ngram_search.h
/* -*- c-basic-offset: 4; indent-tabs-mode: nil -*- */
/* ====================================================================
* Copyright (c) 2008 Carnegie Mellon University. All rights
* reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
*
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in
* the documentation and/or other materials provided with the
* distribution.
*
* This work was supported in part by funding from the Defense Advanced
* Research Projects Agency and the National Science Foundation of the
* United States of America, and the CMU Sphinx Speech Consortium.
*
* THIS SOFTWARE IS PROVIDED BY CARNEGIE MELLON UNIVERSITY ``AS IS'' AND
* ANY EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
* THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY
* NOR ITS EMPLOYEES BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
* SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
* THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
* OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
* ====================================================================
*
*/
/**
* @file ngram_search.h N-Gram based multi-pass search ("FBS")
*/
#ifndef __NGRAM_SEARCH_H__
#define __NGRAM_SEARCH_H__
/* SphinxBase headers. */
#include <sphinxbase/cmd_ln.h>
#include <sphinxbase/logmath.h>
#include <sphinxbase/ngram_model.h>
#include <sphinxbase/listelem_alloc.h>
#include <sphinxbase/err.h>
/* Local headers. */
#include "pocketsphinx_internal.h"
#include "hmm.h"
/**
* Lexical tree node data type.
*
* Not the first HMM for words, which multiplex HMMs based on
* different left contexts. This structure is used both in the
* dynamic HMM tree structure and in the per-word last-phone right
* context fanout.
*/
typedef struct chan_s {
hmm_t hmm; /**< Basic HMM structure. This *must* be first in
the structure because chan_t and root_chan_t are
sometimes used interchangeably */
struct chan_s *next; /**< first descendant of this channel; or, in the
case of the last phone of a word, the next
alternative right context channel */
struct chan_s *alt; /**< sibling; i.e., next descendant of parent HMM */
int32 ciphone; /**< ciphone for this node */
union {
int32 penult_phn_wid; /**< list of words whose last phone follows this one;
this field indicates the first of the list; the
rest must be built up in a separate array. Used
only within HMM tree. -1 if none */
int32 rc_id; /**< right-context id for last phone of words */
} info;
} chan_t;
/**
* Lexical tree node data type for the first phone (root) of each dynamic HMM tree
* structure.
*
* Each state may have a different parent static HMM. Most fields are
* similar to those in chan_t.
*/
typedef struct root_chan_s {
hmm_t hmm; /**< Basic HMM structure. This *must* be first in
the structure because chan_t and root_chan_t are
sometimes used interchangeably. */
chan_t *next; /**< first descendant of this channel */
int32 penult_phn_wid;
int32 this_phn_wid; /**< list of words consisting of this single phone;
actually the first of the list, like penult_phn_wid;
-1 if none */
int16 ciphone; /**< first ciphone of this node; all words rooted at this
node begin with this ciphone */
int16 ci2phone; /**< second ciphone of this node; one root HMM for each
unique right context */
} root_chan_t;
/**
* Back pointer table (forward pass lattice; actually a tree)
*/
typedef struct bptbl_s {
frame_idx_t frame; /**< start or end frame */
uint8 valid; /**< For absolute pruning */
uint8 refcnt; /**< Reference count (number of successors) */
int32 wid; /**< Word index */
int32 bp; /**< Back Pointer */
int32 score; /**< Score (best among all right contexts) */
int32 s_idx; /**< Start of BScoreStack for various right contexts*/
int32 real_wid; /**< wid of this or latest predecessor real word */
int32 prev_real_wid; /**< wid of second-last real word */
int16 last_phone; /**< last phone of this word */
int16 last2_phone; /**< next-to-last phone of this word */
} bptbl_t;
/**
* Segmentation "iterator" for backpointer table results.
*/
typedef struct bptbl_seg_s {
ps_seg_t base; /**< Base structure. */
int32 *bpidx; /**< Sequence of backpointer IDs. */
int16 n_bpidx; /**< Number of backpointer IDs. */
int16 cur; /**< Current position in bpidx. */
} bptbl_seg_t;
/*
* Candidates words for entering their last phones. Cleared and rebuilt in each
* frame.
* NOTE: candidates can only be multi-phone, real dictionary words.
*/
typedef struct lastphn_cand_s {
int32 wid;
int32 score;
int32 bp;
int32 next; /* next candidate starting at the same frame */
} lastphn_cand_t;
/*
* Since the same instance of a word (i.e., <word,start-frame>) reaches its last
* phone several times, we can compute its best BP and LM transition score info
* just the first time and cache it for future occurrences. Structure for such
* a cache.
*/
typedef struct {
int32 sf; /* Start frame */
int32 dscr; /* Delta-score upon entering last phone */
int32 bp; /* Best BP */
} last_ltrans_t;
#define CAND_SF_ALLOCSIZE 32
typedef struct {
int32 bp_ef;
int32 cand;
} cand_sf_t;
/*
* Structure for reorganizing the BP table entries in the current frame according
* to distinct right context ci-phones. Each entry contains the best BP entry for
* a given right context. Each successor word will pick up the correct entry based
* on its first ci-phone.
*/
typedef struct bestbp_rc_s {
int32 score;
int32 path; /* BP table index corresponding to this entry */
int32 lc; /* right most ci-phone of above BP entry word */
} bestbp_rc_t;
#define NO_BP -1
/**
* Various statistics for profiling.
