Advertisement
Not a member of Pastebin yet?
Sign Up,
it unlocks many cool features!
- <?xml version="1.0"?> <!-- XML version -->
- <?xml-stylesheet type="text/xsl" href="SWIrecConfig.xsl"?>
- <!--
- ****************************License********************************
- Copyright (c) 1995-2006 by Nuance Communications, Inc. All rights reserved.
- Nuance, Inc. provides this document without representation or warranty of
- any kind. Nuance, Inc. reserves the right to revise this document and to
- change the information contained in this document without further notice.
- Dialog Modules, OpenSpeech, Productivity Without Boundaries, ScanSoft, the
- ScanSoft logo, SMARTRecognizer, SpeechCare, Speechify, SpeechSecure,
- SpeechSpot, SpeechSite, SpeechWorks, the SpeechWorks logo, and SpeechWorksHere
- are trademarks or registered trademarks of Nuance Communications, Inc. or its
- licensors in the United States and/or other countries. This document may also
- contain other trademarks, which are the property of their respective owners.
- Without limiting the rights under copyright reserved above, no part of this
- document may be reproduced, stored in or introduced into a retrieval system,
- or transmitted in any form or by any means, including, without limitation,
- electronic, mechanical, photocopying, recording, or otherwise, without the
- prior written permission of Nuance Communications, Inc.
- -->
- <!-- Define the structure for the XML -->
- <!DOCTYPE SWIrecConfig SYSTEM "SWIrecConfig.dtd">
- <!-- *************** Organization of Config Parameters ************* -->
- <!-- * Parameters related to INITIALIZATION * -->
- <!-- * Parameters related to CACHING * -->
- <!-- * Parameters related to ENDPOINTER * -->
- <!-- * Parameters related to LEARN * -->
- <!-- * Parameters related to ACCURACY AND LANGUAGE MODELING * -->
- <!-- * Parameters related to PERFORMANCE * -->
- <!-- * Parameters related to DEBUGGING * -->
- <!-- * NEVER CHANGE WITHOUT Nuance authorization * -->
- <SWIrecConfig version="1.0.0">
- <lang name="default">
- <!-- ***************************************************************** -->
- <!-- * Parameters related to INITIALIZATION * -->
- <!-- ***************************************************************** -->
- <!-- 1:enable speech mode in rec result when NOMATCH; 0:disable this mode -->
- <param name="swirec_result_enable_speech_mode">
- <value>0</value>
- </param>
- <!-- 0:include "*#ABCD" in DTMF builtin grammars; 1:exclude "*#ABCD" -->
- <param name="swirec_builtin_grammar_full_dtmf_mode">
- <value>0</value>
- </param>
- <!-- Controls how long the recognizer should wait after the end of
- the prompt for the caller to speak. -->
- <param name="timeout">
- <value>7000</value>
- </param>
- <!-- Controls how much silence (milliseconds) after a successful recognition
- is needed to conclude that a caller is finished. -->
- <param name="completetimeout">
- <value>0</value>
- </param>
- <!-- Controls the level of confidence to determine when the
- application needs to confirm. Values 0-1000. -->
- <param name="confidencelevel">
- <value>0</value>
- </param>
- <!-- Controls the length (milliseconds) of a period of silence after
- callers have spoken to conclude that they finished. This timer is
- also known as the "after end of speech" timer.
