wchmm.c

julius version 4.12.about sound recognition.

	  out_num_prev = out_num_next;
	}	/* end of phone loop */
      }	/* end of multipath block */
      
  } /* new phone node creation loop for this word */


  /*************************************/
  /* Short Pause appending (multipath) */
  /*************************************/
  
  /* if -iwsp, add noise model to the end of word at ntmp */
  if (wchmm->hmminfo->multipath && enable_iwsp && add_tail - add_head + 1 > 0) { /* there are new phones to be created */
    int ntmp_bak;
    
    /* set short pause state info */
    ntmp_bak = ntmp;
    if (wchmm->hmminfo->sp->is_pseudo) {
      for(k = 1;k < hmm_logical_state_num(wchmm->hmminfo->sp) - 1; k++) {
	wchmm->outstyle[ntmp] = AS_LSET;
	wchmm->state[ntmp].out.lset = &(wchmm->hmminfo->sp->body.pseudo->stateset[k]);
	acc_init(wchmm, ntmp);
	wchmm->stend[ntmp] = WORD_INVALID;
	ntmp++;
	if (ntmp >= wchmm->maxwcn) wchmm_extend(wchmm);
      }
    } else {
      for(k = 1;k < hmm_logical_state_num(wchmm->hmminfo->sp) - 1; k++) {
	wchmm->outstyle[ntmp] = AS_STATE;
	wchmm->state[ntmp].out.state = wchmm->hmminfo->sp->body.defined->s[k];
	acc_init(wchmm, ntmp);
	wchmm->stend[ntmp] = WORD_INVALID;
	ntmp++;
	if (ntmp >= wchmm->maxwcn) wchmm_extend(wchmm);
      }
    }
    ntmp = ntmp_bak;
    /* connect incoming arcs from previous phone */
    out_num_next = 0;
    for (ato = 1; ato < hmm_logical_state_num(wchmm->hmminfo->sp); ato++) {
      prob = hmm_logical_trans(wchmm->hmminfo->sp)->a[0][ato];
      if (prob != LOG_ZERO) {
	/* to control short pause insertion, transition probability toward
	 the word-end short pause will be given a penalty */
	prob += wchmm->hmminfo->iwsp_penalty;
	if (ato == hmm_logical_state_num(wchmm->hmminfo->sp) - 1) {
	  /* model has a model skip transition, just inherit them to next */
	  for(kkk=0; kkk<out_num_prev; kkk++) {
	    out_from_next[out_num_next] = out_from[kkk];
	    out_a_next[out_num_next] = out_a[kkk] + prob;
	    out_num_next++;
	  }
	} else {
	  /* connect incoming arcs from previous phone to this phone */
	  for(kkk=0; kkk<out_num_prev; kkk++) {
	    add_wacc(wchmm, out_from[kkk], out_a[kkk] + prob, ntmp + ato - 1);
	  }
	}
      }
    }
    /* if short pause model doesn't have a model skip transition, also add it */
    if (hmm_logical_trans(wchmm->hmminfo->sp)->a[0][hmm_logical_state_num(wchmm->hmminfo->sp)-1] == LOG_ZERO) {
      /* to make insertion sp model to have no effect on the original path,
	 the skip transition probability should be 0.0 (=100%) */
      prob = 0.0;
      for(kkk=0; kkk<out_num_prev; kkk++) {
	out_from_next[out_num_next] = out_from[kkk];
	out_a_next[out_num_next] = out_a[kkk] + prob;
	out_num_next++;
      }
    }
    /* connect arcs within model, and store new outgoing arcs for wordend node */
    for (k = 1; k < hmm_logical_state_num(wchmm->hmminfo->sp) - 1; k++) {
      for (ato = 1; ato < hmm_logical_state_num(wchmm->hmminfo->sp); ato++) {
	prob = hmm_logical_trans(wchmm->hmminfo->sp)->a[k][ato];
	if (prob != LOG_ZERO) {
	  if (ato == hmm_logical_state_num(wchmm->hmminfo->sp) - 1) {
	    out_from_next[out_num_next] = ntmp;
	    out_a_next[out_num_next] = prob;
	    out_num_next++;
	  } else {
	    add_wacc(wchmm, ntmp, prob, ntmp + ato - k);
	  }
	}
      }
      ntmp++;
    }
    /* swap work area for next */
    for(kkk=0;kkk<out_num_next;kkk++) {
      out_from[kkk] = out_from_next[kkk];
      out_a[kkk] = out_a_next[kkk];
    }
    out_num_prev = out_num_next;

