kb.c

CMU大名鼎鼎的SPHINX－3大词汇量连续语音识别系统

			  cmd_ln_float64("-pbeam"),
			  cmd_ln_float64("-wbeam"),
			  cmd_ln_float64("-wend_beam"),
			  cmd_ln_int32("-ptranskip")
			  );

    E_INFO("Parameters used in Beam Pruning of Viterbi Search: Beam= %d, PBeam= %d, WBeam= %d (Skip=%d), WEndBeam=%d\n", kb->beam->hmm, kb->beam->ptrans, kb->beam->word, kb->beam->ptranskip, kb->beam->wordend);

    kb->histprune = histprune_init(cmd_ln_int32("-maxhmmpf"),
				   cmd_ln_int32("-maxhistpf"),
				   cmd_ln_int32("-maxwpf"));

    E_INFO("Parameters used in histogram pruning: Max. HMM per frame=%d, Max. History per frame=%d, Max. Word per frame=%d\n", kb->histprune->maxhmmpf,kb->histprune->maxhistpf,kb->histprune->maxwpf);

    /*Sections of fast GMM computation parameters*/

    kb->fastgmm = fast_gmm_init(cmd_ln_int32("-ds"),
			        cmd_ln_int32("-cond_ds"),
				cmd_ln_int32("-dist_ds"),
				cmd_ln_int32("-gs4gs"),
				cmd_ln_int32("-svq4svq"),
				cmd_ln_float64("-subvqbeam"),
				cmd_ln_float64("-ci_pbeam"),
				cmd_ln_int32("-maxcdsenpf"),
				kb->kbcore->mdef->n_ci_sen
				);

    E_INFO("Parameters used in Fast GMM computation:\n");
    E_INFO("   Frame-level: Down Sampling Ratio %d, Conditional Down Sampling? %d, Distance-based Down Sampling? %d\n",kb->fastgmm->downs->ds_ratio,kb->fastgmm->downs->cond_ds,kb->fastgmm->downs->dist_ds);
    E_INFO("     GMM-level: CI phone beam %d\n",kb->fastgmm->gmms->ci_pbeam);
    E_INFO("Gaussian-level: GS map would be used for Gaussian Selection? =%d, SVQ would be used as Gaussian Score? =%d SubVQ Beam %d\n",kb->fastgmm->gs4gs,kb->fastgmm->svq4svq,kb->fastgmm->gaus->subvqbeam);
    
    /* Not really nice to check it here. Later when we move gs and svq
       out of core. Things will be better. */

    if(kb->fastgmm->downs->cond_ds>0&&kb->kbcore->gs==NULL) 
      E_FATAL("Conditional Down Sampling require the use of Gaussian Selection map\n");

    kb->pl_win=cmd_ln_int32("-pl_window");
    E_INFO("Phoneme look-ahead window size = %d\n",kb->pl_win);

    kb->pl_win_strt=0;

    kb->pl_beam=logs3(cmd_ln_float64("-pl_beam"));
    E_INFO("Phoneme look-ahead beam = %d\n",kb->pl_beam);

    for(cisencnt=0;cisencnt==mdef->cd2cisen[cisencnt];cisencnt++) ;

    kb->cache_ci_senscr=(int32**)ckd_calloc_2d(kb->pl_win,cisencnt,sizeof(int32));
    kb->cache_best_list=(int32*)ckd_calloc(kb->pl_win,sizeof(int32));
    kb->phn_heur_list=(int32*)ckd_calloc(mdef_n_ciphone (mdef),sizeof(int32));

    kb->wordbestscore=(int32*)ckd_calloc(mdef_n_ciphone (mdef),sizeof(int32));
    kb->wordbestexit=(int32*)ckd_calloc(mdef_n_ciphone (mdef),sizeof(int32));
    kb->epl = cmd_ln_int32 ("-epl");

    if ((kb->feat = feat_array_alloc(kbcore_fcb(kbcore),S3_MAX_FRAMES)) == NULL)
	E_FATAL("feat_array_alloc() failed\n");
    
    kb->vithist = vithist_init(kbcore, kb->beam->word, cmd_ln_int32("-bghist"));
    
    ptmr_init (&(kb->tm_sen));
    ptmr_init (&(kb->tm_srch));
    ptmr_init (&(kb->tm_ovrhd));
    kb->tot_fr = 0;
    kb->tot_sen_eval = 0.0;
    kb->tot_gau_eval = 0.0;
    kb->tot_hmm_eval = 0.0;
    kb->tot_wd_exit = 0.0;
    
    kb->hmm_hist_binsize = cmd_ln_int32("-hmmhistbinsize");

    if(lmset)
      n = ((kb->ugtreeMulti[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;
    else
      n = ((kb->ugtree[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;

    n /= kb->hmm_hist_binsize;
    kb->hmm_hist_bins = n+1;
    kb->hmm_hist = (int32 *) ckd_calloc (n+1, sizeof(int32));	/* Really no need for +1 */
    
