Fixed problems with custom pronunciations which caused custom pronunciations to be silently ignored
in a number of cases.
Fixed problem with new grammar caching mechanism where the original GrXML grammar text was not being
correctly stored in the resulting callsre files when using the streaming interface or Media Server.
Fixed problem with grammar parser resolving uris that specified relative paths using the ../
syntax.
Fixed problems with Speech Tuner loading malformed grammars into grammar editor.
Fixed cosmetic problem in Speech Tuner when displaying decode progress with greater than 65535
interactions due to minor overflow issue.
Fixed bug that caused grammar results to sometimes be returned in incorrect case.
Added missing acoustic models to Speech Tuner installation that are needed for phonetic
speller tool.
9.2.300 (February 16, 2010):
Fixed bug in Media Server where an empty result is returned with 0 confidence score when
a bad VAD silence is detected due to insufficient leading silence in the decoded audio.
This was changed to now return no-match if recognized string length is less than 1 and confidence
is set to 0.
Fixed bug in the client where a grammar label of greater than 260 characters could
create a buffer overrun situation and cause an exception.
9.2.200 (February 9, 2010):
A brand new Speech Tuner has been released! Full details are available in the
Tuner documentation
Significant internal changes to allow better handling of grammars. This specifically addresses
large grammar issues but also allows faster processing of all grammars. If you previously had
trouble loading large grammars, please try upgrading to 9.2.200 or later.
Changed default location of the Media Server configuration file in Windows (it will now default
to a folder called config in the Engine installation directory). In the event that your %LVBIN% variable
is set to somewhere other than the Engine installation directory, the Media Server will look in %LVBIN%\config.
Streamlined decode processing code allowing decodes to run faster generally.
Added more verbose logging to grammar loading and parsing.
Better compatibility with alaw audio format in Media Server. Is now faithful to SDP
specified format, not RTP packet markings. Previously, incorrectly marked RTP packets
were not processed correctly if there was an incorrect mismatch between SDP specifier
and RTP specifier. Also improved possible ambiguity when client requests both alaw and
ulaw audio formats at the same time (now first specified will be used).
Fixed a bug in the internal statistical pronunciation model relating to pronunciation of
words (typically nouns) that would not generally be in the language dictionary. This fix
improves accuracy in these cases.
Fixed a problem with possible non-unique Session-ID strings in MRCPv1 sessions.
The Media Server's RTSP interface will now correctly respond with 486 Busy Here when speech port is
unavailable at SETUP. Previously the SETUP would succeed, but subsequent DEFINE-GRAMMAR
or other calls would fail.
Improved the load balancing between client and multiple SRE servers using a new algorithm
designed to better utilize and balance SRE resoruces.
Client side grammar caching has been added. This reduces the amount of time needed for
subsequent grammar load requests from each client. This can significantly speed up loading
GrXML grammars, which need to be converted to ABNF prior to sending to server. Several new
settings have been added to client_property.conf
including: CLIENT_CACHE_ENABLE, CLIENT_CACHE_EXPIRATION, CLIENT_CACHE_MAX_NUMBER and
and CLIENT_CACHE_MAX_MEMORY.
The Media Server was getting into a deadlock state under extreme specific conditions on Linux
platforms. This problem was identified and fixed.
Changed configuration file parser to better handle situations where config file is
missing.
Fixed a problem in the way threads are stopped in Linux. This allows more robust thread
control.
Fixed problem in Media Server in Linux, where locking would occur sometimes due to
blocking socket call.
Changes to Linux shutdown code improve shutdown performance and reduce potential
problems.
More robust handling of configuration values in Media Server to prevent minor rounding
problems on some Linux distributions.
Media Server changed to better handle multiple duplicate SETUP requests in the same
session.
The Engine installation package now includes a new Call Indexer service. This works in conjunction with New
Speech Tuner to index and serve callsre files as needed on Linux and Windows machines
connected to Speech Tuner.
Logging of License Authentication requests was improved to indicate licenses on a per-user basis.
Fixed a grammar reference bug where a non-root-rule was being referred to in a
referenced grammar.
Fixed bug in server side grammar cache, when expiring grammars.
Added fix for NAT problems with certain routers and firewalls that would automatically
close license connections after a period of inactivity. Now the socket will be pinged to
remain open more frequently.
Fixed small memory leak when socket message connectivity is lost in certain conditions.
Fixed a bug where callsre responses were being stored in the incorrect folder under
certain conditions.
License Administrator changes were made to prevent freezing of GUI when connecting to License
Server that is under stress.
Fixed bug with multiple sequential $NULL rules being parsed in a grammar.
Fixed problem with GrXML to ABNF conversion that would not correctly assign grammar
weighting to $NULL rules in the generated ABNF output. Removed unnecessary double $NULL
rules.
