search_bestfirst_main.c

来自「julius version 4.12.about sound recognit」· C语言 代码 · 共 2,134 行 · 第 1/5 页

 * @param framenum [in] 掐蜗フレ〖ム墓
 * </JA>
 * <EN>
 * Update the score envelope using forward score of the given hypothesis.
 * 
 * @param s [i/o] work area for 2nd pass
 * @param n [in] hypothesis
 * @param framenum [in] input frame length
 * </EN>
 */
static void
envl_update(StackDecode *s, NODE *n, int framenum)
{
  int t;
  for(t=framenum-1;t>=0;t--) {
    if (s->framemaxscore[t] < n->g[t]) s->framemaxscore[t] = n->g[t];
  }
}
#endif /* SCAN_BEAM */


/**********************************************************************/
/********** Short pause segmentation **********************************/
/**********************************************************************/

/** 
 * <JA>
 * 千急冯蔡から·肌の掐蜗惰粗の千急を倡幌する狠の介袋帽胳旺悟をセットする. 
 * 譬册胳および簿棱の脚剩を雇胃して介袋帽胳旺悟が疯年される. 
 * 
 * @param hypo [in] 附哼の掐蜗惰粗の千急冯蔡としての矢铬输
 * @param r [in] 千急借妄インスタンス
 * </JA>
 * <EN>
 * Set the previous word context for the recognition of the next input
 * segment from the current recognition result.  The initial context word
 * will be chosen from the current recognition result skipping transparent
 * word and multiplied words.
 * 
 * @param hypo [in] sentence candidate as a recognition result of current
 * input segment
 * @param r [in] recognition process instance
 * </EN>
 *
 * @callgraph
 * @callergraph
 * 
 */
void
segment_set_last_nword(NODE *hypo, RecogProcess *r)
{
  int i;
  WORD_ID w;

  if (r->sp_break_last_nword_allow_override) {
    for(i=0;i<hypo->seqnum;i++) {
      w = hypo->seq[i];
      if (w != r->sp_break_last_word
	  && !is_sil(w, r)
	  && !r->lm->winfo->is_transparent[w]
	  ) {
	r->sp_break_last_nword = w;
	break;
      }
    }
#ifdef SP_BREAK_DEBUG
    printf("sp_break_last_nword=%d[%s]\n", r->sp_break_last_nword, r->lm->winfo->woutput[r->sp_break_last_nword]);
#endif
  } else {
    r->sp_break_last_nword = WORD_INVALID;
  }
}


/**********************************************************************/
/********* Debug output of hypothesis while search ********************/
/**********************************************************************/

/** 
 * <JA>
 * デバッグ脱に簿棱の帽胳误を山绩する. 
 * 
 * @param hypo [in] 簿棱
 * @param winfo [in] 帽胳辑今
 * </JA>
 * <EN>
 * Output word sequence of a hypothesis for debug.
 * 
 * @param hypo [in] hypothesis
 * @param winfo [in] word dictionary
 * </EN>
 */
static void
put_hypo_woutput(NODE *hypo, WORD_INFO *winfo)
{
  int i,w;

  if (hypo != NULL) {
    for (i=hypo->seqnum-1;i>=0;i--) {
      w = hypo->seq[i];
      jlog(" %s", winfo->woutput[w]);
    }
  }
  jlog("\n");  
}

/** 
 * <JA>
 * デバッグ脱に簿棱の帽胳N-gramエントリ叹∈Julianではカテゴリ戎规∷を叫蜗する. 
 * 
 * @param hypo [in] 簿棱
 * @param winfo [in] 帽胳辑今
 * </JA>
 * <EN>
 * Output N-gram entries (or DFA category IDs) of a hypothesis for debug.
 * 
 * @param hypo [in] hypothesis
 * @param winfo [in] word dictionary
 * </EN>
 */
static void
put_hypo_wname(NODE *hypo, WORD_INFO *winfo)
{
  int i,w;

  if (hypo != NULL) {
    for (i=hypo->seqnum-1;i>=0;i--) {
      w = hypo->seq[i];
      jlog(" %s", winfo->wname[w]);
    }
  }
  jlog("\n");  
}

