wchmm.c

julius version 4.12.about sound recognition.

      } else {
	p = LOG_ZERO;
      }
      if (wchmm->lmvar == LM_NGRAM_USER) {
	p = (*(wchmm->uni_prob_user))(wchmm->winfo, i, p);
      }
      if (p >= separate_thres && wchmm->separated_word_count < lmconf->separate_wnum) {
	/* 裳刨の光い帽胳を腾菇陇步しない */
	/* separate_thres は惧疤separate_wnum戎誊のスコア */
	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
    }

    /* 呵も墓く不燎を鼎铜叫丸る帽胳を玫す */
    maxsharelen=0;
    for (j=0;j<i;j++) {
      if (wchmm->category_tree  && wchmm->lmtype == LM_DFA) {
	if (wchmm->winfo->wton[i] != wchmm->winfo->wton[j]) continue;
      }
      sharelen = wchmm_check_match(wchmm->winfo, i, j);
      if (sharelen == wchmm->winfo->wlen[i] && sharelen == wchmm->winfo->wlen[j]) {
       /* word に票不胳が赂哼する */
       /* 涩ず呵络の墓さであり·脚剩カウントを闰けるためここで却ける */
       maxsharelen = sharelen;
       matchword = j;
       break;
      }
      if (sharelen > maxsharelen) {
       matchword = j;
       maxsharelen = sharelen;
      }
    }
    if (wchmm_add_word(wchmm, i, maxsharelen, matchword, lmconf->enable_iwsp) == FALSE) {
      jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n");
      ok_p = FALSE;
    }
  }

#if 0
  /* 腾菇陇を侯らない */
  for (i=0;i<wchmm->winfo->num;i++) {
    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;
    }
  }
#endif  
  jlog("STAT:  %5d words ended     (%6d nodes)\n",i,wchmm->n);

  if (! wchmm->hmminfo->multipath) {
    /* 票办不燎废误を积つ帽胳票晃の leaf node を2脚步して惰侍する */
    num_duplicated = wchmm_duplicate_leafnode(wchmm);
    jlog("STAT:  %d leaf nodes are made unshared\n", num_duplicated);
    
    /* 帽胳の姜眉から嘲への莲败澄唯を滇めておく */
    wchmm_calc_wordend_arc(wchmm);
  }

  /* wchmmの腊圭拉をチェックする */
  check_wchmm(wchmm);

  /* factoring脱に称觉轮に稿鲁帽胳のリストを烧裁する */
  if (!wchmm->category_tree) {

#ifdef UNIGRAM_FACTORING
    if (wchmm->lmtype == LM_PROB) {
      /* 票箕に涟もってfactoring猛を纷换 */
      make_successor_list_unigram_factoring(wchmm);
      jlog("STAT:  1-gram factoring values has been pre-computed\n");
    } else {
      make_successor_list(wchmm);
    }
#else 
    make_successor_list(wchmm);
#endif /* UNIGRAM_FACTORING */
    
    if (wchmm->hmminfo->multipath) {
      /* 菇蜜された factoring 攫鼠をスキップ莲败および矢片矢恕ノ〖ドにコピ〖 */
      adjust_sc_index(wchmm);
    }
    
#ifdef UNIGRAM_FACTORING
    if (wchmm->lmtype == LM_PROB) {
      /* 帽胳粗LMキャッシュが涩妥なノ〖ドのリストを侯る */
      make_iwcache_index(wchmm);
    }
#endif /* UNIGRAM_FACTORING */

    /* sclist2node is no longer used */
    if (wchmm->sclist2node != NULL) {
      free(wchmm->sclist2node);
      wchmm->sclist2node = NULL;
    }

  }

  jlog("STAT: done\n");

  return ok_p;
}

/** 
 * <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 bulid_wchmm() と票じ借妄を乖いますが·
 * 呵介に帽胳を不燎误でソ〖トして不燎误の击た界に帽胳を事べるため·
 * より光庐に腾菇陇步を乖うことができる. とくにオプション回年をしない
 * 嘎り·Julius/Julianではこちらが脱いられる. 
 * 
 * @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 does the same job as build_wchmm(), but it is much
 * faster because finding of the longest matched word to an adding word
 * is done by first sorting all the words in the dictoinary by their phoneme
 * sequence order.  This function will be used instead of build_wchmm()
 * by default.
 * 
 * @param wchmm [i/o] lexicon tree
 * @param lmconf [in] language model (LM) configuration parameters
 * </EN>
 * @callgraph
 * @callergraph
 */  
boolean
build_wchmm2(WCHMM_INFO *wchmm, JCONF_LM *lmconf)
{
  int i,j, last_i;
  int num_duplicated;
  WORD_ID *windex;
#ifdef SEPARATE_BY_UNIGRAM
  LOGPROB separate_thres;
  LOGPROB p;
#endif
  boolean ok_p;
  boolean ret;

