dict2pid.h File Reference

Building triphones for a dictionary. More...

#include <stdio.h>
#include <logmath.h>
#include "s3types.h"
#include "mdef.h"
#include "dict.h"
#include "ctxt_table.h"

Go to the source code of this file.

Classes
struct	dict2pid_t
	Building composite triphone (as well as word internal triphones) with the dictionary. More...
Defines
#define	dict2pid_internal(d, w, p)   ((d)->internal[w][p])
#define	dict2pid_n_comstate(d)   ((d)->n_comstate)
#define	dict2pid_n_comsseq(d)   ((d)->n_comsseq)
#define	dict2pid_is_composite(d)   ((d)->is_composite)
#define	IS_COMPOSITE   1
#define	NOT_COMPOSITE   0
Functions
dict2pid_t *	dict2pid_build (mdef_t *mdef, dict_t *dict, int32 is_composite, logmath_t *logmath)
void	dict2pid_free (dict2pid_t *d2p)
void	dict2pid_comsenscr (dict2pid_t *d2p, int32 *senscr, int32 *comsenscr)
void	dict2pid_comsseq2sen_active (dict2pid_t *d2p, mdef_t *mdef, uint8 *comssid, uint8 *sen)
void	dict2pid_dump (FILE *fp, dict2pid_t *d2p, mdef_t *mdef, dict_t *dict)
void	dict2pid_report (dict2pid_t *d2p)
int32	get_rc_nssid (dict2pid_t *d2p, s3wid_t w, dict_t *dict)
s3cipid_t *	dict2pid_get_rcmap (dict2pid_t *d2p, s3wid_t w, dict_t *dict)

---

Detailed Description

Building triphones for a dictionary.

This is one of the more complicated parts of a cross-word triphone model decoder. The first and last phones of each word get their left and right contexts, respectively, from other words. For single-phone words, both its contexts are from other words, simultaneously. As these words are not known beforehand, life gets complicated. In this implementation, when we do not wish to distinguish between distinct contexts, we use a COMPOSITE triphone (a bit like BBN's fast-match implementation), by clubbing together all possible contexts.

There are 3 cases:

1. Internal phones, and boundary phones without any specific context, in each word. The boundary phones are modelled using composite phones, internal ones using ordinary phones.

2. The first phone of a multi-phone word, for a specific history (i.e., in a 2g/3g/4g... tree) has known left and right contexts. The possible left contexts are limited to the possible last phones of the history. So it can be modelled separately, efficiently, as an ordinary triphone.

3. The one phone in a single-phone word, for a specific history (i.e., in a 2g/3g/4g... tree) has a known left context, but unknown right context. It is modelled using a composite triphone. (Note that right contexts are always composite, left contexts are composite only in the unigram tree.)

A composite triphone is formed as follows. (NOTE: this assumes that all CIphones/triphones have the same HMM topology, specifically, no. of states.) A composite triphone represents a situation where either the left or the right context (or both) for a given base phone is unknown. That is, it represents the set of all possible ordinary triphones derivable from * the unkown context(s). Let us call this set S. It is modelled using the same HMM topology * as the ordinary triphones, but with COMPOSITE states. A composite state (in a given position * in the HMM state topology) is the set of states (senones) at that position derived from S.

Actually, we generally deal with COMPOSITE SENONE-SEQUENCES rather than COMPOSITE PHONES. The former are compressed forms of the latter, by virtue of state sharing among phones. (See mdef.h.)

In 3.6, the composite triphone will only be build when -composite 1 (default) is specified. Other than that, full triphone expansion will be carried out in run-time

---

Define Documentation

#define dict2pid_internal (  d, w, p   )     ((d)->internal[w][p])

return internal dict2pid

#define dict2pid_is_composite (  d   )     ((d)->is_composite)

return whether dict2pid is in composite triphone mode or not

#define dict2pid_n_comsseq (  d   )     ((d)->n_comsseq)

return number of composite state sequence

#define dict2pid_n_comstate (  d   )     ((d)->n_comstate)

return number of composite state

#define IS_COMPOSITE   1

#define NOT_COMPOSITE   0

---

Function Documentation

dict2pid_t* dict2pid_build (  mdef_t *  mdef, dict_t *  dict, int32  is_composite, logmath_t *  logmath )

This segments of code take care of the initialization of internal[0] and ldiph[b][r][l] This part will take care of the initialization of internal[pronlen-1] and rdiph[b][l][r]. Notice that this is symmetric to the first part of the code. Parameters: mdef  A model definition dict  An initialized dictionary is_composite  Whether composite triphones will be built

void dict2pid_comsenscr (  dict2pid_t *  d2p, int32 *  senscr, int32 *  comsenscr )

Populate composite senone score array. The composite senone score is the maximum of its component senones' scores, scaled down by the number of component senones. Parameters: d2p  In: a dict2pid_t structure senscr  In: Ordinary senone scores comsenscr  Out: Composite senone scores

void dict2pid_comsseq2sen_active (  dict2pid_t *  d2p, mdef_t *  mdef, uint8 *  comssid, uint8 *  sen )

Mark senones active based on a set of active composite senones. Parameters: d2p  In: a dict2pid_t structure mdef  In: a mdef_t structure comssid  In: Active flag for each comssid sen  In/Out: Active flags set for senones indicated by the active comssid

void dict2pid_dump (  FILE *  fp, dict2pid_t *  d2p, mdef_t *  mdef, dict_t *  dict )

For debugging Parameters: fp  In: a file pointer d2p  In: a dict2pid_t structure mdef  In: a mdef_t structure dict  In: a dictionary structure

void dict2pid_free (  dict2pid_t *  d2p  )

Free the memory dict2pid structure Parameters: d2p  In: the d2p

s3cipid_t* dict2pid_get_rcmap (  dict2pid_t *  d2p, s3wid_t  w, dict_t *  dict )

Get RC map Parameters: d2p  In: a dict2pid w  In: a wid dict  In: a dictionary

void dict2pid_report (  dict2pid_t *  d2p  )

Report a dict2pid data structure Parameters: d2p  In: a dict2pid_t structure

int32 get_rc_nssid (  dict2pid_t *  d2p, s3wid_t  w, dict_t *  dict )

ARCHAN, A duplicate of get_rc_npid in ctxt_table.h. I doubt whether it is correct because the compressed map has not been checked. Parameters: d2p  In: a dict2pid w  In: a wid dict  In: a dictionary

---