beam.c

julius version 4.12.about sound recognition.

	/* 矢姜位帽胳(silE,剁爬(IPAモデル))なら·silB で倡幌 */
	/* initial word = best last word hypothesis on the last segment */
	/* if silE or sp, begin with silB */
	/*if (is_sil(recog.sp_break_last_word, wchmm->winfo, wchmm->hmminfo)) {*/
	if (r->sp_break_last_word == wchmm->winfo->tail_silwid) {
	  beginword = wchmm->winfo->head_silwid;
	  d->bos.wid = WORD_INVALID;	/* reset initial word context */
	} else {
	  beginword = r->sp_break_last_word;
	}
      } else {
	/* initial segment: initial word set to silB */
	beginword = wchmm->winfo->head_silwid;
      }
    } else {			/* not sp segment mode */
      /* initial word fixed to silB */
      beginword = wchmm->winfo->head_silwid;
    }

#ifdef SP_BREAK_DEBUG
    jlog("DEBUG: startword=[%s], last_nword=[%s]\n",
	   (beginword == WORD_INVALID) ? "WORD_INVALID" : wchmm->winfo->wname[beginword],
	   (d->bos.wid == WORD_INVALID) ? "WORD_INVALID" : wchmm->winfo->wname[d->bos.wid]);
#endif
    /* create the first token at the first node of initial word */
    newid = create_token(d);
    new = &(d->tlist[d->tn][newid]);

    /* initial node = head node of the beginword */
    if (wchmm->hmminfo->multipath) {
      node = wchmm->wordbegin[beginword];
    } else {
      node = wchmm->offset[beginword][0];
    }

    /* set initial LM score */
    if (wchmm->state[node].scid != 0) {
      /* if initial node is on a factoring branch, use the factored score */
      new->last_lscore = max_successor_prob(wchmm, d->bos.wid, node);
    } else {
      /* else, set to 0.0 */
      new->last_lscore = 0.0;
    }
#ifdef FIX_PENALTY
    new->last_lscore = new->last_lscore * d->lm_weight;
#else
    new->last_lscore = new->last_lscore * d->lm_weight + d->lm_penalty;
#endif
    /* set initial word history */
    new->last_tre = &(d->bos);
    new->last_cword = d->bos.wid;
    if (wchmm->hmminfo->multipath) {
      /* set initial score using the initial LM score */
      new->score = new->last_lscore;
    } else {
      /* set initial score using the initial LM score and AM score of the first state */
      new->score = outprob_style(wchmm, node, d->bos.wid, 0, param) + new->last_lscore;
    }
    /* assign the initial node to token list */
    node_assign_token(d, node, newid);

  }

  if (r->lmtype == LM_DFA && r->lmvar == LM_DFA_GRAMMAR) {
  
    /* 介袋簿棱: 矢恕惧矢片に儡鲁しうる帽胳礁圭 */
    /* initial words: all words that can be begin of sentence grammatically */
    /* アクティブな矢恕に掳する帽胳のみ钓す */
    /* only words in active grammars are allowed to be an initial words */
    MULTIGRAM *m;
    int t,tb,te;
    WORD_ID iw;
    boolean flag;
    DFA_INFO *gdfa;

    gdfa = r->lm->dfa;

    flag = FALSE;
    /* for all active grammar */
    for(m = r->lm->grammars; m; m = m->next) {
      if (m->active) {
	tb = m->cate_begin;
	te = tb + m->dfa->term_num;
	for(t=tb;t<te;t++) {
	  /* for all word categories that belong to the grammar */
	  if (dfa_cp_begin(gdfa, t) == TRUE) {
	    /* if the category can appear at beginning of sentence, */
	    flag = TRUE;
	    for (iw=0;iw<gdfa->term.wnum[t];iw++) {
	      /* create the initial token at the first node of all words that belongs to the category */
	      i = gdfa->term.tw[t][iw];
	      if (wchmm->hmminfo->multipath) {
		node = wchmm->wordbegin[i];
	      } else {
		node = wchmm->offset[i][0];
	      }
	      /* in tree lexicon, words in the same category may share the same root node, so skip it if the node has already existed */
	      if (node_exist_token(d, d->tn, node, d->bos.wid) != TOKENID_UNDEFINED) continue;
	      newid = create_token(d);
	      new = &(d->tlist[d->tn][newid]);
	      new->last_tre = &(d->bos);
	      new->last_lscore = 0.0;
	      if (wchmm->hmminfo->multipath) {
		new->score = 0.0;
	      } else {
		new->score = outprob_style(wchmm, node, d->bos.wid, 0, param);
	      }
	      node_assign_token(d, node, newid);
	    }
	  }
	}
      }
    }
    if (!flag) {
      jlog("ERROR: init_nodescore: no initial state found in active DFA grammar\n");
      return FALSE;
    }
  }

