search_bestfirst_v2.c

julius version 4.12.about sound recognition.

      wordend_frame_dst[t] = dwrk->wend_token_frame[tn][0];
      wordend_gscore_dst[t] = dwrk->wend_token_gscore[tn][0];
    }
#endif
    /* 回年された least_frame より黎まで t が渴んでおり·かつこの t において
       スコアエンベロ〖プによって栏き荒ったノ〖ドが办つも痰かった眷圭,
       このフレ〖ムで纷换を虑ち磊りそれ笆惧黎([0..t-1])は纷换しない */
    /* if frame 't' already reached the 'least_frame' and no node was
       survived in this frame (all nodes pruned by score envelope),
       terminate computation at this frame and do not computer further
       frame ([0..t-1]). */
    if (t < least_frame && (!node_exist_p)) {
      /* crear the rest scores */
      for (i=t-1;i>=0;i--) {
	g_new[i] = LOG_ZERO;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
	if (r->graphout) {
	  wordend_frame_dst[i] = -1;
	  wordend_gscore_dst[i] = LOG_ZERO;
	}
#endif
      }
      /* terminate loop */
      break;
    }
    
  } /* end of time loop */

  if (hmminfo->multipath) {
    /* 涟羹きスコアの呵姜猛を纷换 (觉轮 0 から箕粗 0 への莲败) */
    /* compute the total forward score (transition from state 0 to frame 0 */
    if (t < 0) {			/* computed till the end */
      tmpmax = LOG_ZERO;
      for(ac=whmm->state[0].ac;ac;ac=ac->next) {
	tmpscore = dwrk->wordtrellis[tn][ac->arc] + ac->a;
	if (tmpmax < tmpscore) tmpmax = tmpscore;
      }
      *final_g = tmpmax;
    } else {
      *final_g = LOG_ZERO;
    }
  }

  /* free work area */
  free_hmm(whmm);
}

/** 
 * <JA>
 * 呵稿の1不燎に滦して Viterbi 纷换を渴める. 
 * 
 * @param now [in] 鸥倡傅の矢簿棱. 办不燎涟の涟羹きスコアが g[] にあるとする. 
 * @param new [out] 纷换稿の涟羹きスコアが g[] に呈羌される. 
 * @param lastphone [in] Viterbi纷换を乖う不燎HMM
 * @param sp [in] short-pause insertion
 * @param param [in] 掐蜗ベクトル误
 * @param r [in] 千急借妄インスタンス
 * </JA>
 * <EN>
 * Proceed Viterbi for the last one phoneme.
 * 
 * @param now [in] source hypothesis where the forward scores prior to the
 * last one phone is stored at g[]
 * @param new [out] the resulting updated forward scores will be saved to g[]
 * @param lastphone [in] phone HMM for the Viterbi processing
 * @param sp [in] short-pause insertion
 * @param param [in] input vectors
 * @param r [in] recognition process instance
 * </EN>
 */
static void
do_viterbi_next_word(NODE *now, NODE *new, HMM_Logical *lastphone, boolean sp, HTK_Param *param, RecogProcess *r) /* sp is for multipath only */
{
  int t, n;
  LOGPROB a_value;		/* for non multi-path */
  int peseqlen;
  boolean multipath;
  StackDecode *dwrk;

  dwrk = &(r->pass2);

  multipath = r->am->hmminfo->multipath;

  peseqlen = r->peseqlen;
  