- For optimal performance, set this parameter per grammar, as follows:
- 500 for short utterances (yes/no, hot word)
- 750 for simple queries
- 1200 to 2500 For digit strings and long utterances (listings) -->
- <param name="incompletetimeout">
- <value>2500</value>
- </param>
- <param name="swirec_retain_grammar_import_separator">
- <value>0</value>
- </param>
- <param name="swirec_waveform_logging_uniform_name">
- <value>0</value>
- </param>
- <!-- Controls sensitivity of the speech detector when
- looking for speech. -->
- <param name="sensitivity">
- <value>0.5</value>
- </param>
- <!-- Controls the trade-off between processing and accuracy. -->
- <param name="speedvsaccuracy">
- <value>0</value>
- </param>
- <!-- Sets the confidence threshold for recognition results
- computed while the selective_barge_in mode is active. -->
- <param name="swirec_selective_barge_in_conf_thresh">
- <value>700</value>
- </param>
- <!-- Sets the confidence threshold for recognition results
- computed while the magic word endpointer mode is active. -->
- <param name="swirec_magic_word_conf_thresh">
- <value>700</value>
- </param>
- <!-- For Nuance use only. Sets mode of the INTERNAL endpointer. -->
- <!-- OSR sets this param as needed for different audio input. -->
- <!-- For example, Aurora data requires that the endpointer be disabled. -->
- <param name="swirec_internal_endpointer_mode">
- <value>end_only</value>
- </param>
- <!-- This parameter is ignored and is reserved for future use. -->
- <param name="swirec_audio_environment">
- <value>default</value>
- </param>
- <!-- Amount of speech, in milliseconds, that the
- recognizer sees before forcing end of speech. -->
- <param name="swirec_max_speech_duration">
- <value>40000</value>
- </param>
- <!-- Number of entries to return in the nbest list. If >5, additional
- CPU cycles will be needed. -->
- <param name="swirec_nbest_list_length">
- <value>2</value>
- </param>
- <!-- Defines the format of the audio being supplied to the recognizer. -->
- <param name="swirec_audio_media_type">
- <value>audio/basic;rate=8000</value>
- <value>audio/x-alaw-basic;rate=8000</value>
- <!-- value>application/x-feature;rate=8000;encoding=speechpearl</value -->
- </param>
- <!-- Controls whether to log the SWIrslt event -->
- <param name="swirec_log_xmlresult">
- <value>true</value>
- </param>
- <!-- Add a Nuance grammar key to the XML result. -->
- <param name="swirec_extra_nbest_keys">
- <value>SWI_meaning</value>
- <value>SWI_literal</value>
- <value>SWI_grammarName</value>
- </param>
- <!-- Ignore the media type that is returned by the server
- upon fetching a grammar. -->
- <param name="swirec_ignore_grammar_media_type">
- <value>1</value>
- </param>
- <!-- log grammarID in the GRNM token of SWIrcst event -->
- <param name="swirec_enable_grammar_id_logging">
- <value>0</value>
- </param>
- <!-- Max parses evaluated for each literal and grammar
- Valid values are 1-99 -->
- <param name="swirec_max_parses_per_literal">
- <value>10</value>
- </param>
- <!-- Set of characters that are allowed in a sampa pronunciation -->
- <param name="swirec_sampa_characters">
- <value>#123459:;=?@ACDEHIJMNOQRSTUVYZ_abdefghijklmnoprstuvwxyz{~</value>
- </param>
- <!-- Max source grammar size (in bytes) that is allowed to be loaded
- by the online system or non-sgc tools -->
- <param name="swirec_max_source_grammar_size">
- <value>-1</value>
- </param>
- <!-- Max training grammar size (in bytes) that is allowed to be loaded
- by the online system or non-sgc tools -->
- <param name="swirec_max_training_grammar_size">
- <value>-1</value>
- </param>
- <!-- For Nuance use only: Set this to 0 to enable support for answer grammars -->
- <param name="swirec_suppress_answer_grammar">
- <value>1</value>
- </param>
- <!-- ****************************************************************** -->
- <!-- * New License Enforcement initialization parameters * -->
- <!-- ****************************************************************** -->
- <!-- How many osr_swirec licenses will be checked out during SWIrecInit() -->
- <param name="swirec_license_ports">
- <value>4</value>
- </param>
- <!-- How many osr_swiep licenses will be checked out during SWIepInit() -->
- <param name="swiep_license_ports">
- <value>4</value>
- </param>
- <!-- Identifies a threshold of licenses, above which will be considered an overdraft. -->
- <param name="swirec_license_ports_overdraft_thresh">
- <value>-1</value>
- </param>
- <!-- Identifies a threshold of licenses, above which will be considered an overdraft. -->
- <param name="swiep_license_ports_overdraft_thresh">
- <value>-1</value>
- </param>
- <!-- Identifies the features to be checked out by the recognizer -->
- <param name="swirec_licensing_features">
- <value>osr_nl_u, osr_rec_u, dtmf</value>
- </param>
- <!-- Identifies the features to be checked out by the endpointer -->
- <param name="swiep_licensing_features">
- <value/>
- </param>
- <!-- Upon detection of network communicate failure with license server, wait grace
