ngram_decode.c File Reference

Word prediction based on N-gram probability and word trellis index for 2nd pass of Julius. More...

#include <julius.h>

Include dependency graph for ngram_decode.c:

Go to the source code of this file.

Functions
static int	compare_nw (NEXTWORD **a, NEXTWORD **b)
static NEXTWORD *	search_nw (NEXTWORD **nw, WORD_ID w, int num)
static void	set_word_context (WORD_ID *cseq, int n, WORD_INFO *winfo)
static int	pick_backtrellis_words (BACKTRELLIS *bt, WORD_INFO *winfo, NGRAM_INFO *ngram, NEXTWORD **nw, int oldnum, NODE *hypo, short t)
	Extract next word candidates from word trellis.
int	get_backtrellis_words (BACKTRELLIS *bt, WORD_INFO *winfo, NGRAM_INFO *ngram, NEXTWORD **nw, NODE *hypo, short tm, short t_end)
	Look for the next word candidates on the word trellis near the specified time frame.
int	limit_nw (NEXTWORD **nw, NODE *hypo, int num)
int	ngram_firstwords (NEXTWORD **nw, int peseqlen, int maxnw, WORD_INFO *winfo, BACKTRELLIS *bt)
	Return the set of initial word hypotheses at the beginning.
int	ngram_nextwords (NODE *hypo, NEXTWORD **nw, int maxnw, NGRAM_INFO *ngram, WORD_INFO *winfo, BACKTRELLIS *bt)
	Return the list of next word candidate.
boolean	ngram_acceptable (NODE *hypo, WORD_INFO *winfo)
Variables
static WORD_ID	cnword [2]
	Last two non-transparent words.
static int	cnnum
	Num of found non-transparent words (<=2).
static int	last_trans
	Num of skipped transparent words.

---

Detailed Description

Word prediction based on N-gram probability and word trellis index for 2nd pass of Julius.

Author:

Akinobu Lee

Date:

Fri Jul 8 14:57:51 2005

These functions returns next word candidates in the 2nd recognition pass of Julius, i.e. N-gram based stack decoding.

Given a partial sentence hypothesis, it first estimate the beginning frame of the hypothesis based on the word trellis. Then the words in the word trellis around the estimated frame are extracted from the word trellis. They will be returned with their N-gram probabilities.

In Julius, ngram_firstwords(), ngram_nextwords() and ngram_acceptable() are called from main search function wchmm_fbs(). In Julian, corresponding functions in dfa_decode.c will be used instead.

Revision

1.1.1.1

Definition in file ngram_decode.c.

---

Function Documentation

static int compare_nw (  NEXTWORD **  a, NEXTWORD **  b )  [static]

qsort callback function to sort next word candidates by their word ID. Parameters: a  [in] element 1 b  [in] element 2 Returns: 1 if word id of a > that of b, -1 if negative, 0 if equal. Definition at line 71 of file ngram_decode.c. Referenced by get_backtrellis_words().

static NEXTWORD* search_nw (  NEXTWORD **  nw, WORD_ID  w, int  num )  [static]

Find a word from list of next word candidates. Parameters: nw  [in] list of next word candidates w  [in] word id to search for num  [in] length of nw Returns: the pointer to the NEXTWORD data if found, or NULL if not found. Definition at line 101 of file ngram_decode.c. Referenced by pick_backtrellis_words().

static void set_word_context (  WORD_ID *  cseq, int  n, WORD_INFO *  winfo )  [static]

Set last two non-transparent words in the given word sequence and set them to cnword. Parameters: cseq  [in] word sequence n  [in] length of cseq winfo  [in] word dictionary information Definition at line 147 of file ngram_decode.c. Referenced by pick_backtrellis_words().

static int pick_backtrellis_words (  BACKTRELLIS *  bt, WORD_INFO *  winfo, NGRAM_INFO *  ngram, NEXTWORD **  nw, int  oldnum, NODE *  hypo, short  t )  [static]