  } /* end of inter-word short pause appending block */

  /* make mapping: word <-> node on wchmm */
  for (j=0;j<word_len;j++) {
    if (j < add_head) {	/* shared part */
      wchmm->offset[word][j] = wchmm->offset[matchword][j];
    } else if (add_tail < j) { /* shared tail part (should not happen..) */
      wchmm->offset[word][j] = wchmm->offset[matchword][j+(matchword_len-word_len)];
    } else {			/* newly created part */
      wchmm->offset[word][j] = n;
      n += hmm_logical_state_num(wchmm->winfo->wseq[word][j]) - 2;
    }
  }


  if (wchmm->hmminfo->multipath) {
    /* create word-end node */

    /* paranoia check if the short-pause addition has been done well */
    if (enable_iwsp && add_tail - add_head + 1 > 0) {
      n += hmm_logical_state_num(wchmm->hmminfo->sp) - 2;
      if (n != ntmp) j_internal_error("wchmm_add_word: cannot match\n");
    }
    
    /* create word-end node */
    wchmm->wordend[word] = n;	/* tail node of 'word' is 'n' */
    wchmm->stend[n] = word;	/* node 'k' is a tail node of 'word' */
    acc_init(wchmm, n);
    wchmm->state[n].out.state = NULL;
    
    /* connect the final outgoing arcs in out_from[] to the word end node */
    for(k = 0; k < out_num_prev; k++) {
      add_wacc(wchmm, out_from[k], out_a[k], n);
    }
    n++;
    if (n >= wchmm->maxwcn) wchmm_extend(wchmm);
    
    if (matchlen == 0) {
      /* check if the new word has whole word-skipping transition */
      /* (use out_from and out_num_prev temporary) */
      out_num_prev = 0;
      get_outtrans_list(wchmm, word, word_len-1, out_from, out_a, &out_num_prev, wchmm->winfo->maxwn, enable_iwsp);
      for(k=0;k<out_num_prev;k++) {
	if (out_from[k] == wchmm->wordbegin[word]) {
	  jlog("ERROR: *** ERROR: WORD SKIPPING TRANSITION NOT ALLOWED ***\n");
	  jlog("ERROR:   Word id=%d (%s[%s]) has \"word skipping transition\".\n", word, wchmm->winfo->wname[word], wchmm->winfo->woutput[word]);
	  jlog("ERROR:   All HMMs in the word:\n    ");
	  for(kkk=0;kkk<wchmm->winfo->wlen[word];kkk++) {
	    jlog("%s ", wchmm->winfo->wseq[word][kkk]->name);
	  }
	  jlog("\n");
	  jlog("ERROR:  has transitions from initial state to final state.\n");
	  jlog("ERROR:  This type of word skipping is not supported.\n");
	  ok_p = FALSE;
	}
      }
    }

    wchmm->n = n;

  } else {

    wchmm->n = n;
    k = wchmm->offset[word][word_len-1] + hmm_logical_state_num(wchmm->winfo->wseq[word][word_len-1])-2 -1;
    wchmm->wordend[word] = k;	/* tail node of 'word' is 'k' */
    wchmm->stend[k] = word;	/* node 'k' is a tail node of 'word' */
    
    if (matchlen != 0 && add_tail - add_head + 1 > 0) {
      /* new part has been created in the above procedure: */
      /* now make link from shared part to the new part */
      wchmm_link_subword(wchmm, matchword,add_to,word,add_head);	
    }

  }

  return(ok_p);
  
