ngram_decode.c

julius version 4.12.about sound recognition.

 *
 * This function builds a list of next word candidates by looking up
 * the word trellis at specified frame, with lookup_range frame margin.
 * If the same words exists in the near frames, only the one nearest to the
 * specified frame will be chosen.
 * 
 * @param r [in] recognition process instance
 * @param nw [out] pointer to hold the extracted words as list of next word candidates
 * @param hypo [in] partial sentence hypothesis from which the words will be expanded
 * @param tm [in] center time frame to look up the words
 * @param t_end [in] right frame boundary for the lookup.
 * 
 * @return the number of next words candidates stored in @a nw.
 * </EN>
 */
static int
get_backtrellis_words(RecogProcess *r, NEXTWORD **nw, NODE *hypo, short tm, short t_end)
{
  int num = 0;
  int t, t_step;
  int oldnum=0;

  BACKTRELLIS *bt;
  int lookup_range;

  if (tm < 0) return(0);

  bt = r->backtrellis;
  lookup_range = r->config->pass2.lookup_range;

#ifdef PREFER_CENTER_ON_TRELLIS_LOOKUP
  /* fix for 3.2 (01/10/18 by ri) */
  /* before and after (one near center frame has high priority) */
  for (t_step = 0; t_step < lookup_range; t_step++) {
    /* before or center */
    t = tm - t_step;
    if (t < 0 || t > bt->framelen - 1 || t >= t_end) continue;
    num = pick_backtrellis_words(r, nw, oldnum, hypo, t);
    if (num > oldnum) {
      qsort(nw, num, sizeof(NEXTWORD *),
	    (int (*)(const void *,const void *))compare_nw);
      oldnum = num;
    }
    if (t_step == 0) continue;	/* center */
    /* after */
    t = tm + t_step;
    if (t < 0 || t > bt->framelen - 1 || t >= t_end) continue;
    num = pick_backtrellis_words(r, nw, oldnum, hypo, t);
    if (num > oldnum) {
      qsort(nw, num, sizeof(NEXTWORD *),
	    (int (*)(const void *,const void *))compare_nw);
      oldnum = num;
    }
  }

#else

  /* before the center frame */
  for(t = tm; t >= tm - lookup_range; t--) {
    if (t < 0) break;
    num = pick_backtrellis_words(r, nw, oldnum, hypo, t);
    if (num > oldnum) {
      qsort(nw, num, sizeof(NEXTWORD *),
	    (int (*)(const void *,const void *))compare_nw);
      oldnum = num;
    }
  }
  /* after the center frame */
  for(t = tm + 1; t < tm + lookup_range; t++) {
    if (t > bt->framelen - 1) break;
    if (t >= t_end) break;
    num = pick_backtrellis_words(r, nw, oldnum, hypo, t);
    if (num > oldnum) {
      qsort(nw, num, sizeof(NEXTWORD *),
	    (int (*)(const void *,const void *))compare_nw);
      oldnum = num;
    }
  }
#endif

  return num;
}

/** 
 * <JA>
 * @brief  润鸥倡帽胳を近殿. 
 * 
 * 扩腆により鸥倡滦据とならない帽胳をリストから久殿する. 
 * 
 * @param nw [i/o] 肌帽胳礁圭∈礁圭面の鸥倡できない帽胳が久殿される∷
 * @param hypo [in] 鸥倡傅の婶尸矢簿棱
 * @param num [in] @a nw に附哼呈羌されている帽胳眶
 * @param winfo [in] 帽胳辑今
 * 
 * @return 糠たに nw に崔まれる肌帽胳眶
 * </JA>
 * <EN>
 * @brief  Remove non-expansion word from list.
 * 
 * Remove words in the nextword list which should not be expanded.
 * 
 * @param nw [i/o] list of next word candidates (will be shrinked by removing some words)
 * @param hypo [in] partial sentence hypothesis from which the words will be expanded
 * @param num [in] current number of next words in @a nw
 * @param winfo [in] word dictionary
 * 
 * @return the new number of words in @a nw
 * </EN>
 */
static int
limit_nw(NEXTWORD **nw, NODE *hypo, int num, WORD_INFO *winfo)
{
  int src,dst;
  int newnum;