*/
typedef struct ngram_search_stats_s {
int32 n_phone_eval;
int32 n_root_chan_eval;
int32 n_nonroot_chan_eval;
int32 n_last_chan_eval;
int32 n_word_lastchan_eval;
int32 n_lastphn_cand_utt;
int32 n_fwdflat_chan;
int32 n_fwdflat_words;
int32 n_fwdflat_word_transition;
int32 n_senone_active_utt;
} ngram_search_stats_t;
/**
* N-Gram search module structure.
*/
struct ngram_search_s {
ps_search_t base;
ngram_model_t *lmset; /**< Set of language models. */
hmm_context_t *hmmctx; /**< HMM context. */
/* Flags to quickly indicate which passes are enabled. */
uint8 fwdtree;
uint8 fwdflat;
uint8 bestpath;
/* State of procesing. */
uint8 done;
/* Allocators */
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 */
/**
* 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.
*/
root_chan_t *root_chan; /**< Roots of search tree. */
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 */
/**
* 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.
*/
chan_t **word_chan;
bitvec_t *word_active; /**< array of active flags for all words. */
/**
* 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.
*/
int32 *homophone_set;
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 */
/**
* 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.
*/
chan_t ***active_chan_list;
int32 n_active_chan[2]; /**< Number entries in active_chan_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.
*/
int32 **active_word_list;
int32 n_active_word[2]; /**< Number entries in active_word_list */
/*
* FIXME: Document all of these bits.
*/
lastphn_cand_t *lastphn_cand;
int32 n_lastphn_cand;
last_ltrans_t *last_ltrans; /* one per word */
int32 cand_sf_alloc;
cand_sf_t *cand_sf;
bestbp_rc_t *bestbp_rc;
bptbl_t *bp_table; /* Forward pass lattice */
int32 bpidx; /* First free BPTable entry */
int32 bp_table_size;
int32 *bscore_stack; /* Score stack for all possible right contexts */
int32 bss_head; /* First free BScoreStack entry */
int32 bscore_stack_size;
int32 n_frame_alloc; /**< Number of frames allocated in bp_table_idx and friends. */
int32 n_frame; /**< Number of frames actually present. */
int32 *bp_table_idx; /* First BPTable entry for each frame */
int32 *word_lat_idx; /* BPTable index for any word in current frame;
cleared before each frame */
/*
* Flat lexicon (2nd pass) search stuff.
*/
ps_latnode_t **frm_wordlist; /**< List of active words in each frame. */
int32 *fwdflat_wordlist; /**< List of active word IDs for utterance. */
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. */
int32 last_phone_best_score; /**< Best Viterbi path score for last phone. */
int32 renormalized;
/*
* DAG (3rd pass) search stuff.
*/
float32 bestpath_fwdtree_lw_ratio;
float32 ascale; /**< Acoustic score scale for posterior probabilities. */
ngram_search_stats_t st; /**< Various statistics for profiling. */
ptmr_t fwdtree_perf;
ptmr_t fwdflat_perf;
ptmr_t bestpath_perf;
int32 n_tot_frame;
/* A collection of beam widths. */
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;
};
typedef struct ngram_search_s ngram_search_t;
/**
* Initialize the N-Gram search module.
*/
ps_search_t *ngram_search_init(ngram_model_t *lm,
cmd_ln_t *config,
acmod_t *acmod,
dict_t *dict,
dict2pid_t *d2p);
/**
* Finalize the N-Gram search module.
*/
void ngram_search_free(ps_search_t *ngs);
/**
* Record the current frame's index in the backpointer table.
*
* @return the current backpointer index.
*/
int ngram_search_mark_bptable(ngram_search_t *ngs, int frame_idx);
/**
* Enter a word in the backpointer table.
*/
void ngram_search_save_bp(ngram_search_t *ngs, int frame_idx, int32 w,
int32 score, int32 path, int32 rc);
/**
* Allocate last phone channels for all possible right contexts for word w.
*/
void ngram_search_alloc_all_rc(ngram_search_t *ngs, int32 w);
/**
* Allocate last phone channels for all possible right contexts for word w.
*/
void ngram_search_free_all_rc(ngram_search_t *ngs, int32 w);
/**
* Find the best word exit for the current frame in the backpointer table.
*
* @return the backpointer index of the best word exit.
*/
int ngram_search_find_exit(ngram_search_t *ngs, int frame_idx, int32 *out_best_score, int32 *out_is_final);
/**
* Backtrace from a given backpointer index to obtain a word hypothesis.
*
* @return a <strong>read-only</strong> string with the best hypothesis.
*/
char const *ngram_search_bp_hyp(ngram_search_t *ngs, int bpidx);
/**
* Compute language and acoustic scores for backpointer table entries.
*/
void ngram_compute_seg_scores(ngram_search_t *ngs, float32 lwf);
/**
* Construct a word lattice from the current hypothesis.
*/
ps_lattice_t *ngram_search_lattice(ps_search_t *search);
/**
* Get the exit score for a backpointer entry with a given right context.
*/
int32 ngram_search_exit_score(ngram_search_t *ngs, bptbl_t *pbe, int rcphone);
/**
* Sets the global language model.
*
* Sets the language model to use if nothing was passed in configuration
*/
void ngram_search_set_lm(ngram_model_t *lm);
#endif /* __NGRAM_SEARCH_H__ */