}

/*************************************************************/
/**** parse whole structure (after wchmm has been built) *****/
/*************************************************************/

/** 
 * <JA>
 * 腾菇陇步辑今を瘤汉し·帽胳の姜眉觉轮から嘲への肌莲败澄唯のリストを侯喇する. 
 * (non multipath)
 * 
 * @param wchmm [i/o] 腾菇陇步辑今
 * </JA>
 * <EN>
 * Scan the lexicon tree to make list of emission probability from the word end
 * state. (non multipath)
 * 
 * @param wchmm [i/o] tree lexicon
 * </EN>
 */
static void
wchmm_calc_wordend_arc(WCHMM_INFO *wchmm)
{
  WORD_ID w;
  HTK_HMM_Trans *tr;
  LOGPROB a;

  for (w=0;w<wchmm->winfo->num;w++) {
    tr = hmm_logical_trans(wchmm->winfo->wseq[w][wchmm->winfo->wlen[w]-1]);
    a = tr->a[tr->statenum-2][tr->statenum-1];
    wchmm->wordend_a[w] = a;
  }
}

#ifdef SEPARATE_BY_UNIGRAM

/********************************************************************/
/****** for separation (linearization) of high-frequent words *******/
/********************************************************************/

/** 
 * <JA>
 * unigram澄唯でソ〖トするための qsort コ〖ルバック簇眶. 
 * 
 * @param a [in] 妥燎1
 * @param b [in] 妥燎2
 * 
 * @return 遍换の冯蔡の射圭を手す. 
 * </JA>
 * <EN>
 * qsort callback function to sort unigram values.
 * 
 * @param a [in] element #1
 * @param b [in] element #2
 * 
 * @return the result of comparison.
 * </EN>
 */
static int
compare_prob(LOGPROB *a, LOGPROB *b)
{
  if (*a < *b)  return (1);
  if (*a > *b)  return (-1);
  return(0);
}

/** 
 * <JA>
 * 1-gramスコアの惧疤 N 戎誊の猛を滇める. 
 * 
 * @param winfo [in] 帽胳辑今
 * @param n [in] 滇める界疤
 * 
 * @return 惧疤 N 戎誊の uni-gram 澄唯の猛を手す. 
 * </JA>
 * <EN>
 * Get the Nth-best unigram probability from all words.
 * 
 * @param winfo [in] word dictionary
 * @param n [in] required rank
 * 
 * @return the Nth-best unigram probability.
 * </EN>
 */
static LOGPROB
get_nbest_uniprob(WCHMM_INFO *wchmm, int n)
{
  LOGPROB *u_p;
  WORD_ID w;
  LOGPROB x;
  WORD_INFO *winfo;
  NGRAM_INFO *ngram;

  winfo = wchmm->winfo;
  ngram = wchmm->ngram;

  if (n < 1) n = 1;
  if (n > winfo->num) n = winfo->num;

  /* store all unigram probability to u_p[] */
  u_p = (LOGPROB *)mymalloc(sizeof(LOGPROB) * winfo->num);
  for(w=0;w<winfo->num;w++) {
    if (ngram) {
      x = uni_prob(ngram, winfo->wton[w])
#ifdef CLASS_NGRAM
	+ winfo->cprob[w]
#endif
	;
    } else {
      x = LOG_ZERO;
    }
    if (wchmm->lmvar == LM_NGRAM_USER) {
      x = (*(wchmm->uni_prob_user))(wchmm->winfo, w, x);
    }
    u_p[w] = x;
  }

  /* sort them downward */
  qsort(u_p, winfo->num, sizeof(LOGPROB),
	(int (*)(const void *,const void *))compare_prob);

  /* return the Nth value */
  x = u_p[n-1];
  free(u_p);
  return(x);
}

#endif

/**********************************************************/
/****** MAKE WCHMM (LEXICON TREE) --- main function *******/
/**********************************************************/

