[asterisk-speech-rec] In-Grammar and Out-of-Grammar with the LumenVox Speech Engine CRM:00170387

Stephen Keller StephenKeller at LumenVox.com
Tue Dec 19 14:33:12 MST 2006

Recently an issue was brought to our attention that I wanted to share
with the list, in case it's something that has stumped other people.
Basically it has to do with the way the Speech Engine returns results.

The way the Speech Engine works is such that it only ever returns
results that are in-grammar. Even if you said "Maybe" into a yes/no
grammar, the Engine would try and match it with one of your grammar
words (it would just come back with a very low confidence score).

It's worth noting that our Engine is comparing the sounds it hears from
the audio only to the phonemes that are parts of the words in the loaded
grammar, so it is not capable of coming up with out-of-grammar results.
Think of the Engine as a search engine; the space it searches is defined
by the active grammars.

If you're developing speech applications, this is important because the
Engine will always return results to you that are in-grammar, so you
have to reject results with low confidence scores. A good threshold for
rejection is typically in the 450-600 confidence range. The range will
vary depending on the complexity of your grammar and the audio quality.
You probably want to confirm any scores in the 601-750 range, and you
can probably safely accept results that have higher scores without
confirming what the caller said.

This is also important when using our Speech Tuner, freely available to
anyone who has a license for our Speech Engine (e-mail
support at lumenvox.com if you'd like download information). One of the
main components of the Speech Tuner is the Grammar Tester, which allows
you to modify grammars and run tests on audio you've already collected
from your speech application to see how those changes to your grammars
would affect your results.

There is a component of the Tuner called the Transcriber which is used
to transcribe audio. If you do not transcribe audio before you run it
through the Tester, it reports that none of the utterances were
in-grammar. This is because the Tester looks the transcript, not  the
recognition result (since by definition the ASR result is always
in-grammar). If there's no transcript, the Tester compares the null
transcript to the grammar, doesn't match the null, and displays the
utterance as out-of-grammar.

We recognize that this is confusing and we are planning to fix it in
future releases of the Speech Tuner.

I hope this clears up any confusion on the behalf of anyone who has been
using our Engine and Tuner.

Stephen Keller
LumenVox Technical Support
P: 877-977-0707, just say "Stephen Keller" or "Support"
F: 858-707-7072
support at LumenVox.com


Tell us what is important to you! Take our product development survey:

More information about the asterisk-speech-rec mailing list