/** 
 * <EN>
 * Save a hypothesis as a recognition result f 2nd pass.
 * </EN>
 * <JA>
 * 妈2パスの冯蔡として簿棱を瘦赂する. 
 * </JA>
 * 
 * @param hypo [in] hypothesis to save
 * @param r [in] recognition process instance
 * 
 */
static void
store_result_pass2(NODE *hypo, RecogProcess *r)
{
  int i;
  Sentence *s;

  s = &(r->result.sent[r->result.sentnum]);

  s->word_num = hypo->seqnum;
  for (i = 0; i < hypo->seqnum; i++) {
    s->word[i] = hypo->seq[hypo->seqnum - 1 - i];
  }
#ifdef CONFIDENCE_MEASURE
  for (i = 0; i < hypo->seqnum; i++) {
    s->confidence[i] = hypo->cmscore[hypo->seqnum - 1 - i];
  }
#endif

  s->score = hypo->score;
  s->score_lm = hypo->totallscore;
  s->score_am = hypo->score - hypo->totallscore;
  if (r->lmtype == LM_DFA) {
    /* output which grammar the hypothesis belongs to on multiple grammar */
    /* determine only by the last word */
    if (multigram_get_all_num(r->lm) > 0) {
      s->gram_id = multigram_get_gram_from_category(r->lm->winfo->wton[hypo->seq[0]], r->lm);
    } else {
      s->gram_id = 0;
    }
  }

  r->result.sentnum++;
  /* add to tail */
}


/**********************************************************************/
/******** Output top 'ncan' hypotheses in a stack and free all ********/
/**********************************************************************/

/** 
 * <JA>
 * スタックから惧疤の簿棱を艰り叫し·千急冯蔡として叫蜗する. さらに·
 * スタックに呈羌されている链ての簿棱を豺庶する. 
 *
 * 评られた矢铬输は·いったん冯蔡呈羌脱のスタックに呈羌される. 玫瑚姜
 * 位∈"-n" の眶だけ矢铬输が斧つかるか·玫瑚が面们される∷の稿·冯蔡弄
 * に评られた矢铬输の面から惧疤N改∈"-output" で回年された眶∷の簿棱を
 * 叫蜗する.
 *
 * 回年があればアラインメントもここで悸乖する. 
 * 
 * @param r_start [i/o] 冯蔡呈羌脱スタックの黎片ノ〖ドへのポインタ
 * @param r_bottom [i/o] 冯蔡呈羌脱スタックの撵ノ〖ドへのポインタ
 * @param r_stacknum [i/o] スタックに呈羌されているノ〖ド眶へのポインタ
 * @param ncan [in] 叫蜗する惧疤簿棱眶
 * @param r [in] 千急借妄インスタンス
 * @param param [in] 掐蜗パラメ〖タ
 * </JA>
 * <EN>
 * Output top N-best hypotheses in a stack as a recognition result, and
 * free all hypotheses.
 *
 * The sentence candidates found at the 2nd pass will be pushed to the
 * "result stack" instead of immediate output.  After recognition is
 * over (in case the number of found sentences reaches the number
 * specified by "-n", or search has been terminated in other reason),
 * the top N sentence candidates in the stack will be output as a
 * final result (where N should be specified by "-output").  After
 * then, all the hypotheses in the stack will be freed.
 *
 * Additionally, forced alignment for the recognized sentence
 * will be executed here if required.
 * 
 * @param r_start [i/o] pointer to the top node of the result stack
 * @param r_bottom [i/o] pointer to the bottom node of the result stack
 * @param r_stacknum [i/o] number of candidates in the current result stack
 * @param ncan [in] number of sentence candidates to be output
 * @param r [in] recognition process instance
 * @param param [in] input parameter
 * </EN>
 */
static void
result_reorder_and_output(NODE **r_start, NODE **r_bottom, int *r_stacknum, int ncan, RecogProcess *r, HTK_Param *param)
{
  NODE *now;
  int num;