- period milliseconds before disabling features; maximum 3600 -->
- <param name="swilicense_grace_period">
- <value>3600</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Parameters related to CACHING * -->
- <!-- ***************************************************************** -->
- <!-- Amount of memory the internal cache needs to take up before we
- start expiring grammars, in MB -->
- <param name="swirec_memory_cache_size">
- <value>100</value>
- </param>
- <!-- When maximum cache size is reached, what must the cache size be
- reduced to to stop expiring grammars, in MB. Defaults to 85% of
- above size -->
- <!--
- <param name="swirec_memory_cache_low_water_mark">
- <value>34</value>
- </param>
- -->
- <!-- Minimum size of compiled grammar to cache in KB -->
- <param name="swirec_memory_cache_min_entry_size">
- <value>0</value>
- </param>
- <!-- Level to compile newly cached grammars -->
- <param name="swirec_default_optimization">
- <value>6</value>
- </param>
- <!-- Level to compile often-used grammars -->
- <param name="swirec_full_optimization">
- <value>9</value>
- </param>
- <!-- Defines "often-used" grammar. It is the number of times
- a grammar is accessed from a cache before the
- grammar is automatically optimized using the level
- in swirec_disk_cache_full_optimization -->
- <param name="swirec_hits_before_full_optimize">
- <value>3</value>
- </param>
- <!-- User dictionary cache size in MB. The default value is 0 i.e. caching is disabled by default. -->
- <param name="swirec_max_dict_cache_size">
- <value>0</value>
- </param>
- <!-- Defines whether the disk cache should be used. -->
- <param name="swirec_disk_cache_enabled">
- <value>1</value>
- </param>
- <!-- Directory name for the disk cache. -->
- <!-- If relative, the file path is relative to the containing configuration file. -->
- <param name="swirec_disk_cache_directory">
- <value>../GrammarCache</value>
- </param>
- <!-- Maximum size of the disk cache in MB -->
- <param name="swirec_disk_cache_size">
- <value>500</value>
- </param>
- <!-- Minimum size of compiled grammar to cache in KB -->
- <param name="swirec_disk_cache_min_entry_size">
- <value>0</value>
- </param>
- <!-- When maximum cache size is reached, what must the cache size be
- reduced to to stop expiring grammars, in MB. -->
- <param name="swirec_disk_cache_low_water_mark">
- <value>400</value>
- </param>
- <!-- Address a http proxy server to use, e.g. 127.0.0.1 -->
- <param name="swirec_inet_proxy_server">
- <value/>
- </param>
- <!-- Port of the http proxy server to use, e.g. "8080";
- set this only if swirec_inet_proxy_server is set -->
- <param name="swirec_inet_proxy_server_port">
- <value/>
- </param>
- <!-- Delimiters which are used to separate fields in a grammar
- key value query string. The default is ";&", to allow "x=y;z=d" or
- "x=y&z=d" -->
- <param name="swirec_inet_query_delimiters">
- <value>;&</value>
- </param>
- <!-- Table of translations from grammar language code to internal one -->
- <!-- If relative, this file path is relative to baseline (SWISRSDK). -->
- <param name="swirec_language_translation_table">
- <value>generic/langtable.text</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Parameters related to ENDPOINTER * -->
- <!-- ***************************************************************** -->
- <!-- This parameter is ignored and is reserved for future use. -->
- <param name="swiep_audio_environment">
- <value>default</value>
- </param>
- <!-- Backoff is a safety margin to ensure the begin-of-speech is captured.
- The margin extends the BOS marker slightly (milliseconds) sooner in
- the audio sample. -->
- <param name="swiep_BOS_backoff">
- <value>200</value>
- </param>
- <!-- Backoff is a safety margin to ensure the end-of-speech is captured.
- The margin extends the EOS marker slightly (milliseconds) later in
- the audio sample. -->
- <param name="swiep_EOS_backoff">
- <value>240</value>
- </param>
- <!-- Inform the speech detector about the audio format being
- supplied to the recognizer. -->
- <param name="swiep_audio_media_type">
- <value>audio/basic;rate=8000</value>
- <value>audio/x-alaw-basic;rate=8000</value>
- </param>
- <!-- Controls how loudly the caller must speak to interrupt
- prompts (barge-in) in order for speech to be detected. -->
- <param name="swiep_in_prompt_sensitivity_percent">
- <value>50</value>
- </param>
- <!-- Turns barge-in on/off. If a dialog state is not using barge-in, then set this parameter to 0.
- The input received by the endpointer differs depending on whether barge-in is enabled or not.
- When barge-in is enabled, the ratio of silence to speech sent to the endpointer increases
- considerably; also, the sound that follows the first recognition call is usually the residual
- prompt or background noise. When barge-in is not enabled, the first sound processed is usually
- the callers speech. Setting this parameter ensures that the endpointer is optimized
- for either condition.