Added option to automatically clear cache folders one time after installing or upgrading
software to avoid possible version compatibility problems. Upon installation or upgrading, a file
called cached_grammars.key will be generated in the server-side grammar cache folder, and a file
called cached_client_grammars.key will be generated in the client-side grammar cache folder.
If this file is detected by the Engine or client, it will clear the appropriate cache (including
the file). This means that after upgrading, large grammars may take a while to load the first
time as the cache is clear.
Removed medium and high resolution acoustic models from packaging to reduce shipping size
(these models are still available by custom request).
Removed some unwanted/unnecessary logging.
Windows 2000 is no longer a supported operating system; however we have tested on Vista and Windows
7 and now support those operating systems.
Support for Fedora Core 9 has been dropped, but support added for FC12.
Fixed bugs when handling custom pronunciations in (default) continuous decode mode.
9.1 (October 2009):
The LumenVox products now support licensing authentication, allowing for
the License Server to only provide licenses to clients who authenticate with a user name and
password. You can use this if you wish to provide licenses to customers over the public Internet,
or to take advantage of our new
subscription licenses.
American English (en-US) now has three different
resolution acoustic models you can use. These models
offer better accuracy, but use more memory and require more time for decodes. By default, only the lowest
resolution model is loaded.
We have removed the "-di" suffix in a language declaration to indicate that digits-only
acoustic models should be used. The Engine will now make this switch automatically, if it detects that
that all loaded grammars use only digit words. This should offer much greater compatibility when using
the built-in digits grammars, especially with VXML applications. If you continue to specify -di in a language,
it will be ignored.
There have been a number of bug fixes and enhancements that should improve compatibility between the LumenVox
Media Server and various voice platforms that use MRCP.
We have streamlined the Speech Engine such that its memory footprint has been reduced substantially. Under
light loads, you should find that the Engine uses only about 150 MB of memory.
Be sure to note that if you upgrade to 9.1, you should
download and install new license files following
our new upgrade procedures that were put into place in 9.0.
Fixed a bug that caused loads to be improperly distributed when using multiple speech servers.
This is the first release of the LumenVox
Continuous HMM Decoder for use with the Speech Engine.
The Continuous model does not use compression, so it has higher resolution, resulting in increased accuracy.
The continuous models have shown an accuracy increase across various domains, but at the expense of
approximately 15-20% more processing time.
There are new noise reduction options available in 9.0, allowing better
recognition in noisy environments.
The LumenVox Speech Engine version 9.0 offers improved support of the
$GARBAGE rule, which allows grammars to be
defined where utterances before and after the desired phrase can be ignored. This is programmatically
challenging to do correctly, and this new version sets out to improve performance over previous
mplementations of this rule.
The new LumenVox Media Server provides
an interface to our Speech Engine via MRCPv1 (RTSP) and MRCPv2 (SIP) connectivity. These are two commonly used
networking protocols used among a variety of leading speech engines.
The LumenVox License Server will no longer allow old licenses to work with the newer versions of the
Speech Engine. This means that to use the latest versions of the software, you must ensure your software
maintenance is up to date and download and install licenses with the newer maintenance date. For more information,
see Upgrading LumenVox Software.
8.6 (January 2009):
New configuration files have been added to the Engine. This should
allow greater control over Engine settings without having to use the API. A few older configuration files, such as the old
license_client.conf, have been merged into these new files. If you still have the old configuration files in place,
the Engine will prefer the values from those, so it should not break backwards compatibility for any users.
We have revamped all of our example application code. It should be better
commented and more cleanly written. We have also added a few new sample applications.
8.5.100 (May 2008):
Licenses can now be uninstalled, using the latest License
Server. If you have a machine with licenses that were set up before the release of 8.5, you will need to upgrade those licenses.
Users who do not need to uninstall a license should be unaffected by this change.
The location of files on Linux has been restructured. This represents a major revamp of the LumenVox software on
Linux. Existing Linux users will need to take this into account when they upgrade. Note that we have also completely dropped
the environment variables ($LVBIN, $LVLIB, etc.) on our Linux installations, and we have split the Engine into three
separate Linux packages: an Engine client package, an Engine server package, and a "core" package containing files
shared across all products. Please see Linux Directory Structure for information
about where our files are now installed.
Short words (e.g. "back" or "stop") should now have higher confidence scores when correctly recognized.
8.0.300 (December 2007):
Added new utilities related to phonetic
spellings and our internal dictionary of words. You can now use
LV_SRE_CheckWordsInDictionary
(C API) / CheckWordsInDictionary
(C++) to determine whether a word or string of words is in the dictionary of a given language.
LV_SRE_GetPhoneticPronunciationCount
(C API) / GetPhoneticPronunciationCount (C++) returns the number of pronunciations the Engine has for a string of words, and
LV_SRE_GetPhoneticPronunciation
(C API) / GetPhoneticPronunciation (C++ API) returns the actual phonemes for a string of words.
Improved the Engine's performance in a number of ways, particularly relating to memory use on Linux. Fixed some potential
memory leaks.