), individually as needed.
We also illustrated in brief how one can define multiple like-sounding utterance values that return the same interpretation value, and defined an for our alphaRule1 entry simply to show how this can be done: The task of taking this framework, and turning it into a grammar that satisfies any given project rests in the hands of you, the capable IVR developer.
=^)
Till next time,
Matthew Henry
Director of Customer Support
Voxeo Corporation
Useful Links
Technorati Tags:
voicexml, voxeo, srgs, sisr, grammars, ivr, applications, tutorials
RSS Feed