  if (r->lmtype == LM_DFA && r->lmvar == LM_DFA_WORD) {
    /* アクティブな矢恕に掳する帽胳のみ钓す */
    /* only words in active grammars are allowed to be an initial words */
    MULTIGRAM *m;

    for(m = r->lm->grammars; m; m = m->next) {
      if (m->active) {
	for(i = m->word_begin; i < m->word_begin + m->winfo->num; i++) {
	  if (wchmm->hmminfo->multipath) {
	    node = wchmm->wordbegin[i];
	  } else {
	    node = wchmm->offset[i][0];
	  }
	  if (node_exist_token(d, d->tn, node, d->bos.wid) != TOKENID_UNDEFINED) continue;
	  newid = create_token(d);
	  new = &(d->tlist[d->tn][newid]);
	  new->last_tre = &(d->bos);
	  new->last_lscore = 0.0;
	  if (wchmm->hmminfo->multipath) {
	    new->score = 0.0;
	  } else {
	    new->score = outprob_style(wchmm, node, d->bos.wid, 0, param);
	  }
	  node_assign_token(d, node, newid);
	}
      }
    }
  }

  return TRUE;

}

/******************************************************/
/* フレ〖ム票袋ビ〖ム玫瑚の悸乖 --- 呵介のフレ〖ム脱  */
/* frame synchronous beam search --- first frame only */
/******************************************************/

/** 
 * <JA>
 * @brief  フレ〖ム票袋ビ〖ム玫瑚の介袋步
 *
 * ここではビ〖ムサ〖チに脱いるワ〖クエリアの澄瘦と介袋步を乖う. 
 * 介袋步棱の栏喇は init_nodescore() で乖われる. 
 * 
 * @param param [in] 掐蜗ベクトル误攫鼠 (呵介の１フレ〖ム誊のみ脱いられる)
 * @param r [i/o] 不兰千急借妄インスタンス
 * </JA>
 * <EN>
 * @brief  Initialization of the frame synchronous beam search
 *
 * This function will initialize work area for the 1st pass.
 * Generation of initial hypotheses will be performed in init_nodescore().
 * 
 * @param param [in] input vectors (only the first frame will be used)
 * @param r [i/o] recognition process instance
 * </EN>
 *
 * @callergraph
 * @callgraph
 * 
 */
boolean
get_back_trellis_init(HTK_Param *param,	RecogProcess *r)
{
  WCHMM_INFO *wchmm;
  BACKTRELLIS *backtrellis;
  FSBeam *d;

  wchmm = r->wchmm;
  backtrellis = r->backtrellis;
  d = &(r->pass1);

  /* Viterbi遍换脱ワ〖クエリアのスイッチャ〖 tl,tn の介袋猛肋年 */
  /* tn: このフレ〖ム脱ID   tl: １フレ〖ム涟のID */
  /* initialize switch tl, tn for Viterbi computation */
  /* tn: this frame  tl: last frame */
  d->tn = 0;
  d->tl = 1;

  /* 冯蔡の帽胳トレリスを呈羌するバックトレリス菇陇挛を介袋步 */
  /* initialize backtrellis structure to store resulting word trellis */
  bt_prepare(backtrellis);

  /* 纷换脱ワ〖クエリアを介袋步 */
  /* initialize some data on work area */

  if (r->lmtype == LM_PROB) {
    d->lm_weight  = r->config->lmp.lm_weight;
    d->lm_penalty = r->config->lmp.lm_penalty;
  }
  if (r->lmtype == LM_DFA) {
    d->penalty1 = r->config->lmp.penalty1;
  }
#if defined(WPAIR) && defined(WPAIR_KEEP_NLIMIT)
  d->wpair_keep_nlimit = r->config->pass1.wpair_keep_nlimit;
#endif