- -->
- <param name="bargein">
- <value>1</value>
- </param>
- <!-- Sets the speech detector into various endpointer modes. -->
- <param name="swiep_mode">
- <value>begin_only</value>
- </param>
- <!-- Must be set if SWIepRead is used by platform -->
- <param name="swiep_use_swiep_read">
- <value>1</value>
- </param>
- <!-- Disable pre-endpointed waveform logging entirely. -->
- <param name="swiep_suppress_waveform_logging">
- <value>1</value>
- </param>
- <!-- Disable logging of endpointer events. -->
- <param name="swiep_suppress_event_logging">
- <value>0</value>
- </param>
- <!-- Max channels to simultaneously save pre-endpointed waveforms. -->
- <param name="swiep_waveform_logging_max_channels">
- <value>-1</value>
- </param>
- <!-- Timeout (seconds) for inactive channel when max_channels
- is set -->
- <param name="swiep_waveform_logging_channel_timeout">
- <value>300</value>
- </param>
- <!-- Prevent thrashing on tiny input buffers, each of which
- has a constant fixed cost. Too big hurts response time;
- too small increases cpu; 800 recommended for 8khz ulaw
- legal values 1 to 1999 -->
- <param name="swiep_min_bytes_to_process">
- <value>800</value>
- </param>
- <!-- Controls how much memory is spent holding samples waiting for read
- to catch up to write; in msecs -->
- <param name="swiep_max_read_lag">
- <value>10000</value>
- </param>
- <!-- suppresses barge-in for this many msecs; can stop early barge-ins -->
- <param name="swiep_suppress_barge_in_time">
- <value>0</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Advanced endpointer configuration -->
- <!-- * In general, you should not need to adjust these parameters; the -->
- <!--* endpointer is optimized to provide the best performance with -->
- <!--* default settings. Nuance will continue to revamp these settings -->
- <!--* based on field results. -->
- <!-- ***************************************************************** -->
- <!-- ** Start of the call
- Nuance provides parameters that allow the endpointer to reject speech-like residual
- echoes, which are typically strongest at the start of the call. ** -->
- <!-- This parameter specifies the speech level (on a dBm scale) of the most quiet utterance
- that can trigger the endpointer at the start of the first recognition. This parameter keeps
- the endpointer from being triggered by prompt echoes at the beginning of the first recognition,
- before the echo cancellation might have converged. This minimum level gradually decreases to
- the value specified by swiep_bargein_min_noise_floor according to the timing values specified by
- swiep_bargein_initial_hold_seconds and swiep_bargein_initial_decay_seconds. This parameter is used only
- when bargein is set to 1. -->
- <param name="swiep_bargein_initial_noise_floor">
- <value>-25</value>
- </param>
- <!-- This parameter specifies how long (in seconds) the initial noise floor set by
- swiep_bargein_initial_noise_floor is held. After this time, the initial noise floor starts
- to transition to the value specified in swiep_bargein_min_noise_floor. The transition lasts for the
- duration specified in swiep_bargein_initial_decay_seconds. This parameter is used only when bargein
- is set to 1. -->
- <param name="swiep_bargein_initial_hold_seconds">
- <value>1.5</value>
- </param>
- <!-- This parameter specifies the duration (in seconds) during which the minimum noise floor transitions
- from the value specified in swiep_bargein_initial_noise_floor to the value specified in
- swiep_bargein_min_noise_floor. This parameter is used only when bargein is set to 1. -->
- <param name="swiep_bargein_initial_decay_seconds">
- <value>0.5</value>
- </param>
- <!-- This parameter specifies whether to automatically truncate the initial hold time when
- the endpointer is notified that the prompt has completed -->
- <param name="swiep_stop_initial_hold_on_prompt_done">
- <value>1</value>
- </param>
- <!-- This parameter specifies the absolute minimum level of speech (on a dBm scale) that can trigger
- the endpointer. This parameter is always used, regardless of the value in parameter bargein. -->
- <param name="swiep_bargein_min_noise_floor">
- <value>-45</value>
- </param>
- <!-- Detection sensitivity
- generalized sensitivity settings are offered through 'sensitivity' and 'swiep_in_prompt_sensitivity_percent'.
- Sensitivity can be fine-tuned through adjustment of parameters in this section. -->
- <!-- Set this parameter to 0 to disable 'sensitivity' and 'swiep_in_prompt_sensitivity_percent'
- before tuning the remaining parameters in this section.
- -->
- <param name="swiep_use_sensitivity">
- <value>1</value>
- </param>
- <!-- This parameter is the primary parameter controlling the sensitivity of the endpointer.
