Configuration for tuning

After deployment to a production system, applications accept calls from real users (not test users), and the system generates call logs for each session that occurs. The logs represent a history of in-production performance, and are critical for gaining insight to application performance.

Application developers use the call logs (and call center reports from users and agents) to identify problems. The solution to a problem might involve a change to configuration, application logic, prompts, or communications between components (such as database interactions). Fixing problems with the configuration is the least expensive technique. No recompilation or deployment is needed. It is only necessary to keep a record of changes on the production system so they can be propagated to any mirrored systems, and replicated in the original development environment for future updates.

For an overview of the tuning lifecycle, see Tuning applications. The tuning mechanisms are the same as during application development (Configuration during application development).

Tuning initialization speed and licensing bottlenecks

To optimize the speed of Recognizer initialization, and subsequent availability of licenses, check the values of the parameters:

Parameter	Description	Default
swiep_license_ports	Number of licenses to check out during initialization of the speech detector (endpointer).	8 (licenses)
swirec_license_ports	The number of licenses to check out during initialization.	4 (licenses)
swirec_preload_file	File that loads grammars during initialization.	$SWISRSDK/config/SWIgrmPreload.xml
swirec_shmem_enabled	Determines whether recognition servers share memory.	0 (disabled)

Tuning user utterances

These parameters tune the collection of user speech:

Parameter	Description	Default
bargein	Allows callers to interrupt prompts.	1 (enabled)
completetimeout	How long to wait before concluding that a caller is finished speaking.	0 (timer disabled)
incompletetimeout	Duration of silence to determine that callers have finished speaking.	1500 (milliseconds)
maxspeechtimeout	Maximum duration of an utterance collected from users.	-1 (no timeout)
swiep_suppress_barge_in_time	Disables barge-in briefly at the beginning of a prompt.	0 (no delay)
timeout	How long to wait for speech after a prompt ends.	7000 (milliseconds)

These parameters tune the collected audio:

Parameter	Description	Default
sensitivity	Sensitivity of the speech detector when looking for speech.	0.5
swiep_BOS_backoff	Safety margin to ensure that the begin-of-speech is captured.	200 (milliseconds)
swiep_EOS_backoff	Safety margin to ensure the end-of-speech is captured.	350 (milliseconds)
swiep_in_prompt_sensitivity_percent	Controls how loudly callers must speak to interrupt prompts (barge-in) and detect speech.	50 (percent)
swirec_max_cpu_time	Maximum CPU time used to recognize an answer.	20000 (milliseconds)

These parameters control magic word recognition, which enables the application to recognize a specific speech or DTMF sequence and ignore anything else.

Parameter	Description	Default
swiep_magic_word_max_msec	Maximum duration of a magic word candidate for recognition.	800 (milliseconds)
swiep_magic_word_min_msec	Minimum duration of a magic word candidate for recognition.	200 (milliseconds)
swiep_mode	Sets special recognition modes (such as magic word) in the endpointer.	begin_only
swirec_barge_in_mode	Sets special recognition modes in Recognizer.	normal

Tuning accuracy, memory, and CPU load

These parameters accept multiple values to maximize accuracy given the available CPU capacity. They can improve accuracy during periods of reduced load, and slightly reduce accuracy during periods of heavy load to avoid delays in response time.

Parameter	Description	Default
swirec_load_adjusted_cpu_ranges	Defines ranges of system activity (idle, normal, and busy) based on CPU capacity.	0, 14, 40, 101 (percentages)
swirec_load_adjusted_speedvsaccuracy	Overrides the automatic detection of CPU load, and forces specific values for parameters that balance speed and accuracy.	on
swirec_max_arcs	Maximum number of active FSM arcs.	10000, 5000, 3000
swirec_phoneme_lookahead_beam	Provides a secondary guide to the Viterbi beam search.	-30, -60, -60
swirec_silence_prune_offset	Limits search paths that end in a silence model during pruning.	56, 56, 56
swirec_state_beam	Primary guide for the Viterbi beam search.	0, -15, -35
swirec_acoustic_adapt_suppress_adaptation	Temporarily stops self-learning activities for one or more languages.	(depends on language)

These parameters balance accuracy and CPU load:

Parameter	Description	Default
swirec_astar_max_paths	Maximum number of nodes visited during the a-star search.	100000 (nodes visited)
swirec_max_search_time	Maximum CPU time for the search phase of recognition.	5000 (milliseconds)
swirec_max_sentences_tried	Maximum number of candidates for filling the n-best list.	999999 (sentences)
swirec_word_lattice_density	Density of the word lattice.	100.0
swirec_word_posterior_pruning	Density of the word lattice.	7.0

These parameter optimize memory usage when you run more than one recognition server on a host:

Parameter	Description	Default
swirec_shmem_address	Base address in memory for Recognizers to use for shared memory.	(empty, the default address is used)
swirec_shmem_enabled	Determines whether recognition servers share memory.	0 (disabled)
swirec_shmem_resource_load_timeout	Maximum time for a process to wait for a resource in shared memory.	30 (seconds)
swirec_shmem_size	Size of the memory region shared by recognition server processes.	0 (automatically determined)
swirec_shmem_warn_threshold	Generates warnings when shared memory is getting filled.	90 (percent)

This parameter is useful when Nuance provides a custom language model for an application:

Parameter	Description	Default
swirec_lmweight	Adds weight to match the dynamic ranges of language and acoustic models.	1.0

These parameters are useful when Nuance provides custom acoustic models for an application:

Parameter	Description	Default
swirec_secondpass_allophone_mapfile_name	Defines allophone maps for secondpass processing in Recognizer.	(default mapfiles used)
swirec_secondpass_global_fsm_name	Defines finite state machines for secondpass processing in Recognizer.	(default fsm files used)
swirec_secondpass_model_name	Acoustic models for secondpass processing in Recognizer.	(default models used)

Cache performance

These parameters control Recognizer’s caching behavior:

Parameter	Description	Default
swirec_disk_cache_enabled	Enables (or disables) the disk and inet caches.	1 (enabled)
swirec_full_optimization	Default optimization level for fully-optimized grammars.	9
swirec_hits_before_full_optimize	Number of previous activations before fully optimizing a grammar.	3
swirec_disk_cache_low_water_mark	Largest size of the disk and inet caches after cache cleanup.	0 (adaptation enabled)
swirec_disk_cache_min_entry_size	Minimum size of disk and inet caches entries.	0 (KB)
swirec_disk_cache_size	Desired maximum size of the disk and inet caches.	500 (MB)
swirec_lock_preload_grammars	Keeps preloaded grammars in the memory cache.	0 (flushing allowed)
swirec_memory_cache_low_water_mark	Maximum size of the memory cache after removing grammars to create available space.	85 (MB)
swirec_memory_cache_min_entry_size	Minimum size for memory cache entries.	0 (KB)
swirec_memory_cache_size	Desired size of the memory cache.	100 (MB)

For details on caching, see Understanding grammar caching.

Troubleshooting licensing issues

Performance-related licensing problems can be detected these parameters:

Parameter	Description	Default
swiep_license_ports_overdraft_thresh	Number of licenses the endpointer can claim before the system generates a warning.	-1 (disabled)
swirec_license_ports_overdraft_thresh	The number of licenses Recognizer can claim before the system generates a warning.	-1 (disabled)

In addition to checking Recognizer’s diagnostic log for licensing issues, check the logs from the connected License Manager.

Configuration for tuning

Related topics