- Generated by
1.8.5
|
PocketSphinx
5prealpha
|
N-Gram search module structure. More...
#include <ngram_search.h>
Data Fields | |
| ps_search_t | base |
| ngram_model_t * | lmset |
| Set of language models. More... | |
| hmm_context_t * | hmmctx |
| HMM context. More... | |
| uint8 | fwdtree |
| uint8 | fwdflat |
| uint8 | bestpath |
| uint8 | done |
| listelem_alloc_t * | chan_alloc |
| For chan_t. | |
| listelem_alloc_t * | root_chan_alloc |
| For root_chan_t. | |
| listelem_alloc_t * | latnode_alloc |
| For latnode_t. | |
| root_chan_t * | root_chan |
| Search structure of HMM instances. More... | |
| int32 | n_root_chan_alloc |
| Number of root_chan allocated. | |
| int32 | n_root_chan |
| Number of valid root_chan. | |
| int32 | n_nonroot_chan |
| Number of valid non-root channels. | |
| int32 | max_nonroot_chan |
| Maximum possible number of non-root channels. | |
| root_chan_t * | rhmm_1ph |
| Root HMMs for single-phone words. | |
| chan_t ** | word_chan |
| Channels associated with a given word (only used for right contexts, single-phone words in fwdtree search, and word HMMs in fwdflat search). More... | |
| bitvec_t * | word_active |
| array of active flags for all words. More... | |
| int32 * | homophone_set |
| Each node in the HMM tree structure may point to a set of words whose last phone would follow that node in the tree structure (but is not included in the tree structure for reasons explained above). More... | |
| int32 * | single_phone_wid |
| list of single-phone word ids | |
| int32 | n_1ph_words |
| Number single phone words in dict (total) | |
| int32 | n_1ph_LMwords |
| Number single phone dict words also in LM; these come first in single_phone_wid. | |
| chan_t *** | active_chan_list |
| Array of active channels for current and next frame. More... | |
| int32 | n_active_chan [2] |
| Number entries in active_chan_list. | |
| int32 ** | active_word_list |
| Array of active multi-phone words for current and next frame. More... | |
| int32 | n_active_word [2] |
| Number entries in active_word_list. | |
| lastphn_cand_t * | lastphn_cand |
| int32 | n_lastphn_cand |
| last_ltrans_t * | last_ltrans |
| int32 | cand_sf_alloc |
| cand_sf_t * | cand_sf |
| bestbp_rc_t * | bestbp_rc |
| bptbl_t * | bp_table |
| int32 | bpidx |
| int32 | bp_table_size |
| int32 * | bscore_stack |
| int32 | bss_head |
| int32 | bscore_stack_size |
| int32 | n_frame_alloc |
| Number of frames allocated in bp_table_idx and friends. More... | |
| int32 | n_frame |
| Number of frames actually present. More... | |
| int32 * | bp_table_idx |
| int32 * | word_lat_idx |
| ps_latnode_t ** | frm_wordlist |
| List of active words in each frame. More... | |
| int32 * | fwdflat_wordlist |
| List of active word IDs for utterance. More... | |
| bitvec_t * | expand_word_flag |
| int32 * | expand_word_list |
| int32 | n_expand_words |
| int32 | min_ef_width |
| int32 | max_sf_win |
| float32 | fwdflat_fwdtree_lw_ratio |
| int32 | best_score |
| Best Viterbi path score. More... | |
| int32 | last_phone_best_score |
| Best Viterbi path score for last phone. More... | |
| int32 | renormalized |
| float32 | bestpath_fwdtree_lw_ratio |
| float32 | ascale |
| Acoustic score scale for posterior probabilities. More... | |
| ngram_search_stats_t | st |
| Various statistics for profiling. More... | |
| ptmr_t | fwdtree_perf |
| ptmr_t | fwdflat_perf |
| ptmr_t | bestpath_perf |
| int32 | n_tot_frame |
| int32 | beam |
| int32 | dynamic_beam |
| int32 | pbeam |
| int32 | wbeam |
| int32 | lpbeam |
| int32 | lponlybeam |
| int32 | fwdflatbeam |
| int32 | fwdflatwbeam |
| int32 | fillpen |
| int32 | silpen |
| int32 | wip |
| int32 | nwpen |
| int32 | pip |
| int32 | maxwpf |
| int32 | maxhmmpf |
N-Gram search module structure.
Definition at line 197 of file ngram_search.h.
| chan_t*** ngram_search_s::active_chan_list |
Array of active channels for current and next frame.
In any frame, only some HMM tree nodes are active. active_chan_list[f mod 2] = list of nonroot channels in the HMM tree active in frame f.
Definition at line 275 of file ngram_search.h.
Referenced by ngram_fwdtree_deinit(), and ngram_fwdtree_finish().
| int32** ngram_search_s::active_word_list |
Array of active multi-phone words for current and next frame.
Similarly to active_chan_list, active_word_list[f mod 2] = list of word ids for which active channels exist in word_chan in frame f.