#ifdef CM_NBEST 
  /* compute CM from the N-best sentence candidates */
  cm_compute_from_nbest(&(r->pass2), *r_start, *r_stacknum, r->config);
#endif
  num = 0;
  while ((now = get_best_from_stack(r_start,r_stacknum)) != NULL && num < ncan) {
    num++;
    /* output result */
    store_result_pass2(now, r);

    /* if in sp segmentation mode, */
    /* set the last context-aware word for next recognition */
    if (r->lmtype == LM_PROB && r->config->successive.enabled && num == 1) segment_set_last_nword(now, r);

    free_node(now);
  }
  /* free the rest */
  if (now != NULL) free_node(now);
  free_all_nodes(*r_start);
}  

/** 
 * <EN>
 * @brief  Post-process of 2nd pass when no result is obtained.
 *
 * This is a post-process for the 2nd pass which should be called when
 * the 2nd pass has no result.  This will occur when the 2nd pass was
 * executed but failed with no sentence candidate, or skipped by
 * an option.
 *
 * When the 2nd argument is set to TRUE, the result of the 1st pass
 * will be copied as final result of 2nd pass and the recognition
 * status flag is set to SUCCESS.  If FALSE, recognition status will
 * be set to FAILED.  On sp-segment decoding, the initial hypothesis
 * marker for the next input segment will be set up from the 1st pass
 * result also.
 * 
 * @param r [in] recognition process instance
 * @param use_1pass_as_final [in] when TRUE the 1st pass result will be used as final recognition result of 2nd pass.
 * 
 * </EN>
 * <JA>
 * @brief  妈2パスの豺が评られない眷圭の姜位借妄
 *
 * 妈2パスが己窃した眷圭や妈2パスが悸乖されない肋年の眷圭の
 * 千急姜位借妄を乖うˉuse_1pass_as_final が TRUE のとき·
 * 妈1パスの冯蔡を妈2パスの冯蔡としてコピ〖して呈羌し·千急喇根とするˉ
 * FALSE箕は千急己窃とするˉ
 * また·sp-segment 箕は·肌の千急惰粗脱の介袋簿棱肋年も妈1パスの
 * 冯蔡から乖うˉ
 * 
 * @param r [in] 千急借妄インスタンス
 * @param use_1pass_as_final [in] TRUE 箕妈1パスの冯蔡を妈2パス冯蔡に呈羌する
 * 
 * </JA>
 */
void
pass2_finalize_on_no_result(RecogProcess *r, boolean use_1pass_as_final)
{
  NODE *now;
  int i, j;

  /* 玫瑚己窃 */
  /* search failed */

  /* make temporal hypothesis data from the result of previous 1st pass */
  now = newnode(r);
  for (i=0;i<r->pass1_wnum;i++) {
    now->seq[i] = r->pass1_wseq[r->pass1_wnum-1-i];
  }
  now->seqnum = r->pass1_wnum;
  now->score = r->pass1_score;
#ifdef CONFIDENCE_MEASURE
  /* fill in null values */
#ifdef CM_MULTIPLE_ALPHA
  for(j=0;j<jconf->annotate.cm_alpha_num;j++) {
    for(i=0;i<now->seqnum;i++) now->cmscore[i][j] = 0.0;
  }
#else
  for(i=0;i<now->seqnum;i++) now->cmscore[i] = 0.0;
#endif
#endif /* CONFIDENCE_MEASURE */
  
  if (r->lmtype == LM_PROB && r->config->successive.enabled) {
    /* if in sp segment mode, */
    /* find segment restart words from 1st pass result */
    segment_set_last_nword(now, r);
  }
    
  if (use_1pass_as_final) {
    /* 妈1パスの冯蔡をそのまま叫蜗する */
    /* output the result of the previous 1st pass as a final result. */
    store_result_pass2(now, r);
    r->result.status = J_RESULT_STATUS_SUCCESS;
  } else {
    /* store output as failure */
    r->result.status = J_RESULT_STATUS_FAIL;
    //callback_exec(CALLBACK_RESULT, r);
  }
  
  free_node(now);
}


/**********************************************************************/
/********* Main stack decoding function *******************************/
/**********************************************************************/