#define COUNT_STEP 500         ///< Word count step for debug progress output

/** 
 * <JA>
 * 涂えられた帽胳辑今と咐胳モデルから腾菇陇步辑今を菇蜜する. この簇眶は
 * 借妄が觅く·Julianで"-oldtree"オプション回年箕のみ蝗脱されます. オプション
 * 润回年箕およびJuliusでは洛わりに build_wchmm2() が脱いられます. 
 * 
 * @param wchmm [i/o] 腾菇陇步辑今
 * @param lmconf [in] 咐胳モデル(LM)肋年パラメ〖タ
 * </JA>
 * <EN>
 * Build a tree lexicon from given word dictionary and language model.
 * This function is slow and only used when "-oldtree" option is specified
 * in Julian.  Julian without that option and Julius uses build_wchmm2()
 * instead of this.
 * 
 * @param wchmm [i/o] lexicon tree
 * @param lmconf [in] language model (LM) configuration parameters
 * </EN>
 * @callgraph
 * @callergraph
 */
boolean
build_wchmm(WCHMM_INFO *wchmm, JCONF_LM *lmconf)
{
  int i,j;
  int matchword=0, sharelen=0, maxsharelen=0;
  int num_duplicated;
#ifdef SEPARATE_BY_UNIGRAM
  LOGPROB separate_thres;
  LOGPROB p;
#endif
  boolean ok_p;

  /* lingustic infos must be set before build_wchmm() is called */
  /* check if necessary lingustic info is already assigned (for debug) */
  if (wchmm->winfo == NULL
      || (wchmm->lmvar == LM_NGRAM && wchmm->ngram == NULL)
      || (wchmm->lmvar == LM_DFA_GRAMMAR && wchmm->dfa == NULL)
      ) {
    jlog("ERROR: wchmm: linguistic info not available!!\n");
    return FALSE;
  }

  ok_p = TRUE;
  
#ifdef SEPARATE_BY_UNIGRAM
  /* 惧疤[separate_wnum]戎誊の1-gramスコアを滇める */
  /* 1-gramスコアがこの猛笆惧のものは腾から尸ける */
  separate_thres = get_nbest_uniprob(wchmm, lmconf->separate_wnum);
#endif

#ifdef PASS1_IWCD
#ifndef USE_OLD_IWCD
  if (wchmm->category_tree) {
    if (wchmm->ccd_flag) {
      /* 链てのカテゴリID烧き lcd_set を侯喇 */
      lcdset_register_with_category_all(wchmm);
    }
  }
#endif
#endif /* PASS1_IWCD */
  

  /* wchmmを介袋步 */
  wchmm_init(wchmm);

  /* カウンタリセット */
  wchmm->separated_word_count=0;

  jlog("STAT: wchmm: Building HMM lexicon tree (left-to-right)\n");
  for (i=0;i<wchmm->winfo->num;i++) {

    if (wchmm->lmtype == LM_PROB) {
      if (i == wchmm->winfo->head_silwid || i == wchmm->winfo->tail_silwid) {
	/* 黎片/琐萨の痰不モデルは腾菇陇步せず·
	 * 黎片の痰不帽胳の黎片への莲败·琐萨帽胳の琐萨からの莲败は侯らない*/
	/* sharelen=0でそのまま */
	if (wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp) == FALSE) {
	  jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n");
	  ok_p = FALSE;
	}
	continue;
      }
#ifndef NO_SEPARATE_SHORT_WORD
      if (wchmm->winfo->wlen[i] <= SHORT_WORD_LEN) {
	/* 墓さの没い帽胳を腾菇陇步しない(ここでは1不泪) */
	/* sharelen=0でそのまま */
	if (wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp) == FALSE) {
	  jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n");
	  ok_p = FALSE;
	}
	wchmm->separated_word_count++;
	continue;
      }
#endif
#ifdef SEPARATE_BY_UNIGRAM
      if (wchmm->ngram) {
	p = uni_prob(wchmm->ngram, wchmm->winfo->wton[i])
#ifdef CLASS_NGRAM
	  + wchmm->winfo->cprob[i]
#endif
	  ;