(JUNE 2009) — LumenVox 9.0, with improved accuracy, new noise reduction, better logging, a new media server application, and a streamlined API is now available.
We also have a 9.0 announcement at LumenVox.com that outlines some changes, and some improvements such as noise reduction were documented in last month's Tech Bulletin.
This is a major release of the LumenVox software, and it paves the way for some exciting changes in future releases. In this Tech Bulletin, we want to concentrate on all of the good things that are available now.
LumenVox 9.0 is a free download to any customers with current software maintenance. If your maintenance has expired, you can now log into your LumenVox.com account and upgrade your maintenance online. You can also feel free to call us at +1–858–707–7700 and ask for "Sales" to talk to a representative about renewing your maintenance or Contact Us.
And if you have any legacy applications running with a different speech recognition provider, is the perfect time to switch them to LumenVox.
Undoubtedly the biggest change in 9.0 is the addition of our continuous models, which allow for higher quality recognition. By default, a new installation of the 9.0 Speech Engine is set to use the continuous models, so when upgrading you should automatically see the improved quality. These models also use about 1/3 of the memory than previous models, so you should see a significant reduction in the memory used by the Engine.
One important note is that, so far, American English and Australian English are the only languages with continuous models built for them. We are working on adding continuous models for the other languages, but for now you will have to revert to semi–continuous mode if you use another language.
To do this, simply edit your sre_server.conf file and change HMM_TYPE from CONT to SEMI and restart the Engine.
The special grammar rule called the $GARBAGE rule is used for dealing with out of grammar responses. If a speaker says something otherwise not in the grammar, the Engine attempts to match the $GARBAGE rule, making it easier to know the caller said something not contained in the grammar.
Speech Engine 9.0 features improved handling of the $GARBAGE rule, meaning that developers who use it in their grammars will find their success rate should increase. We've expanded our help documentation on the special grammar rules so you can get started using this feature if you're not already.
The logs generated for use with the Speech Tuner — called "callSRE files" due to their .callsre file extensions — are some of the most useful logs that LumenVox generates. These contain the audio from callers that is used to help tune the application.
Previously, the Engine could only save audio after the voice activity detection had detected barge–in. This means that while the Engine could usually record what a caller said, it would be unable to create a .callsre file with the audio if barge–in was never triggered.
This made it somewhat difficult to troubleshoot problems relating to barge–in. If barge–in was not being triggered, there was no audio to review and attempt to diagnose the issue. In LumenVox 9.0, the Engine can be set to record all audio streaming into it, even before barge–in is triggered.
Doing so makes it easier to understand what's happening when barge-in or end of speech are not being properly detected. To enable this extra logging, edit the client_property.conf file in your LumenVox configuration directory and change SAVE_SOUND_FILES to equal ADVANCED. The old method can be enabled by changing the parameter to BASIC, and there is a new debugging setting called ALL that can be used to provide LumenVox technical support with even more information, though you should probably not enable this setting unless directed to by LumenVox.
One interesting new addition in 9.0 is the introduction of our new LumenVox Media Server application. This is a replacement for our older MRCPv2 Server, and our intention is that it will ultimately replace our MRCPv1 Server as well.
The LumenVoxMedia Server uses less memory, is faster, more stable, and generally superior to the previous MRCPv2 Server. We have performed interoperability testing with a number of our partners to ensure it is compliant with the MRCPv2 specification and a variety of popular voice platforms.
As we conduct more testing with the LumenVox Media Server against platforms that use MRCPv1 to communicate, we will eventually phase out the MRCPv1 Server in favor of a single, unified server that can handle SIP and RTSP traffic to support both versions of MRCP.
One change that will help LumenVox developers is the streamlining of our C API. We have moved a number of older functions into a new deprecated header file called LV_SRE_Deprecated.h. We have also moved some of the more advanced functions into a new header file called LV_SRE_Advanced.h.
Both of these files are linked from the normal LV_SRE.h header, so existing applications should just need to be recompiled without code changes. New developers, however, will find a cleaner and simpler SpeechPort interface.
If you are a developer of existing code using the LumenVox API, you may wish to review the functions in the deprecated header. These are functions we may remove in future releases, so your feedback about them is appreciated, especially if there are functions you use and would like retained.
It is likely we will keep most of them in that deprecated file, and in almost all cases, those functions represent an older method of doing something that has a cleaner representation in the new LV_SRE.h.
© 2016 LumenVox, LLC. All rights reserved.