  /* <s>からは部も鸥倡しない */
  /* no hypothesis will be generated after "<s>" */
  if (hypo->seq[hypo->seqnum-1] == winfo->head_silwid) {
    return(0);
  }

  dst = 0;
  for (src=0; src<num; src++) {
    if (nw[src]->id == winfo->tail_silwid) {
      /* </s> は鸥倡しない */
      /* do not expand </s> (it only appears at start) */
      continue;
    }
#ifdef FIX_35_INHIBIT_SAME_WORD_EXPANSION
    /* 木涟帽胳と票じトレリス帽胳は鸥倡しない */
    /* inhibit expanding the exactly the same trellis word twice */
    if (nw[src]->tre == hypo->tre) continue;
#endif
    
    if (src != dst) memcpy(nw[dst], nw[src], sizeof(NEXTWORD));
    dst++;
  }
  newnum = dst;

  return newnum;
}
	

/** 
 * <JA>
 * @brief  介袋帽胳簿棱礁圭を滇める. 
 *
 * N-gramベ〖スの玫瑚では·介袋簿棱は帽胳琐萨の痰不帽胳に盖年されている. 
 * ただし·ショ〖トポ〖ズセグメンテ〖ション箕は·妈1パスで呵姜フレ〖ムに姜眉が
 * 荒った帽胳の面で锑刨呵络の帽胳となる. 
 * 
 * @param nw [out] 肌帽胳铬输リスト∈评られた介袋帽胳簿棱を呈羌する∷
 * @param peseqlen [in] 掐蜗フレ〖ム墓
 * @param maxnw [in] @a nw に呈羌できる帽胳の呵络眶
 * @param r [in] 千急借妄インスタンス
 * 
 * @return @a nw に呈羌された帽胳铬输眶を手す. 
 * </JA>
 * <EN>
 * @brief  Get initial word hypotheses at the beginning.
 *
 * on N-gram based recogntion, the initial hypothesis is fixed to the tail
 * silence word.  Exception is that, in short-pause segmentation mode, the
 * initial hypothesis will be chosen from survived words on the last input
 * frame in the first pass.
 * 
 * @param nw [out] pointer to hold the initial word candidates
 * @param peseqlen [in] input frame length
 * @param maxnw [in] maximum number of words that can be stored in @a nw
 * @param r [in] recognition process instance
 * 
 * @return the number of words extracted and stored to @a nw.
 * </EN>
 *
 * @callgraph
 * @callergraph
 */
int
ngram_firstwords(NEXTWORD **nw, int peseqlen, int maxnw, RecogProcess *r)
{

  if (r->config->successive.enabled) {
    /* in sp segment mode  */
    if (r->sp_break_2_begin_word != WORD_INVALID) {
      /* 介袋簿棱は 呵姜フレ〖ムに荒った帽胳トレリス惧の呵锑帽胳 */
      /* the initial hypothesis is the best word survived on the last frame of
	 the segment */
      nw[0]->id = r->sp_break_2_begin_word;
    } else {
      /* 呵姜セグメント: 介袋簿棱は 帽胳の琐萨の痰不帽胳(=winfo->tail_silwid) */
      /* we are in the last of sentence: initial hypothesis is word-end silence word */
      nw[0]->id = r->lm->winfo->tail_silwid;
    }
  } else {
    /* initial hypothesis should be word-end silence word */
    nw[0]->id = r->lm->winfo->tail_silwid;
  }

#ifdef FIX_PENALTY
  nw[0]->lscore = 0.0;
#else
  nw[0]->lscore = r->config->lmp.lm_penalty2;
#endif