/** 
 * <JA>
 * 妈2玫瑚パスであるスタックデコ〖ディングを乖うメイン簇眶
 *
 * 苞眶のうち cate_bgn, cate_num は帽胳N-gramでは痰浑される. 
 * 
 * @param param [in] 掐蜗パラメ〖タベクトル误
 * @param r [i/o] 千急借妄インスタンス
 * @param cate_bgn [in] 鸥倡滦据とすべきカテゴリの倡幌戎规
 * @param cate_num [in] 鸥倡滦据とすべきカテゴリの眶
 * </JA>
 * <EN>
 * Main function to perform stack decoding of the 2nd search pass.
 *
 * The cate_bgn and cate_num (third and fourth argument) will have
 * no effect when N-gram is used.
 * 
 * @param param [in] input parameter vector
 * @param r [i/o] recognition process instance
 * @param cate_bgn [in] category id to allow word expansion from (ignored in Julius)
 * @param cate_num [in] num of category to allow word expansion from (ignored in Julius)
 * </EN>
 *
 * @callgraph
 * @callergraph
 * 
 */
void
wchmm_fbs(HTK_Param *param, RecogProcess *r, int cate_bgn, int cate_num)
{
  /* 矢簿棱スタック */
  /* hypothesis stack (double-linked list) */
  int stacknum;			/* current stack size */
  NODE *start = NULL;		/* top node */
  NODE *bottom = NULL;		/* bottom node */

  /* 千急冯蔡呈羌スタック(冯蔡はここへいったん礁められる) */
  /* result sentence stack (found results will be stored here and then re-ordered) */
  int r_stacksize;
  int r_stacknum;
  NODE *r_start = NULL;
  NODE *r_bottom = NULL;

  /* ワ〖クエリア */
  /* work area */
  NEXTWORD fornoise; /* Dummy NEXTWORD data for short-pause insertion handling */

  NEXTWORD **nextword, *nwroot;	/* buffer to store predicted words */
  int maxnwnum;			/* current allocated number of words in nextword */
  int nwnum;			/* current number of words in nextword */
  NODE *now, *new;		/* popped current hypo., expanded new hypo. */
  NODE *now_noise;	       /* for inserting/deleting noise word */
  boolean now_noise_calced;
  boolean acc;
  int t;
  int w,i,j;
  LOGPROB last_score = LOG_ZERO;
  /* for graph generation */
  LOGPROB prev_score;
  WordGraph *wordgraph_root = NULL;
  boolean merged_p;
#ifdef GRAPHOUT_DYNAMIC
  int dynamic_merged_num = 0;
  WordGraph *wtmp;
  LOGPROB lscore_prev;
#endif
#ifdef GRAPHOUT_SEARCH
  int terminate_search_num = 0;
#endif

  /* local temporal parameter */
  int stacksize, ncan, maxhypo, peseqlen;
  JCONF_SEARCH *jconf;
  WORD_INFO *winfo;
  NGRAM_INFO *ngram;
  DFA_INFO *gdfa;
  BACKTRELLIS *backtrellis;
  StackDecode *dwrk;

  if (r->lmtype == LM_DFA) {
    if (debug2_flag) jlog("DEBUG: only words in these categories will be expanded: %d-%d\n", cate_bgn, cate_bgn + cate_num-1);
  }

  /*
   * 介袋步