Extract next word candidates from word trellis. This function extracts the list of trellis words whose word end has survived in the word trellis at the specified frame. The N-gram probabilities of them are then computed and added to the current next word candidates data. Parameters: bt  [in] word trellis structure winfo  [in] word dictionary structure ngram  [in] N-gram data structure nw  [in] list of next word candidates (new words will be appended at oldnum) oldnum  [in] number of words already stored in nw hypo  [in] the source sentence hypothesis t  [in] specified frame Returns: the total number of words currently stored in the nw. Definition at line 205 of file ngram_decode.c. Referenced by get_backtrellis_words().

int get_backtrellis_words (  BACKTRELLIS *  bt, WORD_INFO *  winfo, NGRAM_INFO *  ngram, NEXTWORD **  nw, NODE *  hypo, short  tm, short  t_end )

Look for the next word candidates on the word trellis near the specified time frame. This function builds a list of next word candidates by looking up the word trellis at specified frame, with lookup_range frame margin. If the same words exists in the near frames, only the one nearest to the specified frame will be chosen. Parameters: bt  [in] word trellis structure winfo  [in] word dictionary structure ngram  [in] word N-gram structure nw  [out] pointer to hold the extracted words as list of next word candidates hypo  [in] partial sentence hypothesis from which the words will be expanded tm  [in] center time frame to look up the words t_end  [in] right frame boundary for the lookup. Returns: the number of next words candidates stored in nw. Definition at line 314 of file ngram_decode.c. Referenced by ngram_nextwords().

int limit_nw (  NEXTWORD **  nw, NODE *  hypo, int  num )

Remove words in the nextword list which should not be expanded. Parameters: nw  [i/o] list of next word candidates (will be shrinked by removing some words) hypo  [in] partial sentence hypothesis from which the words will be expanded num  [in] current number of next words in nw Returns: the new number of words in nw Definition at line 396 of file ngram_decode.c. Referenced by ngram_nextwords().

int ngram_firstwords (  NEXTWORD **  nw, int  peseqlen, int  maxnw, WORD_INFO *  winfo, BACKTRELLIS *  bt )

Return the set of initial word hypotheses at the beginning. on N-gram based recogntion, the initial hypothesis is fixed to the tail silence word. Exception is that, in short-pause segmentation mode, the initial hypothesis will be chosen from survived words on the last input frame in the first pass. Parameters: nw  [out] pointer to hold the initial word candidates peseqlen  [in] input frame length maxnw  [in] maximum number of words that can be stored in nw winfo  [in] word dictionary information bt  [in] word trellis structure Returns: the number of words extracted and stored to nw. Definition at line 467 of file ngram_decode.c. Referenced by wchmm_fbs().

int ngram_nextwords (  NODE *  hypo, NEXTWORD **  nw, int  maxnw, NGRAM_INFO *  ngram, WORD_INFO *  winfo, BACKTRELLIS *  bt )

Return the list of next word candidate. Given a partial sentence hypothesis "hypo", it returns the list of next word candidates. Actually, it extracts from word trellis the list of words whose word-end node has survived near the estimated beginning-of-word frame of last word "hypo->estimated_next_t", and store them to "nw" with their N-gram probabilities. Parameters: hypo  [in] source partial sentence hypothesis nw  [out] pointer to store the list of next word candidates (should be already allocated) maxnw  [in] maximum number of words that can be stored to nw ngram  [in] word N-gram winfo  [in] word dictionary bt  [in] word trellis structure Returns: the number of extracted next word candidates in nw. Definition at line 553 of file ngram_decode.c. Referenced by wchmm_fbs().

boolean ngram_acceptable (  NODE *  hypo, WORD_INFO *  winfo )

Return whether the given partial hypothesis is acceptable as a sentence and can be treated as a final search candidate. In N-gram mode, it checks whether the last word is the beginning-of-sentence silence (silhead). Parameters: hypo  [in] partial sentence hypothesis to be examined winfo  [in] word dictionary Returns: TRUE if acceptable as a sentence, or FALSE if not. Definition at line 606 of file ngram_decode.c. Referenced by wchmm_fbs().

---