  /* ワ〖クエリアを澄瘦 */
  /* malloc work area */
  /* 蝗脱するト〖クン翁 = viterbi箕に莲败黎となる觉轮铬输の眶
   * 徒卢: ビ〖ム眶 x 2 (极甘莲败+肌觉轮) + 腾菇陇步辑今のル〖トノ〖ド眶
   */
  /* assumed initial number of needed tokens: beam width x 2 (self + next trans.)
   * + root node on the tree lexicon (for inter-word trans.)
   */
  if (d->totalnodenum != wchmm->n) {
    free_nodes(d);
  }
  if (d->nodes_malloced == FALSE) {
    malloc_nodes(d, wchmm->n, r->trellis_beam_width * 2 + wchmm->startnum);
  }
  prepare_nodes(d, r->trellis_beam_width);
  
  /* 介袋スコアを nodescore[tn] にセット */
  /* set initial score to nodescore[tn] */
  if (init_nodescore(param, r) == FALSE) {
    jlog("ERROR: get_back_trellis_init: failed to set initial node scores\n");
    return FALSE;
  }

  sort_token_no_order(d, r->trellis_beam_width, &(d->n_start), &(d->n_end));

  /* 敛肌叫蜗を乖なう眷圭のインタ〖バルを纷换 */
  /* set interval frame for progout */
  r->config->output.progout_interval_frame = (int)((float)r->config->output.progout_interval / ((float)param->header.wshift / 10000.0));

  if (r->config->successive.enabled) {
    /* ショ〖トポ〖ズセグメンテ〖ション脱パラメ〖タの介袋步 */
    /* initialize parameter for short pause segmentation */
    d->in_sparea = TRUE;		/* assume beginning is silence */
    r->am->mfcc->sparea_start = d->tmp_sparea_start = 0; /* set start frame to 0 */
#ifdef SP_BREAK_RESUME_WORD_BEGIN
    d->tmp_sp_break_last_word = WORD_INVALID;
#endif
    r->sp_break_last_word = WORD_INVALID;
    /* 呵介のセグメント: 肌の润ポ〖ズフレ〖ムで妈2パスへ败乖しない */
    /* the first end of pause segment should be always silB, so
       skip the first segment */
    d->first_sparea = TRUE;
    r->sp_break_2_begin_word = WORD_INVALID;
  }

#ifdef DETERMINE
  if (r->lmvar == LM_DFA_WORD) {
    /* initialize */
    determine_word(r, 0, NULL, 0, 0);
  }
#endif

  return TRUE;
}

/*****************************************************/
/* frame synchronous beam search --- proceed 1 frame */
/* フレ〖ム票袋ビ〖ム玫瑚の悸乖 --- 1フレ〖ム渴める  */
/*****************************************************/

/** 
 * <EN>
 * Propagate a token to next node.
 * 
 * </EN>
 * <JA>
 * ト〖クンを肌ノ〖ドに帕嚷する. 
 * 
 * </JA>
 * 
 * @param d [i/o] work area for 1st pass recognition processing
 * @param next_node [in] next node id
 * @param next_score [in] score when transmitted to the next node
 * @param last_tre [in] previous word context for the next node
 * @param last_cword [in] previous context-valid word for the next node
 * @param last_lscore [in] LM score to be propagated
 * 
 */
static void
propagate_token(FSBeam *d, int next_node, LOGPROB next_score, TRELLIS_ATOM *last_tre, WORD_ID last_cword, LOGPROB last_lscore)
{
  TOKEN2 *tknext;
  TOKENID tknextid;