  if (! multipath) {

    /* もし鸥倡傅簿棱の呵稿の帽胳の不燎墓が 1 であれば·その不燎は
       木涟の scan_word で纷换されていない. この眷圭, now->g[] に笆涟の
       介袋猛が呈羌されている. 
       もし不燎墓が１笆惧であれば·now->g[] はその缄涟まで纷换した觉轮
       のスコアが掐っているので,now->g[t] から介袋猛を肋年する涩妥がある */
    /* If the length of last word is 1, it means the last phone was not
       scanned in the last call of scan_word().  In this case, now->g[]
       keeps the previous initial value, so start viterbi with the old scores.
       If the length is more than 1, the now->g[] keeps the values of the
       scan result till the previous phone, so make initial value
       considering last transition probability. */
    if (r->lm->winfo->wlen[now->seq[now->seqnum-1]] > 1) {
      n = hmm_logical_state_num(lastphone);
      a_value = (hmm_logical_trans(lastphone))->a[n-2][n-1];
      for(t=0; t<peseqlen-1; t++) dwrk->g[t] = now->g[t+1] + a_value;
      dwrk->g[peseqlen-1] = LOG_ZERO;
    } else {
      for(t=0; t<peseqlen; t++) dwrk->g[t] = now->g[t];
    }

  } else {
  
    for(t=0; t<peseqlen; t++) dwrk->g[t] = now->g[t];
    dwrk->phmmseq[0] = lastphone;
    if (r->lm->config->enable_iwsp) dwrk->has_sp[0] = sp;

  }
  
  do_viterbi(dwrk->g, new->g,
	     multipath ? dwrk->phmmseq : &lastphone,
	     (r->lm->config->enable_iwsp && multipath) ? dwrk->has_sp : NULL,
	     1, param, peseqlen, now->estimated_next_t, &(new->final_g)
#ifdef GRAPHOUT_PRECISE_BOUNDARY
	     , now->wordend_frame, new->wordend_frame
	     , now->wordend_gscore, new->wordend_gscore
#else
	     , NULL, NULL
	     , NULL, NULL
#endif
	     , r
	     );

#ifdef GRAPHOUT_PRECISE_BOUNDARY
  if (! multipath) {
    if (r->graphout) {
      /* 肌搀の next_word 脱に董肠攫鼠を拇腊 */
      /* proceed word boundary for one step for next_word */
      new->wordend_frame[r->peseqlen-1] = new->wordend_frame[0];
      new->wordend_gscore[r->peseqlen-1] = new->wordend_gscore[0];
      for (t=0;t<r->peseqlen-1;t++) {
	new->wordend_frame[t] = new->wordend_frame[t+1];
	new->wordend_gscore[t] = new->wordend_gscore[t+1];
      }
    }
  }
#endif
}

/** 
 * <JA>
 * 呵稿の1帽胳の涟羹きトレリスを纷换して·矢簿棱の涟羹き锑刨を构糠する. 
 * 
 * @param now [i/o] 矢簿棱
 * @param param [in] 掐蜗パラメ〖タ误
 * @param r [in] 千急借妄インスタンス
 * </JA>
 * <EN>
 * Compute the forward viterbi for the last word to update forward scores
 * and ready for word connection.
 * 
 * @param now [i/o] hypothesis
 * @param param [in] input parameter vectors
 * @param r [in] recognition process instance
 * </EN>
 * @callgraph
 * @callergraph
 */
void
scan_word(NODE *now, HTK_Param *param, RecogProcess *r)
{
  int   i,t;
  WORD_ID word;
  int phmmlen;
  HMM_Logical *tailph;

  /* store global values to local for rapid access */
  WORD_INFO *winfo;
  HTK_HMM_INFO *hmminfo;
  int peseqlen;
  boolean ccd_flag;
  boolean enable_iwsp;		/* multipath */
  StackDecode *dwrk;

  dwrk = &(r->pass2);
  winfo = r->lm->winfo;
  hmminfo = r->am->hmminfo;
  peseqlen = r->peseqlen;
  ccd_flag = r->ccd_flag;
  if (hmminfo->multipath) {
    enable_iwsp = r->lm->config->enable_iwsp;
  }
  
#ifndef GRAPHOUT_PRECISE_BOUNDARY
  if (r->graphout) {
    if (ccd_flag) {
      now->tail_g_score = now->g[now->bestt];
    }
  }
#endif

