s3_align.c File Reference

Engine for Sphinx 3 aligner. More...

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>
#include <feat.h>
#include <strfuncs.h>
#include <s3types.h>
#include "mdef.h"
#include "tmat.h"
#include "dict.h"
#include "logs3.h"
#include "s3_align.h"

Classes
struct	pnode_s
struct	plink_s
struct	history_s
struct	snode_s
struct	slink_s
Defines
#define	ACTIVE_LIST_SIZE_INCR   16380
Typedefs
typedef pnode_s	pnode_t
typedef plink_s	plink_t
typedef history_s	history_t
typedef snode_s	snode_t
typedef slink_s	slink_t
Functions
int32	align_build_sent_hmm (char *wordstr, int insert_sil)
int32	align_destroy_sent_hmm (void)
void	align_sen_active (uint8 *senlist, int32 n_sen)
int32	align_start_utt (char *uttid)
int32	align_frame (int32 *senscr)
int32	align_end_utt (align_stseg_t **stseg_out, align_phseg_t **phseg_out, align_wdseg_t **wdseg_out)
int32	align_init (mdef_t *_mdef, tmat_t *_tmat, dict_t *_dict, cmd_ln_t *_config, logmath_t *_logmath)
void	align_free (void)

---

Detailed Description

Engine for Sphinx 3 aligner.

---

Define Documentation

#define ACTIVE_LIST_SIZE_INCR   16380

---

Typedef Documentation

typedef struct history_s history_t

Viterbi search history for each state at each time.

typedef struct plink_s plink_t

A may have links (transitions) to several successor or predecessor nodes. They are captured by a list of the following plink_t type.

typedef struct pnode_s pnode_t

Phone-level sentence HMM structures: pnode_t: nodes of phones forming sentence HMM. plink_t: a link between two pnode_t nodes. A phone node may have multiple successors and/or predecessors because of multiple alternative pronunciations for a word, as well as the presence of OPTIONAL filler words. Assumptions: No cycles in phone level sentence HMM.

typedef struct slink_s slink_t

typedef struct snode_s snode_t

State DAG structures similar to phone DAG structures.

---

Function Documentation

int32 align_build_sent_hmm (  char *  wordstr, int  insert_sil )

Build a sentence HMM for the given transcription (wordstr). A two-level DAG is built: phone-level and state-level. <s> and </s> always added at the beginning and end of sentence to form an augmented transcription. Optional <sil> and noise words added between words in the augmented transcription. wordstr must contain only the transcript; no extraneous stuff such as utterance-id. Phone-level HMM structure has replicated nodes to allow for different left and right context CI phones; hence, each pnode corresponds to a unique triphone in the sentence HMM. Return 0 if successful, <0 if any error (eg, OOV word encountered). Parameters: wordstr  In: Word transcript insert_sil  In: Whether to insert silences/fillers

int32 align_destroy_sent_hmm (  void   )

int32 align_end_utt (  align_stseg_t **  stseg_out, align_phseg_t **  phseg_out, align_wdseg_t **  wdseg_out )

All frames consumed. Trace back best Viterbi state sequence and dump it out. Parameters: stseg_out  Out: list of state segmentation phseg_out  Out: list of phone segmentation wdseg_out  Out: list of word segmentation

int32 align_frame (  int32 *  senscr  )

One frame of Viterbi time alignment. Parameters: senscr  In: array of senone scores this frame

void align_free (  void   )

int32 align_init (  mdef_t *  _mdef, tmat_t *  _tmat, dict_t *  _dict, cmd_ln_t *  _config, logmath_t *  _logmath )

void align_sen_active (  uint8 *  senlist, int32  n_sen )

Flag the active senones. Parameters: senlist  Out: senlist[s] TRUE iff active in frame n_sen  In: Size of senlist[] array

int32 align_start_utt (  char *  uttid  )

Start Viterbi alignment using the sentence HMM previously built. Assumes that each utterance will only be aligned once; state member variables initialized during sentence HMM building.

---