  if ((tknextid = node_exist_token(d, d->tn, next_node, last_tre->wid)) != TOKENID_UNDEFINED) {
    /* 莲败黎ノ〖ドには贷に戮ノ〖ドから帕嚷貉み: スコアが光いほうを荒す */
    /* the destination node already has a token: compare score */
    tknext = &(d->tlist[d->tn][tknextid]);
    if (tknext->score < next_score) {
      /* その莲败黎ノ〖ドが积つト〖クンの柒推を惧今きする(糠惮ト〖クンは侯らない) */
      /* overwrite the content of existing destination token: not create a new token */
      tknext->last_tre = last_tre; /* propagate last word info */
      tknext->last_cword = last_cword; /* propagate last context word info */
      tknext->last_lscore = last_lscore; /* set new LM score */
      tknext->score = next_score; /* set new score */
    }
  } else {
    /* 莲败黎ノ〖ドは踏帕嚷: 糠惮ト〖クンを侯って充り烧ける */
    /* token unassigned: create new token and assign */
    if (next_score > LOG_ZERO) { /* valid token */
      tknextid = create_token(d); /* get new token */
    }
    tknext = &(d->tlist[d->tn][tknextid]);
    tknext->last_tre = last_tre; /* propagate last word info */
    tknext->last_cword = last_cword; /* propagate last context word info */
    tknext->last_lscore = last_lscore;
    tknext->score = next_score; /* set new score */
    node_assign_token(d, next_node, tknextid); /* assign this new token to the next node */
  }
}

/** 
 * <JA>
 * 帽胳柒のあるノ〖ド粗の莲败を乖う. 
 * 
 * @param wchmm [in] 腾菇陇步辑今
 * @param d [i/o] 妈1パスワ〖クエリア
 * @param tk_ret [i/o] 帕嚷傅のト〖クン∈柒婶でポインタ构糠箕は惧今き∷
 * @param j [in] @a tk_ret の傅のト〖クンリストのID
 * @param next_node [in] 莲败黎のノ〖ド戎规
 * @param next_a [in] 莲败澄唯
 * </JA>
 * <EN>
 * Word-internal transition for a set of nodes.
 * 
 * @param wchmm [in] tree lexicon
 * @param d [i/o] work area for the 1st pass
 * @param tk_ret [in] source token (if pointer updated, overwrite this)
 * @param j [in] the token ID of @a tk_ret
 * @param next_node [in] id of next node
 * @param next_a [in] transition probability
 * 
 * </EN>
 */
static void
beam_intra_word_core(WCHMM_INFO *wchmm, FSBeam *d, TOKEN2 **tk_ret, int j, int next_node, LOGPROB next_a)
{
  int node; ///< Temporal work to hold the current node number on the lexicon tree
  LOGPROB tmpsum;
  LOGPROB ngram_score_cache;
  TOKEN2 *tk;

  tk = *tk_ret;

  node = tk->node;

  /* now, 'node' is the source node, 'next_node' is the destication node,
     and ac-> holds transition probability */
  /* tk->score is the accumulated score at the 'node' on previous frame */
  
  /******************************************************************/
  /* 2.1.1 莲败黎へのスコア纷换(莲败澄唯≤咐胳スコア)               */
  /*       compute score of destination node (transition prob + LM) */
  /******************************************************************/
  tmpsum = tk->score + next_a;
  ngram_score_cache = LOG_ZERO;
  /* the next score at 'next_node' will be computed on 'tmpsum', and
     the new LM probability (if updated) will be stored on 'ngram_score_cache' at below */
  
  if (!wchmm->category_tree) {
    /* 咐胳スコア factoring:
       arcが极甘莲败でない帽胳柒の莲败で·かつ莲败黎にsuccessorリスト
       があれば·lexicon tree の尸呆婶尸の莲败である */
    /* LM factoring:
       If this is not a self transition and destination node has successor
       list, this is branching transition
    */
    if (next_node != node) {
      if (wchmm->state[next_node].scid != 0
#ifdef UNIGRAM_FACTORING
	  /* 1-gram factoring 蝗脱箕は, 剩眶で鼎铜される晦では
	     wchmm->state[node].scid は砷の猛となり·その冷滦猛を
	     藕机として wchmm->fscore[] に帽胳礁圭の1-gramの呵络猛が呈羌
	     されている. 琐眉の晦(剩眶帽胳で鼎铜されない)では·
	     wchmm->state[node].scid は赖の猛となり·
	     １帽胳を sc として积つのでそこで赖澄な2-gramを纷换する */
	  /* When uni-gram factoring,
	     wchmm->state[node].scid is below 0 for shared branches.
	     In this case the maximum uni-gram probability for sub-tree
	     is stored in