  /* ----------------------- prepare phoneme sequence ------------------ */
  /* triphoneなら黎片の1不燎はここでは滦据嘲(あとでnext_wordでやる) */
  /*             琐萨の1不燎はコンテキストにしたがって弥垂 */
  /* with triphone, modify the tail phone of the last word according to the
     previous word, and do not compute the head phone here (that will be
     computed later in next_word() */
  word = now->seq[now->seqnum-1];
  
#ifdef TCD
    jlog("DEBUG: w=");
    for(i=0;i<winfo->wlen[word];i++) {
      jlog(" %s",(winfo->wseq[word][i])->name);
    }
    if (ccd_flag) {
      if (now->last_ph != NULL) {
	jlog(" | %s", (now->last_ph)->name);
      }
    }
    jlog("\n");
#endif /* TCD */
    
  if (ccd_flag) {
    
    /* the tail triphone of the last word varies by context */
    if (now->last_ph != NULL) {
      tailph = get_right_context_HMM(winfo->wseq[word][winfo->wlen[word]-1], now->last_ph->name, hmminfo);
      if (tailph == NULL) {
	/* fallback to the original bi/mono-phone */
	/* error if the original is pseudo phone (not explicitly defined
	   in hmmdefs/hmmlist) */
	/* exception: word with 1 phone (triphone may exist in the next expansion */
	if (winfo->wlen[word] > 1 && winfo->wseq[word][winfo->wlen[word]-1]->is_pseudo){
	  error_missing_right_triphone(winfo->wseq[word][winfo->wlen[word]-1], now->last_ph->name);
	}

	tailph = winfo->wseq[word][winfo->wlen[word]-1];
      }
    } else {
      tailph = winfo->wseq[word][winfo->wlen[word]-1];
    }
    /* 墓さ1の帽胳は肌のnextwordでさらに恃步するのでここではscanしない */
    /* do not scan word if the length is 1, as it further varies in the
       following next_word() */
    if (winfo->wlen[word] == 1) {
      now->last_ph = tailph;
      if (enable_iwsp && hmminfo->multipath) now->last_ph_sp_attached = TRUE;
#ifdef GRAPHOUT_PRECISE_BOUNDARY
      if (r->graphout) {
	/* 帽胳董肠帕嚷攫鼠を介袋步 */
	/* initialize word boundary propagation info */
	for (t=0;t<peseqlen;t++) {
	  now->wordend_frame[t] = t;
	  now->wordend_gscore[t] = now->g[t];
	}
      }
#endif
#ifdef TCD
      jlog("DEBUG: suspended as %s\n", (now->last_ph)->name);
#endif
      return;
    }

    /* scan认跋の不燎误を洁洒 */
    /* prepare HMM of the scan range */
    phmmlen = winfo->wlen[word] - 1;
    if (phmmlen > dwrk->phmmlen_max) {
      j_internal_error("scan_word: num of phonemes in a word exceed phmmlenmax (%d) ?\n", dwrk->phmmlen_max);
    }
    for (i=0;i<phmmlen-1;i++) {
      dwrk->phmmseq[i] = winfo->wseq[word][i+1];
    }
    dwrk->phmmseq[phmmlen-1] = tailph;
    if (enable_iwsp && hmminfo->multipath) {
      for (i=0;i<phmmlen-1;i++) dwrk->has_sp[i] = FALSE;
      dwrk->has_sp[phmmlen-1] = TRUE;
    }

  } else {			/* ~ccd_flag */

    phmmlen = winfo->wlen[word];
    for (i=0;i<phmmlen;i++) dwrk->phmmseq[i] = winfo->wseq[word][i];
    if (enable_iwsp && hmminfo->multipath) {
      for (i=0;i<phmmlen;i++) dwrk->has_sp[i] = FALSE;
      dwrk->has_sp[phmmlen-1] = TRUE;
    }