- Increasing it requires a larger change in energy to trigger the endpointer, and decreasing
- it makes the endpointer more sensitive. -->
- <param name="swiep_threshold_snr">
- <value>14</value>
- </param>
- <!-- The value of this parameter is added to the threshold specified by parameter
- swiep_threshold_snr to detect the first start-of-speech in a call. A larger positive
- number increases the detection threshold for the first utterance. After one successful
- recognition, this parameter is ignored. This parameter does not apply to the detection of
- end-of-speech -->
- <param name="swiep_threshold_snr_first_offset">
- <value>3</value>
- </param>
- <!-- The value of this parameter is applied to the threshold specified by parameter
- swiep_threshold_snr to detect the end-of-speech. A larger negative number makes it
- more likely that phrases ending weakly are classified as speech. -->
- <param name="swiep_threshold_snr_end_offset">
- <value>-3</value>
- </param>
- <!-- The value of this parameter specifies the search range for a smart back-up that
- looks for low-energy consonants (msec) -->
- <param name="swiep_consonant_search_width">
- <value>550</value>
- </param>
- <!-- The value of this parameter specifies the difference between the measured phrase strength
- and the minimum strength required to trigger the endpointer, that may represent a consonant. -->
- <param name="swiep_consonant_offset">
- <value>3</value>
- </param>
- <!-- ** Timing
- This section describes new and updated timing parameters used to determine the
- begin-of-speech and end-of-speech. ** -->
- <!-- To trigger the start-of-speech, a certain percentage of frames
- (specified by swiep_start_percentage) during a time window
- (specified by swiep_start_seconds) must be above the threshold specified by
- swiep_threshold_snr. The default values mean that the duration of speech activity required
- to determine that speech has started is about 0.045 seconds. -->
- <param name="swiep_start_percentage">
- <value>0.15</value>
- </param>
- <param name="swiep_start_seconds">
- <value>0.3</value>
- </param>
- <!-- To trigger the end-of-speech, a certain percentage of frames (specified by swiep_end_percentage)
- during a time window (specified by incompletetimeout) must be below the detection thresholds. -->
- <param name="swiep_end_percentage">
- <value>1.0</value>
- </param>
- <!-- ** Adapting speech models
- The endpointer keeps an adapting speech model of the signal level for the current caller.
- To trigger the endpointer, the current signal must be reasonably close to or above the adapting
- speech model. This section describes the parameters that control the adapting speech models.** -->
- <!-- If the spectral estimate of the input signal is lower than the speech model of the speaker, it
- has to be within the threshold specified by this parameter to trigger the endpointer. For example, if
- the speech model of the speaker is 46 and parameter swiep_voice_threshold is set to 5, the input signal
- must be 41 or higher to trigger the endpointer. Increasing this value (typically to 10) will let more
- quiet utterances and weak onsets trigger the endpointer when the speech model for the current caller
- is loud. -->
- <param name="swiep_voice_threshold">
- <value>5</value>
- </param>
- <!-- This parameter provides an initial estimate (on a dBm scale) of how loud the speaker will be.
- This is used by the endpointer when comparing the current level of the signal to an internally-adapting
- model of the speech frames that have been processed so far. Increasing this parameter can lead to fewer
- false triggers, and decreasing it can lead to fewer misses. This parameter is independent of the
- swiep_threshold_snr and noise floor-related parameters. -->
- <param name="swiep_initial_voice_bin">
- <value>-35</value>
- </param>
- <!-- For Nuance use only. -->
- <param name="swiep_dbg_verbosity">
- <value>0</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Parameters related to HOT INSERT and Grammar Preloading-->
- <!-- ***************************************************************** -->
- <!--Number of seconds between attempts to update baseline data -->
- <param name="swirec_update_interval">
- <value>600</value>
- </param>
- <!-- If this file exists, hot insert will NOT be done, use this to
- temporarily stop hot insert while a group of files is updated. -->
- <param name="swirec_update_lockfile">
- <value>lockfile.dat</value>
- </param>
- <!-- If this is set, preloaded grammars will never be removed from
- the cache, even if they normally would expire or SWIrecGrammarFree() is called -->
- <param name="swirec_lock_preload_grammars">
- <value>0</value>
- </param>
- <!-- Set default properties used for SWIrecGrammarLoad,Activate,Free(), See
- Properties (VXIinet) in Reference Manual for configurable properties -->
- <param name="swirec_default_loadfree_properties">
- <value/>
- <!-- Example:
- <value>swirec_mem_flush_on_free=1</value>
- <value>inet.maxage=0</value>
- -->
- </param>
- <param name="swirec_preload_file">
- <value>SWIgrmPreload.xml</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Parameters related to LEARN * -->
- <!-- ***************************************************************** -->
- <!-- For Nuance use only. -->
- <param name="swirec_learn_timeout">
- <value/>
- </param>
- <!-- For Nuance use only. -->
- <param name="swirec_learn_server">
- <value/>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Parameters related to ACCURACY AND LANGUAGE MODELING * -->
- <!-- ***************************************************************** -->
- <!-- For Nuance use only. Sets a maximum number of models (e.g. male/female).
- Typically never changed. -->
- <param name="swirec_num_parallel_models">
- <value>2</value>
- </param>
- <!-- Language model scale. Scale used to get language model within same
- dynamic range as acoustic models. -->
- <param name="swirec_lmweight">
- <value>1.0</value>
- </param>
- <!-- For Nuance use only: Weight of lm during rescore phase. Absolute value (not dependent on lmweight). -->
- <param name="swirec_lmrescore_weight">
- <value>2.0</value>
- </param>
- <!-- For Nuance use only -->
- <param name="swirec_max_arc_cost">
- <value>6.5</value>
- </param>
- <!-- For Nuance use only -->
- <param name="swirec_max_epsilon_arc_cost">
- <value>6.0</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * SSM defaults * -->
- <!-- ***************************************************************** -->
- <!-- If confidence of ssm is below this number, do not return result.
