Browse
 
Tools
Rss Categories

Vendor-Specific Recognizer Properties (MRCPV1)

Reference Number: AA-01108 Views: 7413 0 Rating/ Voters

Vendor-Specific Recognizer Properties

The following parameters are LumenVox-specific extensions to the MRCPv1 specification. They can be controlled via the mrcp.config file, located in the config directory of the MRCPv1 installation folder. By default, this location is C:\Program Files\Lumenvox\MRCPServer\config\

They may also be set with the appropriate header as part of a RECOGNITION or SET-PARAMS method; see Specifying Vendor-Specific Properties via MRCP Headers below.

See Configuring Parameters for more information about changing various MRCPv1 parameters.

choose_model

If this property is set to 1, then the client will decide which acoustic model is most appropriate for the server to use, based on a frequency analysis of the speaker's voice. Otherwise, two decodes will be done simultaneously, and an answer will be selected based on which model had better "coverage" for the speaker's voice.

Default: 0

enable_lattice_scoring

The lattice-based confidence score is a slightly slower, but more accurate confidence score.

Default: 1

end_of_speech_delay

Deprecated in favor of speech_complete_timeout, the standard MRCP parameter that works in the same way. See Streaming Media Properties.

Default: 500

wind_back_time

The length of audio wound back at the beginning of voice. It helps in the situation of weak speech onset. The resolution of this parameter is 1/8 sec, i.e. 125 ms, which means setting this value to 249 ms is the same as setting it to 125. It is specified in milliseconds.

Default: 250

initial_silence_trimmed

This parameter tells the Engine's voice activity detection technology whether the audio stream contains leading silence or not. By default, the Engine expects to receive an audio stream that contains silence before speech. If your hardware is trimming the silence before sending the Engine audio, you should set this to 1.

Default: 0

burst_thrsld

Barge-in will trigger only after it has detected speech for the period specified by this parameter. This helps prevent burst noise from triggering barge-in. It is specified in milliseconds.

Default: 100

barge_in_timeout

If no speech was detected in the time specified by this property, the recognition request is cancelled. It is specified in milliseconds.

Default: 10000

end_of_speech_timeout

After barge-in, if end-of-speech is not detected in the time specified by this property, the recognition request is cancelled. It is specified in milliseconds.

Default: 10000

Specifying Vendor-Specific Properties via MRCP Headers

As mentioned previously, you may specify the above parameters in an MRCP header. You must use the following format. Note that a semicolon (";" ) is used as the delimiter:

Vendor-Specific: com.lumenvox.barge-in-lvl=30;com.lumenvox.noise-floor=200\r\n

This header field may be specified in RECOGNIZE or SET-PARAMS method during an MRCP session. The following header field names may be used:

com.lumenvox.choose-model
com.lumenvox.lattice-score
com.lumenvox.end-of-speech-delay
com.lumenvox.noise-floor
com.lumenvox.wind-back-time
com.lumenvox.initial-silence-trimmed
com.lumenvox.barge-in-timeout
com.lumenvox.end-of-speech-timeout