    /* Open hypseg file if specified */
    str = cmd_ln_str("-hypseg");
    kb->matchsegfp = NULL;
    if (str) {
#ifdef WIN32
	if ((kb->matchsegfp = fopen(str, "wt")) == NULL)
#else
	if ((kb->matchsegfp = fopen(str, "w")) == NULL)
#endif
	    E_ERROR("fopen(%s,w) failed; use FWDXCT: from std logfile\n", str);
    }

    str = cmd_ln_str("-hyp");
    kb->matchfp = NULL;
    if (str) {
#ifdef WIN32
	if ((kb->matchfp = fopen(str, "wt")) == NULL)
#else
	if ((kb->matchfp = fopen(str, "w")) == NULL)
#endif
	    E_ERROR("fopen(%s,w) failed; use FWDXCT: from std logfile\n", str);
    }
    
    /* Setting for MLLR matrix */
    kb->prevmllrfn=(char*)ckd_calloc(1024,sizeof(char));
    kb->prevmllrfn[0]='\0';
    
}

void kb_set_uttid(char *_uttid,kb_t* _kb)
{
  assert(_kb != NULL);
  assert(_uttid!=NULL);

  if (_kb->uttid != NULL) {
    ckd_free(_kb->uttid);
    _kb->uttid = NULL;
  }
  if ((_kb->uttid = ckd_malloc(strlen(_uttid) + 1)) == NULL) {
    E_FATAL("Failed to allocate space for utterance id.\n");
  }
  strcpy(_kb->uttid,_uttid);
}

void kb_setlm(char* lmname,kb_t* kb)
{
  lmset_t* lms;
  kbcore_t* kbc=NULL;
  int i = 0;
  int j;
  int n;
  /*  s3wid_t dictid;*/

  kbc=kb->kbcore;
  lms=kbc->lmset;
  kbc->lm=NULL;
  for(j=0;j<kb->n_lextree;j++){
    kb->ugtree[j]=NULL;
  }

  if(lms!=NULL || cmd_ln_str("-lmctlfn")){
    for(i=0;i<kbc->n_lm;i++){
      if(!strcmp(lmname,lms[i].name)){
	/* Point the current lm to a particular lm */
	kbc->lm=lms[i].lm;

	for(j=0;j<kb->n_lextree;j++){
	  kb->ugtree[j]=kb->ugtreeMulti[i*kb->n_lextree+j];
	}

	break;
      }
    }
    if(kbc->lm==NULL){
      E_ERROR("LM name %s cannot be found in the preloaded lm ctl file (spefied by -lmctlfile)! Fall back to use base LM model\n",lmname);
      kbc->lm=lms[0].lm;
      /* Also, don't forget to reset 'i', since if the code reaches
       * this point, i has a value of 'n_lm + 1'
       */
      i = 0;
      for(j=0;j<kb->n_lextree;j++){
	kb->ugtree[j]=kb->ugtreeMulti[j];
      }
    }
  }
  /*  Just to make sure we're not trying to point beyond the limit of
   *  array lms.
   */
  assert (i < kbc->n_lm);


  if((kb->vithist->lms2vh_root=
     (vh_lms2vh_t**)ckd_realloc(kb->vithist->lms2vh_root,
				lm_n_ug(kbc->lm)*sizeof(vh_lms2vh_t *)
				))==NULL) 
    {
      E_FATAL("failed to allocate memory for vithist\n");
    }


  n = ((kb->ugtree[0]->n_node) + (kb->fillertree[0]->n_node)) * kb->n_lextree;
  n /= kb->hmm_hist_binsize;
  kb->hmm_hist_bins = n+1;
  kb->hmm_hist = (int32 *) ckd_realloc (kb->hmm_hist,(n+1)*sizeof(int32));	/* Really no need for +1 */