  return 1;			/* number of words = 1 */
}

/** 
 * <JA>
 * @brief 肌帽胳簿棱礁圭を手す. 
 *
 * 涂えられた婶尸矢簿棱から·肌に儡鲁しうる帽胳の礁圭を手す. 悸狠には·
 * 妈1パスの冯蔡であるトレリス帽胳礁圭 bt 惧で·鸥倡傅の婶尸矢簿棱の呵姜帽胳の
 * ∈夸年された∷幌眉フレ〖ム hypo->estimated_next_t の涟稿に赂哼する
 * 帽胳礁圭を艰叫し·それらの N-gram 儡鲁澄唯を纷换して手す. 
 * 艰り叫された肌帽胳簿棱は·あらかじめ maxnm の墓さだけ
 * 挝拌が澄瘦されている nw に呈羌される. 
 * 
 * @param hypo [in] 鸥倡傅の矢簿棱
 * @param nw [out] 肌帽胳铬输リストを呈羌する挝拌へのポインタ
 * @param maxnw [in] @a nw の呵络墓
 * @param r [in] 千急借妄インスタンス
 * 
 * @return 藐叫され nw に呈羌された肌帽胳簿棱の眶を手す. 
 * </JA>
 * <EN>
 * @brief  Return the list of next word candidate.
 *
 * Given a partial sentence hypothesis "hypo", it returns the list of
 * next word candidates.  Actually, it extracts from word trellis the
 * list of words whose word-end node has survived near the estimated
 * beginning-of-word frame of last word "hypo->estimated_next_t", and store
 * them to "nw" with their N-gram probabilities. 
 * 
 * @param hypo [in] source partial sentence hypothesis
 * @param nw [out] pointer to store the list of next word candidates (should be already allocated)
 * @param maxnw [in] maximum number of words that can be stored to @a nw
 * @param r [in] recognition process instance
 * 
 * @return the number of extracted next word candidates in @a nw.
 * </EN>
 * @callgraph
 * @callergraph
 */
int
ngram_nextwords(NODE *hypo, NEXTWORD **nw, int maxnw, RecogProcess *r)
{
  int num, num2;

  if (hypo->seqnum == 0) {
    j_internal_error("ngram_nextwords: hypo contains no word\n");
  }

  /* 簿棱の夸年姜眉箕癸において backtrellis柒に荒っている帽胳を评る */
  /* get survived words on backtrellis at the estimated end frame */
  num = get_backtrellis_words(r, nw, hypo, hypo->estimated_next_t, hypo->bestt);

  /* 鸥倡できない帽胳をチェックして嘲す */
  /* exclude unallowed words */
  num2 = limit_nw(nw, hypo, num, r->lm->winfo);

  if (debug2_flag) jlog("DEBUG: ngram_decode: %d-%d=%d unfolded\n",num, num-num2,num2);

  return(num2);
}

/** 
 * <JA>
 * @brief  减妄冉年
 * 
 * 涂えられた婶尸矢簿棱が·矢∈すなわち玫瑚姜位∷として
 * 减妄材墙であるかどうかを手す. N-gram では矢片に滦炳する痰不帽胳
 * (silhead) であれば减妄する. 
 * 
 * @param hypo [in] 婶尸矢簿棱
 * @param r [in] 千急借妄インスタンス
 * 
 * @return 矢として减妄材墙であれば TRUE·稍材墙なら FALSE を手す. 
 * </JA>
 * <EN>
 * @brief  Acceptance check.
 * 
 * Return whether the given partial hypothesis is acceptable as a sentence
 * and can be treated as a final search candidate.  In N-gram mode, it checks
 * whether the last word is the beginning-of-sentence silence (silhead).
 * 
 * @param hypo [in] partial sentence hypothesis to be examined
 * @param r [in] recognition process instance
 * 
 * @return TRUE if acceptable as a sentence, or FALSE if not.
 * </EN>
 * @callgraph
 * @callergraph
 */
boolean
ngram_acceptable(NODE *hypo, RecogProcess *r)
{

  if (r->config->successive.enabled) {
    /* 呵稿の簿棱が妈１パス呵锑簿棱の呵介の帽胳と办米しなければならない */
    /* the last word should be equal to the first word on the best hypothesis on 1st pass */
    if (hypo->seq[hypo->seqnum-1] == r->sp_break_2_end_word) {
      return TRUE;
    }
  } else {
    /* 呵稿の簿棱が矢片痰不帽胳でなければならない */
    /* the last word should be head silence word */
    if (hypo->seq[hypo->seqnum-1] == r->lm->winfo->head_silwid) {
      return TRUE;
    }
  }
  return FALSE;
}

/* end of file */