Powered by Universal Speech Solutions LLC

Menu
Announcements

Kaldi SR Plugin 1.3.0 Released


Kaldi Speech Recognition (SR) Plugin 1.3.0 to the UniMRCP Server (UMS) has been released.

The plugin is based on the following components:

  • UniMRCP Server 1.6.0
  • Kaldi GStreamer Server
  • Libevent 2.1.8
  • Rapidjson 1.1.0

The binaries are currently available for the following Linux distributions:

  • Red Hat / CentOS 7 (unimrcp-kaldi-sr-1.6.1-1.el7.x86_64.rpm)
  • Ubuntu 16.04 LTS (unimrcp-kaldi-sr_1.6.1-xenial_amd64.deb)

This release brings numerous enhancements and fixes to the Kaldi SR plugin.

The detailed list of changes introduced in this release follows.
New Features
  • Speech contexts now can optionally be used as a basic grammar.
  • Allow to specify the format of the confidence score returned in the NLSML results based on a new configuration parameter.
  • Allow to control whether or not to include alternatives with a confidence score below the specified threshold in the returned NLSML results.
  • Keep track of the maximum number of channels used concurrently. This number is logged by default within the statement 'KaldiSR Usage' and can also be written in the status file, if enabled.
  • Added support for 'xml:lang' attribute optionally specified in SRGS XML grammars. This method has the highest precedence in selection of recognition language. The language can be otherwise specified either by the global configuration parameter or the header field 'Speech-Language'. See Section 4.2 in the Usage Guide.
  • Added support for an optional language parameter passed to a built-in grammar. Fixed interoperability with Aspect Prophecy. See the Section 4.2 in the Usage Guide.
  • Added support for predefined metadata in SRGS XML grammars that allows to reference a built-in grammar and/or specify various input parameters. See Sections 4.2, 5.1, 5.2, 5.3, 5.4 in the Usage Guide.
  • A speech context is now can be dynamically specified based on the 'one-of' construct in SRGS XML grammar. See Section 5.2 in the Usage Guide.
  • A no-match event is triggered if transcription result does not literally match any phrase specified in SRGS XML.
  • Set an alarm in the status file if the license server is not reachable for a certain period of time, but the service is not yet affected. Clear the alarm as soon as the license server becomes available. See Section 6.2 in the Usage Guide.
Fixed Problems
  • Fixed an issue in the SDI detector, encountered when 'speech-incomplete-timeout' is significantly longer than 'speech-complete-timeout'.
  • Fixed processing of the header field 'Recognition-Timeout' to set the SDI 'speech-input-timeout' accordingly. Only the global configuration parameter was in effect before.
  • In case of DTMF input, allow both 'maxlength' and 'term char' be specified at the same time. Before, 'malxength' was not observed, if 'term char' was specified.
  • Fixed confidence level set in an NLSML result for DTMF to conform to the format specified in configuration.
  • Fixed processing of malformed parameters passed to a built-in grammar.
  • If an annual node-bound license is used, the expiration time of the license could be provisioned incorrectly, requiring a restart of the service in order to continue normal operation.
Configuration Parameters
  • Added a new attribute 'confidence-format' to the element 'streaming-recognition', which defaults to 'auto' and also accepts 'mrcpv2' and 'mrcpv1'.
  • Added a new boolean attribute 'alternatives-below-threshold' to the element 'streaming-recognition', which defaults to 'false'.
Miscellaneous
  • Updated the Usage Guide to reflect the changes introduced in this release.

Visit the Kaldi SR plugin page for more information.

http://www.unimrcp.org/kaldisr

Thank you for using UniMRCP.

--
Arsen Chaloyan
Author of UniMRCP
http://www.unimrcp.org

Announcement Group

Subscribe to the Announcements Group if you prefer to receive news and press releases only. Please note that announcements are normally posted to the Discussion Group as well.