  /* 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;
  
  wchmm->separated_word_count = 0;
  
  jlog("STAT: Building HMM lexicon tree\n");
  
  if (wchmm->lmtype == LM_PROB) {
#ifdef SEPARATE_BY_UNIGRAM
    /* compute score threshold beforehand to separate words from tree */
    /* here we will separate best [separate_wnum] words from tree */
    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) {
      /* when Julian mode (category-tree) and triphone is used,
	 make all category-indexed context-dependent phone set (cdset) here */
      /* these will be assigned on the last phone of each word on tree */
      lcdset_register_with_category_all(wchmm);
    }
  }
#endif
#endif /* PASS1_IWCD */

 /* initialize wchmm */
  wchmm_init(wchmm);

  /* make sorted word index ordered by phone sequence */
  windex = (WORD_ID *)mymalloc(sizeof(WORD_ID) * wchmm->winfo->num);
  for(i=0;i<wchmm->winfo->num;i++) windex[i] = i;

  if (wchmm->category_tree && wchmm->lmtype == LM_DFA) {

    /* sort by category -> sort by word ID in each category */
    wchmm_sort_idx_by_category(wchmm->winfo, windex, wchmm->winfo->num);
    {
      int last_cate;
      last_i = 0;
      last_cate = wchmm->winfo->wton[windex[0]];
      for(i = 1;i<wchmm->winfo->num;i++) {
	if (wchmm->winfo->wton[windex[i]] != last_cate) {
	  wchmm_sort_idx_by_wseq(wchmm->winfo, windex, last_i, i - last_i);
	  last_cate = wchmm->winfo->wton[windex[i]];
	  last_i = i;
	}
      }
      wchmm_sort_idx_by_wseq(wchmm->winfo, windex, last_i, wchmm->winfo->num - last_i);
    }

  } else {

    /* sort by word ID for whole vocabulary */
    wchmm_sort_idx_by_wseq(wchmm->winfo, windex, 0, wchmm->winfo->num);

  }

/* 
 *   {
 *     int i,w;
 *     for(i=0;i<wchmm->winfo->num;i++) {
 *	 w = windex[i];
 *	 printf("%d: cate=%4d wid=%4d %s\n",i, wchmm->winfo->wton[w], w, wchmm->winfo->woutput[w]);
 *     }
 *   }
 */

  /* incrementaly add words to lexicon tree */
  /* now for each word, the previous word (last_i) is always the most matched one */
  last_i = WORD_INVALID;
  for (j=0;j<wchmm->winfo->num;j++) {
    i = windex[j];

    if (wchmm->lmtype == LM_PROB) {

      /* start/end silence word should not be shared */
      if (i == wchmm->winfo->head_silwid || i == wchmm->winfo->tail_silwid) {
	/* add whole word as new (sharelen=0) */
	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
      /* separate short words from tree */
      if (wchmm->winfo->wlen[i] <= SHORT_WORD_LEN) {
	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
	  ;
      } else {
	p = LOG_ZERO;
      }
      if (wchmm->lmvar == LM_NGRAM_USER) {
	p = (*(wchmm->uni_prob_user))(wchmm->winfo, i, p);
      }
      /* separate high-frequent words from tree (threshold = separate_thres) */
      if (p >= separate_thres && wchmm->separated_word_count < lmconf->separate_wnum) {
	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
    }

    if (last_i == WORD_INVALID) { /* first word */
      ret = wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp);
    } else {
      /* the previous word (last_i) is always the most matched one */
      if (wchmm->category_tree && wchmm->lmtype == LM_DFA) {
	if (wchmm->winfo->wton[i] != wchmm->winfo->wton[last_i]) {
	  ret = wchmm_add_word(wchmm, i, 0, 0, lmconf->enable_iwsp);
	} else {
	  ret = wchmm_add_word(wchmm, i, wchmm_check_match(wchmm->winfo, i, last_i), last_i, lmconf->enable_iwsp);
	}
      } else {
	ret = wchmm_add_word(wchmm, i, wchmm_check_match(wchmm->winfo, i, last_i), last_i, lmconf->enable_iwsp);
      }
    }
    if (ret == FALSE) {
      jlog("ERROR: wchmm: failed to add word #%d to lexicon tree\n");
      ok_p = FALSE;
    }
    last_i = i;
    