Statically allocated single-phone words are always active and should not appear in this list.
Definition at line 287 of file ngram_search.h.
Referenced by ngram_fwdflat_search(), ngram_fwdflat_start(), ngram_fwdtree_finish(), ngram_search_free(), and ngram_search_init().
| float32 ngram_search_s::ascale |
Acoustic score scale for posterior probabilities.
Definition at line 333 of file ngram_search.h.
| int32 ngram_search_s::best_score |
Best Viterbi path score.
Definition at line 325 of file ngram_search.h.
Referenced by ngram_fwdflat_search(), ngram_fwdflat_start(), ngram_fwdtree_search(), ngram_fwdtree_start(), and ngram_search_lattice().
| ps_latnode_t** ngram_search_s::frm_wordlist |
List of active words in each frame.
Definition at line 316 of file ngram_search.h.
Referenced by ngram_fwdflat_deinit(), ngram_fwdflat_init(), and ngram_search_mark_bptable().
| int32* ngram_search_s::fwdflat_wordlist |
List of active word IDs for utterance.
Definition at line 317 of file ngram_search.h.
Referenced by ngram_fwdflat_deinit(), ngram_fwdflat_init(), ngram_fwdflat_reinit(), and ngram_fwdflat_search().
| hmm_context_t* ngram_search_s::hmmctx |
HMM context.
Definition at line 200 of file ngram_search.h.
Referenced by ngram_fwdflat_search(), ngram_search_alloc_all_rc(), ngram_search_free(), and ngram_search_init().
| int32* ngram_search_s::homophone_set |
Each node in the HMM tree structure may point to a set of words whose last phone would follow that node in the tree structure (but is not included in the tree structure for reasons explained above).
The channel node points to one word in this set of words. The remaining words are linked through homophone_set[].
Single-phone words are not represented in the HMM tree; they are kept in word_chan.
Specifically, homophone_set[w] = wid of next word in the same set as w.
Definition at line 263 of file ngram_search.h.
| int32 ngram_search_s::last_phone_best_score |
Best Viterbi path score for last phone.
Definition at line 326 of file ngram_search.h.
| ngram_model_t* ngram_search_s::lmset |
Set of language models.
Definition at line 199 of file ngram_search.h.
Referenced by ngram_search_free(), and ngram_search_init().
| int32 ngram_search_s::n_frame |
Number of frames actually present.
Definition at line 308 of file ngram_search.h.
Referenced by ngram_fwdflat_search(), ngram_fwdflat_start(), ngram_fwdtree_search(), ngram_fwdtree_start(), ngram_search_find_exit(), and ngram_search_lattice().
| int32 ngram_search_s::n_frame_alloc |
Number of frames allocated in bp_table_idx and friends.
Definition at line 307 of file ngram_search.h.
Referenced by ngram_fwdflat_init(), ngram_search_init(), and ngram_search_mark_bptable().
| root_chan_t* ngram_search_s::root_chan |
Search structure of HMM instances.
The word triphone sequences (HMM instances) are transformed into tree structures, one tree per unique left triphone in the entire dictionary (actually diphone, since its left context varies dyamically during the search process). The entire set of trees of channels is allocated once and for all during initialization (since dynamic management of active CHANs is time consuming), with one exception: the last phones of words, that need multiple right context modelling, are not maintained in this static structure since there are too many of them and few are active at any time. Instead they are maintained as linked lists of CHANs, one list per word, and each CHAN in this set is allocated only on demand and freed if inactive.Roots of search tree.
Definition at line 231 of file ngram_search.h.
Referenced by ngram_fwdtree_finish().
| ngram_search_stats_t ngram_search_s::st |
Various statistics for profiling.
Definition at line 335 of file ngram_search.h.
Referenced by ngram_fwdflat_finish(), ngram_fwdflat_search(), ngram_fwdflat_start(), ngram_fwdtree_finish(), ngram_fwdtree_search(), and ngram_fwdtree_start().
| bitvec_t* ngram_search_s::word_active |
array of active flags for all words.
Definition at line 247 of file ngram_search.h.
Referenced by ngram_fwdflat_finish(), ngram_fwdflat_search(), ngram_fwdtree_finish(), ngram_search_free(), and ngram_search_init().
| chan_t** ngram_search_s::word_chan |
Channels associated with a given word (only used for right contexts, single-phone words in fwdtree search, and word HMMs in fwdflat search).
WARNING: For single-phone words and fwdflat search, this actually contains pointers to root_chan_t, which are allocated using root_chan_alloc. This is a suboptimal state of affairs.
Definition at line 246 of file ngram_search.h.
Referenced by ngram_fwdflat_reinit(), ngram_fwdflat_start(), ngram_fwdtree_finish(), ngram_fwdtree_reinit(), ngram_fwdtree_start(), ngram_search_alloc_all_rc(), ngram_search_free(), ngram_search_free_all_rc(), and ngram_search_init().
1.8.5