(NOV 2007) — As we've covered in previous tech bulletins, the LumenVox Speech Engine has a large dictionary of words for every language it supports. It uses this dictionary to know how words in a grammar are pronounced, and many of these dictionary words are names.
But even with close to 100,000 names in the English dictionary, there is always the chance that your speech application needs to use a name that is not in our dictionary.
In that case, the Speech Engine turns to its phonetic speller. The phonetic speller takes a word loaded in the grammar and tries to determine how it should be pronounced.
Improvements in the latest version of the phonetic speller have shown significant gains — our accuracy rate went up by about 10 percentage points — on our tests that include large numbers of names.
Previously, the LumenVox phonetic spellers were built by programming in the phonetic spelling rules for a given language. With a language like Spanish, this is relatively easy, as Spanish spelling follows standard phonetic rules.
In languages like English, this is not the case, as many words do not follow the standard pronunciation rules. Even in very phonetic languages like Spanish, foreign words (especially names) do not follow the normal spelling rules.
To accommodate this, we have changed the way the phonetic speller works. Rather than using hard–coded phonetic rules, LumenVox is switching to use statistical models to generate phonetic spellings.
We train the new phonetic spellers on a set of written words and their pronunciations, in a process similar to how we build acoustic models. The phonetic speller is able to learn the statistical correlation between letter combinations and sounds.
Because our data sets for a language include names and foreign words, the phonetic speller learns the relationship between spelling and pronunciation for these sorts of words.
This allows the speller to figure out how words should be pronounced without any developers having to hard code phonetic rules, making it more accurate for domains such as names from diverse backgrounds. We see the most improvement in these sorts of domains where there are likely to be words not common the language of the specified acoustic model, as this is where the old phonetic rules were the least effective.
Our guide to phonetic spellings has more information on this topic.
The October release of the LumenVox Speech Engine includes the new pronunciation generator, allowing you to achieve better accuracy for many applications.
Users with current software maintenance can always download the latest release from the LumenVox customer portal at http://www.lumenvox.com/customers/ by clicking "Your Software Downloads."
For information about renewing software maintenance agreement, please contact us.
cache–check–interval: This parameter controls how often the Engine should check to see if any cached grammars have reached the time limit specified by time–since–lastaccess and thus need to be deleted. Times are specified in HH:MM:SS. By default it is set to 00:00:50 (50 seconds).
LumenVox now has builds of the Speech Engine, License Server, and MRCP Server available for Debian 4 (Debian Etch). These are available for immediate download. Instructions are available for using apt–get to download and install on Debian. More >
With the October release, we are also introducing support for 64–bit operating systems. We will begin creating 64–bit builds for our supported Linux distributions.
As part of our upcoming integration with the Holly Voice Portal, the LumenVox Speech Engine (via our MRCPv1 Server) passed the rigorous VoiceXML Conformance Test with a 100 percent success rate.
The test suite is designed by the VoiceXML Forum and was administered by Holly Connects on their platform.
This test helps confirm LumenVox's commitment to supporting standards such as VXML and MRCP, and enables to us to add Holly to our list of voice platforms that support the LumenVox Speech Engine.
Stacy Ford will give a 30–minute talk about using speech recognition to differentiate an Asterisk system, Oct. 30 at noon and again Oct. 31 at 2:30 p.m.
Stephen Keller will discuss localization issues when using speech recognition as part of the localization panel on Oct. 31 from 3:30 to 4:40 p.m.
© 2016 LumenVox, LLC. All rights reserved.