- Range:0.0-1.0 -->
- <param name="swissm_confidence_threshold">
- <value>0.0</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Word level confidence parameters * -->
- <!-- ***************************************************************** -->
- <!-- Enabled/Disabled word level confidence estimates -->
- <param name="swirec_word_confidence_enabled">
- <value>0</value>
- </param>
- <!-- Word posterior pruning. Controls the density of the word lattice
- by pruning all word hypothesis with an initial posterior estimate
- larger than this value. Also relates to the amount of CPU spent
- computing the word lattice an ultimately the word level confidence
- estimates -->
- <param name="swirec_word_posterior_pruning">
- <value>7.0</value>
- </param>
- <!-- Controls the density of the lattice used for computing word consensus
- hypothesis. Need to keep this value relatively low to mimize CPU -->
- <param name="swirec_word_lattice_density">
- <value>100.0</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Slot level confidence parameters * -->
- <!-- ***************************************************************** -->
- <!-- Used to calculate SP sentence confidence -->
- <param name="swirec_slot_confidence_norm_factor">
- <value>1.0</value>
- </param>
- <!-- Interpolation factor to combine OSR and SP slot confidence values -->
- <param name="swirec_combined_slot_confidence_factor">
- <value>1.0</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Parameters related to PERFORMANCE * -->
- <!-- ***************************************************************** -->
- <!-- Speed vs Accuracy tradeoff. Valid settings are idle, normal, busy, pegged and on.
- Other parameters that use one of several values based on the setting of this
- parameter will choose their value in the order above (idle=first_value, normal=
- second_value, busy=third_value, and pegged=fourth_value.
- "On" means the load_adjusted speed vs accuracy is dynamic. -->
- <param name="swirec_load_adjusted_speedvsaccuracy">
- <value>on</value>
- </param>
- <!-- Maximum number of dictionary pronunciations to use per word. -->
- <param name="swirec_max_dict_prons">
- <value>8</value>
- </param>
- <!-- Maximum number of automatically generated pronunciations per word. -->
- <param name="swirec_max_auto_prons">
- <value>1</value>
- </param>
- <!-- For Nuance use only. Controls the second pass re-scoring stage.
- The will only execute on N-best entries from the first pass that are
- within the specified delta from each other. Concept: if the first pass
- gives a big delta, you can be fairly confident that the second
- N-best is not the correct answer, and a second pass re-score
- will probably not force it to the top either. Higher numbers
- can increase accuracy, but will slow response time. -->
- <param name="swirec_max_rescore_delta">
- <value>500</value>
- <value>500</value>
- <value>500</value>
- <value>500</value>
- </param>
- <!-- For Nuance use only. Related to the second pass re-scoring stage.
- Controls the absolute time spent for rescoring. -->
- <param name="swirec_max_rescore_time">
- <value>20000</value>
- <value>20000</value>
- <value>20000</value>
- <value>20000</value>
- </param>
- <!-- The maximum amount of time that the recognizer spends
- to recognize an answer. Milliseconds. -->
- <param name="swirec_max_cpu_time">
- <value>20000</value>
- </param>
- <!-- The maximum time of the search phase of recognition. Milliseconds. -->
- <param name="swirec_max_search_time">
- <value>5000</value>
- </param>
- <!-- Defines an approximate usage percentage as "normal" CPU load.
- When usage is normal, the recognizer can dynamically use more CPU
- to improve accuracy. -->
- <param name="swirec_normal_cpu_start">
- <value>15</value>
- </param>
- <!-- Defines an approximate usage percentage as "busy" CPU load.
- When usage is busy, the recognizer does not use extra CPU
- to refine its results. -->
- <param name="swirec_busy_cpu_start">
- <value>40</value>
- </param>
- <!-- Defines an approximate usage percentage as "maxium" CPU load.
- When usage is pegged, the recognizer minimizes CPU usage. -->
- <param name="swirec_pegged_cpu_start">
- <value>101</value>
- </param>
- <!-- The maximum number of candidates to consider when attempting
- to fill the nbest list. This is not the preferred way to control
- nbest processing and should only be set small when there is a specific
- reason each sentence is very expensive (like a computationally intense
- ECMAscript). Otherwise accuracy and confidence can be seriously impaired
- for no reason. By default (value of 99999999) this parameter has no effect.
- This parameter can also be set via SetParameter or a <meta> in a grammar. -->
- <param name="swirec_max_sentences_tried">
- <value>999999</value>
- </param>
- <!-- This is the main search parameter guiding the Viterbi beam search.
