Provides a secondary guide to the Viterbi beam search.
|
Value |
Float: -100–100. As many as three values are allowed. Typically set to the same value as swirec_silence_prune_offset. DEFAULT: -30, -60, -60 |
|
How to set |
Shown in precedence order (see Rules of parameter precedence): Set in the browser, using the Vendor-Specific-Parameters header in a RECOGNIZE method. In a parameter grammar. In a grammar using a <meta> tag (see the text on recognition-time parameters in Parameters set in grammar files). In the application's session xml in the <speechrecog> section of <sessionparameters>. If using the Management Station, set on the Nuance recognition service. If not using Management Station, set in a Recognizer configuration file. |
|
Usage |
This is an advanced feature for developers who are familiar with Recognizer and its underlying models. It tunes Recognizer’s search for matches: Nuance might remove this parameter from the product or change its behavior in the future. |
|
Performance |
Changing the default value can seriously impair accuracy and confidence. After changing the value, you must thoroughly test recognition accuracy before applying the change to a production system. |
When set via VoiceXML or MRCP, the setting is valid for the next recognition event for all active grammars. When set via a grammar, the setting is valid until the grammar is de-activated. After the event or deactivation, the settings return to the values set in a parameter grammar, in the individual grammars, or in a configuration file.
This parameter specifies as many as three possible values. The value used depends on the setting of swirec_load_adjusted_speedvsaccuracy and on the system load at recognition time. When the speed versus accuracy parameter is turned on, Recognizer automatically refines and improves recognition results by using extra CPU cycles if they are available. By default, Recognizer uses the first value defined and then it uses the second or third value as CPU load increases. (The CPU thresholds are labeled as idle, normal, and busy.) If fewer than three values are specified, the value in the final position is assumed to be repeated.
Conceivably, you can achieve better accuracy for the amount of CPU processing using a tighter beam (lower value) for this parameter. For example, if swirec_state_beam is -30, then swirec_phoneme_lookahead_beam might be set to -40 or -50.
Browsers usually set this parameter in a RECOGNIZE request, and the parameter automatically reverts to the previous value after the recognition event. (Alternatively, browsers can use the SET-PARAMS method. See Configuring Recognizer with the browser.)
Vendor-Specific-Parameters: swirec_phoneme_lookahead_beam=-30 -60 -60
Sample definition in a grammar file:
<meta name="swirec_phoneme_lookahead_beam" content="-30 -60 -60"/>
Sample definition in the <speechrecog> section of a session.xml file:
<param name="swirec_phoneme_lookahead_beam">
<value>-30</value>
<value>-60</value>
<value>-60</value>
</param>
Related parameters