  }

  /* 傅のg[]をいったん略闰しておく */
  /* temporally keeps the original g[] */
  for (t=0;t<peseqlen;t++) dwrk->g[t] = now->g[t];

#ifdef GRAPHOUT_PRECISE_BOUNDARY
  if (r->graphout) {
    /* 帽胳董肠帕嚷攫鼠を介袋步 */
    /* initialize word boundary propagation info */
    for (t=0;t<peseqlen;t++) {
      dwrk->wef[t] = t;
      dwrk->wes[t] = now->g[t];
    }
  }
#endif

  /* viterbiを悸乖して g[] から now->g[] を构糠する */
  /* do viterbi computation for phmmseq from g[] to now->g[] */
  do_viterbi(dwrk->g, now->g, dwrk->phmmseq, (enable_iwsp && hmminfo->multipath) ? dwrk->has_sp : NULL, 
	     phmmlen, param, peseqlen, now->estimated_next_t, &(now->final_g)
#ifdef GRAPHOUT_PRECISE_BOUNDARY
	     /* 帽胳董肠攫鼠 we[] から now->wordend_frame[] を构糠する */
	     /* propagate word boundary info from we[] to now->wordend_frame[] */
	     , dwrk->wef, now->wordend_frame
	     , dwrk->wes, now->wordend_gscore
#else
	     , NULL, NULL
	     , NULL, NULL
#endif
	     , r
	     );
#ifdef GRAPHOUT_PRECISE_BOUNDARY
  if (! hmminfo->multipath) {
    if (r->graphout) {
      /* 肌搀の next_word 脱に董肠攫鼠を拇腊 */
      /* proceed word boundary for one step for next_word */
      now->wordend_frame[peseqlen-1] = now->wordend_frame[0];
      now->wordend_gscore[peseqlen-1] = now->wordend_gscore[0];
      for (t=0;t<peseqlen-1;t++) {
	now->wordend_frame[t] = now->wordend_frame[t+1];
	now->wordend_gscore[t] = now->wordend_gscore[t+1];
      }
    }
  }
#endif

  if (ccd_flag) {
    /* 肌搀のために now->last_ph を构糠 */
    /* update 'now->last_ph' for future scan_word() */
    now->last_ph = winfo->wseq[word][0];
    if (enable_iwsp && hmminfo->multipath) now->last_ph_sp_attached = FALSE; /* wlen > 1 here */
#ifdef TCD
    jlog("DEBUG: last_ph = %s\n", (now->last_ph)->name);
#endif
  }
}


/**************************************************************************/
/*** 糠簿棱の鸥倡とヒュ〖リスティックを芬いだ链挛スコアを纷换           ***/
/*** Expand new hypothesis and compute the total score (with heuristic) ***/
/**************************************************************************/

/** 
 * <JA>
 * 鸥倡傅簿棱に肌帽胳を儡鲁して糠しい簿棱を栏喇する. 肌帽胳の帽胳トレリス惧の
 * スコアから呵锑儡鲁爬を滇め·簿棱スコアを纷换する. 
 * 
 * @param now [in] 鸥倡傅簿棱
 * @param new [out] 糠たに栏喇された簿棱が呈羌される
 * @param nword [in] 儡鲁する肌帽胳の攫鼠
 * @param param [in] 掐蜗パラメ〖タ误
 * @param r [in] 千急借妄インスタンス
 * </JA>
 * <EN>
 * Connect a new word to generate a next hypothesis.  The optimal connection
 * point and new sentence score of the new hypothesis will be estimated by
 * looking up the corresponding words on word trellis.
 * 
 * @param now [in] source hypothesis
 * @param new [out] pointer to save the newly generated hypothesis
 * @param nword [in] next word to be connected
 * @param param [in] input parameter vector
 * @param r [in] recognition process instance
 * </EN>
 * @callgraph
 * @callergraph