- It is used to prune all active paths. These are paths either active
- during Hmm state propagation or paths active during FSM state propagation. -->
- <param name="swirec_state_beam">
- <value>0</value>
- <value>-15</value>
- <value>-45</value>
- <value>-45</value>
- </param>
- <!-- This is equivalent to the above state_beam, but for the parallel
- deocding of the GSM only. -->
- <param name="swirec_gsm_state_beam">
- <value>-50</value>
- </param>
- <!-- This determines the vocab threshold that decides whether we use the full or small
- gsm. However, we always use the small gsm in busy -->
- <param name="swirec_gsm_vocab_size_threshold">
- <value>100</value>
- </param>
- <!-- Determines how much the phoneme look-ahead influences the Viterbi beam search.
- Typically this parameter is set to match the beam search parameter.
- Conceivably a better trade-off can be achieved by using a tighter beam
- for the phoneme look-ahead compared to the main Viterbi beam. -->
- <param name="swirec_phoneme_lookahead_beam">
- <value>-30</value>
- <value>-45</value>
- <value>-60</value>
- <value>-60</value>
- </param>
- <!-- Determines how much to reduce the state beam when the current best path i
- ends in a silence model. This effectively reduces the search space by
- tighter pruning when the search is confident of being in a silence region. -->
- <param name="swirec_silence_prune_offset">
- <value>56</value>
- <value>56</value>
- <value>56</value>
- <value>56</value>
- </param>
- <!-- Maximum number of FSM arcs that can be active at any time. This parameter
- is used to guide the histogram pruning algorithm that dynamically adjusts
- the state beam to ensure the total number of active arcs does not
- exceed swirec_max_arcs. -->
- <param name="swirec_max_arcs">
- <value>10000</value>
- <value>6000</value>
- <value>3000</value>
- <value>3000</value>
- </param>
- <!-- For Nuance use only. The word beam applies only at word endings. The
- hypothesis is that search paths at end of words can be pruned tighter than
- other paths, without loss of accuracy. -->
- <param name="swirec_word_beam">
- <value>-50</value>
- </param>
- <!-- For Nuance use only. Penalty for self loop transitions in the Hmm model -->
- <param name="swirec_loop_penalty">
- <value>0.3</value>
- </param>
- <!-- For Nuance use only. Penalty for next state transition in Hmm model. -->
- <param name="swirec_next_penalty">
- <value>0.2</value>
- </param>
- <!-- For Nuance use only. Penalty to skip a Hmm state. -->
- <param name="swirec_skip_penalty">
- <value>0.6</value>
- </param>
- <!-- For Nuance use only. Penalty for entering silence model. -->
- <param name="swirec_pause_penalty">
- <value>0.6</value>
- </param>
- <!-- For Nuance use only. Self loop penalty for the Hmm silence state. -->
- <param name="swirec_silence_penalty">
- <value>0.0</value>
- </param>
- <!-- For Nuance use only. Penalty for entering a whole word unit model. -->
- <param name="swirec_whole_word_unit_penalty">
- <value>2.0</value>
- </param>
- <!-- For Nuance use only. Word transition penalty. -->
- <param name="swirec_word_penalty">
- <value>1.0</value>
- </param>
- <!-- For Nuance use only. Start-of-sentence cost. -->
- <param name="swirec_sentence_initial_cost">
- <value>1.0</value>
- </param>
- <!-- For Nuance use only. Forward search forces a
- model restart if this number of frames have passed -->
- <param name="swirec_restart_neighborhood">
- <value>25</value>
- </param>
- <!-- For Nuance use only. Generate lattice states for
- one frame over two; value 0 is turn off -->
- <param name="swirec_lattice_decimation">
- <value>0</value>
- </param>
- <!-- For Nuance use only: Phonetic Pruning activation -->
- <param name="swirec_ppr_active">
- <value>1</value>
- </param>
- <!-- For Nuance use only.
- threshold of phone posterior to prune start arc -->
- <param name="swirec_ppr_start_thres">
- <value>0.00001</value>
- </param>
- <!-- For Nuance use only.
- threshold of phone posterior to prune existing arc -->
- <param name="swirec_ppr_exist_thres">
- <value>0.0</value>
- </param>
- <!-- For Nuance use only.
- maximum threshold of phone posterior to prune arc after correction
- based on position in the beam -->
- <param name="swirec_ppr_max_thres">
- <value>0.01</value>
- </param>
- <!-- For Nuance use only.