  E_INFO("Current LM name %s\n",lms[i].name);
  E_INFO("LM ug size %d\n",kbc->lm->n_ug);
  E_INFO("LM bg size %d\n",kbc->lm->n_bg);
  E_INFO("LM tg size %d\n",kbc->lm->n_tg);
  E_INFO("HMM history bin size %d\n", n+1);

  for(j=0;j<kb->n_lextree;j++){
    E_INFO("Lextrees(%d), %d nodes(ug)\n",
	   kb->n_lextree, lextree_n_node(kb->ugtree[j]));
  }

}

void kb_setmllr(char* mllrname,kb_t* kb)
{
/*  int32 veclen;*/
  E_INFO("Using MLLR matrix %s\n", mllrname);
  
  if(strcmp(kb->prevmllrfn,mllrname)!=0){ /* If there is a change of mllr file name */
    /* Reread the gaussian mean from the file again */
    E_INFO("Reloading mean\n");
    mgau_mean_reload(kbcore_mgau(kb->kbcore),cmd_ln_str("-mean"));

    /* Read in the mllr matrix */

#if MLLR_DEBUG
    /*This generates huge amount of information */
    /*    mgau_dump(kbcore_mgau(kb->kbcore),1);*/
#endif

    mllr_read_regmat(mllrname,
		     &(kb->regA),
		     &(kb->regB),
		     mgau_veclen(kbcore_mgau(kb->kbcore)));




    /* Transform all the mean vectors */

    mllr_norm_mgau(kbcore_mgau(kb->kbcore),kb->regA,kb->regB,kbcore_mdef(kb->kbcore));

#if MLLR_DEBUG
    /*#if 1*/
    mllr_dump(kb->regA,kb->regB,mgau_veclen(kbcore_mgau(kb->kbcore)));
    /*This generates huge amount of information */
    /*mgau_dump(kbcore_mgau(kb->kbcore),1);*/
#endif 


    /* allocate memory for the prevmllrfn if it is too short*/
    if(strlen(mllrname)*sizeof(char) > 1024){
      kb->prevmllrfn=(char*)ckd_calloc(strlen(mllrname), sizeof(char));
    }

    strcpy(kb->prevmllrfn,mllrname);
  }else{
    /* No need to change anything for now */
  }
}


/*
 * Make the next_active information within all lextrees be the current one, after blowing
 * away the latter; in preparation for moving on to the next frame.
 */
void kb_lextree_active_swap (kb_t *kb)
{
    int32 i;
    
    for (i = 0; i < kb->n_lextree; i++) {
	lextree_active_swap (kb->ugtree[i]);
	lextree_active_swap (kb->fillertree[i]);
    }
}

/* RAH 4.15.01 Lots of memory is allocated, but never freed, this function will clean up.
 * First pass will get the low hanging fruit.*/
void kb_free (kb_t *kb)
{
  vithist_t *vithist = kb->vithist;

  if (kb->sen_active)
    ckd_free ((void *)kb->sen_active);
  if (kb->ssid_active) 
    ckd_free ((void *)kb->ssid_active);
  if (kb->comssid_active)
    ckd_free ((void *)kb->comssid_active);
  if (kb->fillertree) 
    ckd_free ((void *)kb->fillertree);
  if (kb->hmm_hist) 
    ckd_free ((void *)kb->hmm_hist);
  

  /* vithist */
  if (vithist) {
    ckd_free ((void *) vithist->entry);
    ckd_free ((void *) vithist->frame_start);
    ckd_free ((void *) vithist->bestscore);
    ckd_free ((void *) vithist->bestvh);
    ckd_free ((void *) vithist->lms2vh_root);    
    ckd_free ((void *) kb->vithist);
  }


  kbcore_free (kb->kbcore);

  if (kb->feat) {
    ckd_free ((void *)kb->feat[0][0]);
    ckd_free_2d ((void **)kb->feat);
  }

  if (kb->cache_ci_senscr) {
    ckd_free_2d ((void **)kb->cache_ci_senscr);
  }
  if( kb->cache_best_list) {
    ckd_free((void*) kb->cache_best_list);
  }
  if(kb->phn_heur_list) {
    ckd_free((void*) kb->phn_heur_list);
  }

  if (kb->matchsegfp) fclose(kb->matchsegfp);
  if (kb->matchfp) fclose(kb->matchfp);

  if(kb->prevmllrfn) ckd_free((char*) kb->prevmllrfn);
  if(kb->regA && kb->regB) mllr_free_regmat(kb->regA, kb->regB);
}