#include <srch_word_switch_tree.h>
Public Attributes | |
int32 | n_static_lextree |
lextree_t * | curroottree |
lextree_t ** | expandtree |
lextree_t * | curfillertree |
lextree_t ** | expandfillertree |
lextree_t ** | roottree |
hash_table_t * | active_word |
glist_t | empty_tree_idx_stack |
int32 | no_active_word |
lmset_t * | lmset |
int32 | isLMLA |
histprune_t * | histprune |
A note by ARCHAN at 20050703
In general, what made a LVCSR search to be unique has three major factors. They are 1) how high level knowledge source (LM, FSG) is applied, 2) how the lexicon is organized, 3, how cross word triphones is realized and implemented. It is important to realize the interplay between these 3 components to realize how a search actually really works.
hash_table_t* srch_WST_graph_t::active_word |
Hash table that map word end to the index of the expandtree.
lextree_t* srch_WST_graph_t::curfillertree |
The first filler tree
lextree_t* srch_WST_graph_t::curroottree |
The first unigram tree that used in the search for this utterance.
glist_t srch_WST_graph_t::empty_tree_idx_stack |
Store a pool of indices which indicate a tree is empty. It acts like a stack
lextree_t** srch_WST_graph_t::expandfillertree |
The expanded fillertrees
lextree_t** srch_WST_graph_t::expandtree |
The expanded trees
histprune_t* srch_WST_graph_t::histprune |
Structure that wraps up parameters related to
int32 srch_WST_graph_t::isLMLA |
Is LM lookahead used?
lmset_t* srch_WST_graph_t::lmset |
The LM set
int32 srch_WST_graph_t::n_static_lextree |
Number of static lexical tree for word switching
int32 srch_WST_graph_t::no_active_word |
No of active word
lextree_t** srch_WST_graph_t::roottree |
The pool of trees that stores all word trees. An array with dimension. #lm