- Scaling factor use to correct the ppr threshold -->
- <param name="swirec_ppr_thres_scale_factor">
- <value>170.667</value>
- </param>
- <!-- For Nuance use only. Whether to share contents of public rules referenced by other public
- rules. Turning on speeds grammar compilation and lowers compiled memory
- size but results in recognition-time CPU increase. -->
- <param name="swirec_shared_publics">
- <value>0</value>
- </param>
- <!-- For Nuance use only. Max number of nodes that can be visited during the astar search
- before it is aborted -->
- <param name="swirec_astar_max_paths">
- <value>100000</value>
- </param>
- <!-- Turning on causes an implicit-root (of all grammar publics) to
- be defined if an explicit root rule is not declared in the grammar.
- Note that this behavior does not seem to be compatible with
- the relevant w3c specifications. -->
- <param name="swirec_compute_implicit_root">
- <value>0</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * Parameters related to DEBUGGING * -->
- <!-- ***************************************************************** -->
- <!-- Maximum n-best entries logged in event log. -->
- <param name="swirec_max_logged_nbest">
- <value>2</value>
- </param>
- <!-- For Nuance use only. Turns on/off the writing of a computation statistics
- file. These statistics show the amounts of time spent during various
- recognition steps. -->
- <param name="swirec_save_comp_stats">
- <value>0</value>
- </param>
- <!-- Disable post-endpointed waveform logging entirely. -->
- <param name="swirec_suppress_waveform_logging">
- <value>0</value>
- </param>
- <!-- Disable logging of recognition events. -->
- <param name="swirec_suppress_event_logging">
- <value>0</value>
- </param>
- <!-- Current recognition should be treated as sensitive information
- and thus suppressed according to suppression rules. -->
- <param name="secure_context">
- <value>0</value>
- </param>
- <!-- Suppressed sensitive information from URI according to suppression
- rules. Not settable at runtime. Default to '0' (do not suppress) -->
- <param name="secure_context_uri">
- <value>0</value>
- </param>
- <!-- Max channels to simultaneously save waveforms. -->
- <param name="swirec_waveform_logging_max_channels">
- <value>-1</value>
- </param>
- <!-- Timeout (seconds) for inactive channel when max_channels
- is set, Specifies the number of seconds during which a rec
- handle must be idle in order to be considered inactive and
- therefore not logging any waveforms -->
- <param name="swirec_waveform_logging_channel_timeout">
- <value>300</value>
- </param>
- <!-- Allows/disallows obtaining waveforms by calling SWIrecGetWaveform() -->
- <param name="swirec_return_waveform">
- <value>1</value>
- </param>
- <!-- Length of leading silence (msecs) in returned waveform
- when calling SWIrecGetWaveformEx with 'begin' silence_suppression -->
- <param name="swirec_waveform_begin_silence">
- <value>0</value>
- </param>
- <!-- Length of trailing silence (msecs) in returned waveform
- when calling SWIrecGetWaveformEx with 'end' silence_suppression -->
- <param name="swirec_waveform_end_silence">
- <value>0</value>
- </param>
- <!-- Maximum length of interword silence (msecs) in returned waveform
- when calling SWIrecGetWaveformEx with 'interword' silence_suppression -->
- <param name="swirec_waveform_interword_max">
- <value>0</value>
- </param>
- <!-- Percentage of the maximum waveform amplitude to be used
- as threshold for detecting speech -->
- <param name="swirec_waveform_speech_thresh">
- <value>10.0</value>
- </param>
- <!-- Safety margin (msecs) extending the BOS and EOS in the waveform -->
- <param name="swirec_waveform_speech_backoff">
- <value>50</value>
- </param>
- <!-- ***************************************************************** -->
- <!-- * NEVER CHANGE WITHOUT Nuance authorization * -->
- <!-- ***************************************************************** -->
- <!-- For Nuance use only. -->
- <!-- Wait if completetimeout is gt incompletetimeout -->
- <param name="swirec_allow_wait_completetimeout">
- <value>0</value>
- </param>
- <!-- For Nuance use only. -->
- <!-- If relative, this file path is relative to baseline (SWISRSDK). -->
- <param name="swifeature_config_name">
- <value>generic/models/feature.txt</value>
- </param>
- <!-- For Nuance use only. -->
- <!-- File containing data related to confidence and rejection. -->
- <param name="swirec_reject_info_name">
- <value>generic/models/reject.text</value>
- </param>
- <!-- For Nuance use only. Sets up path to OSR language-specific configuration files. -->
- <!-- If relative, this file path is relative to baseline (SWISRSDK). -->
- <param name="swirec_language_baseline">
- <value>@LANGUAGENAME@/@[email protected]</value>
- </param>
- <param name="swirec_rp_info_name">
- <value>generic/rp_info.txt</value>
- </param>
- <param name="swirec_fixed_score_delta_conf">
- <value>0</value>
- </param>
- </lang>
- </SWIrecConfig>
Advertisement
Add Comment
Please, Sign In to add comment
Advertisement