  } /* end of add word loop */
  
  /*j_printerr("\r %5d words ended     (%6d nodes)\n",j,wchmm->n);*/

  /* free work area */
  free(windex);

  if (wchmm->hmminfo->multipath) {
    jlog("STAT: lexicon size: %d nodes\n", wchmm->n);
  } else {
    /* duplicate leaf nodes of homophone/embedded words */
    jlog("STAT: lexicon size: %d", wchmm->n);
    num_duplicated = wchmm_duplicate_leafnode(wchmm);
    jlog("+%d=%d\n", num_duplicated, wchmm->n);
  }

  if (! wchmm->hmminfo->multipath) {
    /* calculate transition probability of word end node to outside */
    wchmm_calc_wordend_arc(wchmm);
  }

  /* check wchmm coherence (internal debug) */
  check_wchmm(wchmm);

  /* make successor list for all branch nodes for N-gram factoring */
  if (!wchmm->category_tree) {

#ifdef UNIGRAM_FACTORING
    if (wchmm->lmtype == LM_PROB) {
      /* for 1-gram factoring, we can compute the values before search */
       make_successor_list_unigram_factoring(wchmm);
       jlog("STAT:  1-gram factoring values has been pre-computed\n");
    } else {
      make_successor_list(wchmm);
    }
#else
    make_successor_list(wchmm);
#endif /* UNIGRAM_FACTORING */
    if (wchmm->hmminfo->multipath) {
      /* Copy the factoring data according to the skip transitions and startword nodes */
      adjust_sc_index(wchmm);
    }
#ifdef UNIGRAM_FACTORING
    if (wchmm->lmtype == LM_PROB) {
      /* make list of start nodes that needs inter-word LM cache */
      make_iwcache_index(wchmm);
    }
#endif /* UNIGRAM_FACTORING */

    /* sclist2node is no longer used */
    if (wchmm->sclist2node != NULL) {
      free(wchmm->sclist2node);
      wchmm->sclist2node = NULL;
    }

  }

  //jlog("STAT: done\n");

#ifdef WCHMM_SIZE_CHECK
  if (debug2_flag) {
    /* detailed check of lexicon tree size (inaccurate!) */
    jlog("STAT: --- memory size of word lexicon ---\n");
    jlog("STAT: wchmm: %d words, %d nodes\n", wchmm->winfo->num, wchmm->n);
    jlog("STAT: %9d bytes: wchmm->state[node] (exclude ac, sc)\n", sizeof(WCHMM_STATE) * wchmm->n);
    {
      int count1 = 0;
      int count2 = 0;
      int count3 = 0;
      for(i=0;i<wchmm->n;i++) {
	if (wchmm->self_a[i] != LOG_ZERO) count1++;
	if (wchmm->next_a[i] != LOG_ZERO) count2++;
	if (wchmm->ac[i] != NULL) count3++;
      }
      jlog("STAT: %9d bytes: wchmm->self_a[node] (%4.1f%% filled)\n", sizeof(LOGPROB) * wchmm->n, 100.0 * count1 / (float)wchmm->n);
      jlog("STAT: %9d bytes: wchmm->next_a[node] (%4.1f%% filled)\n", sizeof(LOGPROB) * wchmm->n, 100.0 * count2 / (float)wchmm->n);
      jlog("STAT: %9d bytes: wchmm->ac[node] (%4.1f%% used)\n", sizeof(A_CELL2 *) * wchmm->n, 100.0 * count3 / (float)wchmm->n);
    }
    jlog("STAT: %9d bytes: wchmm->stend[node]\n", sizeof(WORD_ID) * wchmm->n);
    {
      int w,count;
      count = 0;
      for(w=0;w<wchmm->winfo->num;w++) {
	count += wchmm->winfo->wlen[w] * sizeof(int) + sizeof(int *);
      }
      jlog("STAT: %9d bytes: wchmm->offset[w][]\n", count);
    }
    if (wchmm->hmminfo->multipath) {
      jlog("STAT: %9d bytes: wchmm->wordbegin[w]\n", wchmm->winfo->num * sizeof(int));
    }
    jlog("STAT: %9d bytes: wchmm->wordend[w]\n", wchmm->winfo->num * sizeof(int));
    jlog("STAT: %9d bytes: wchmm->startnode[]\n", wchmm->startnum * sizeof(int));
    if (wchmm->category_tree) {
      jlog("STAT: %9d bytes: wchmm->start2wid[]\n", wchmm->startnum * sizeof(WORD_ID));
    }
#ifdef UNIGRAM_FACTORING
    if (wchmm->lmtype == LM_PROB) {
      jlog("STAT: %9d bytes: wchmm->start2isolate[]\n", wchmm->isolatenum * sizeof(int));
    }
#endif
    if (!wchmm->hmminfo->multipath) {
      jlog("STAT: %9d bytes: wchmm->wordend_a[]\n", wchmm->winfo->num * sizeof(LOGPROB));
    }
#ifdef PASS1_IWCD
    jlog("STAT: %9d bytes: wchmm->outstyle[]\n", wchmm->n * sizeof(unsigned char));
    {