static void
phmm_free(allphone_search_t * allphs)
{
    s3cipid_t ci;
    bin_mdef_t *mdef;

    if (!allphs->ci_phmm)
        //nothing to free
        return;
    ckd_free(allphs->ci_phmm[0]->lc);
    mdef = ((ps_search_t *) allphs)->acmod->mdef;
    for (ci = 0; ci < mdef_n_ciphone(mdef); ++ci) {
        phmm_t *p, *next;

        for (p = allphs->ci_phmm[ci]; p; p = next) {
            plink_t *l, *lnext;

            next = p->next;
            for (l = p->succlist; l; l = lnext) {
                lnext = l->next;
                ckd_free(l);
            }
            hmm_deinit(&(p->hmm));
            ckd_free(p);
        }
    }
    ckd_free(allphs->ci_phmm);
}
/*
 * Build a glist of triphone senone-sequence IDs (ssids) derivable from [b] as a single
 * phone word.  If no triphone found in mdef, include the ssid for basephone b.
 * Return the generated glist.
 */
static glist_t single_comsseq (mdef_t *mdef, int32 b)
{
    int32 l, r, p, ssid;
    glist_t g;
    
    g = NULL;
    for (l = 0; l < mdef_n_ciphone(mdef); l++) {
	for (r = 0; r < mdef_n_ciphone(mdef); r++) {
	    p = mdef_phone_id (mdef, (s3cipid_t)b, (s3cipid_t)l, (s3cipid_t)r, WORD_POSN_SINGLE);
	    
	    if (IS_S3PID(p)) {
		ssid = mdef_pid2ssid(mdef, p);
		if (! glist_chkdup_int32 (g, ssid))
		    g = glist_add_int32 (g, ssid);
	    }
	}
    }
    if (! g)
	g = glist_add_int32 (g, mdef_pid2ssid(mdef, b));
    
    return g;
}
/*
 * Build a glist of triphone senone-sequence IDs (ssids) derivable from [r][b] at the word
 * end position.  If no triphone found in mdef, include the ssid for basephone b.
 * Return the generated glist.
 */
static glist_t rdiph_comsseq (mdef_t *mdef, int32 b, int32 l)
{
    int32 r, p, ssid;
    glist_t g;
    
    g = NULL;
    for (r = 0; r < mdef_n_ciphone(mdef); r++) {
	p = mdef_phone_id (mdef, (s3cipid_t)b, (s3cipid_t)l, (s3cipid_t)r, WORD_POSN_END);
	
	if (IS_S3PID(p)) {
	    ssid = mdef_pid2ssid(mdef, p);
	    if (! glist_chkdup_int32 (g, ssid))
		g = glist_add_int32 (g, ssid);
	}
    }
    if (! g)
	g = glist_add_int32 (g, mdef_pid2ssid(mdef, b));
    
    return g;
}
Exemplo n.º 4
0
Arquivo: kb.c Projeto: 4auka/cmusphinx
void
kb_init(kb_t * kb, cmd_ln_t *config)
{
    kbcore_t *kbcore;
    mdef_t *mdef;
    dict_t *dict;
    dict2pid_t *d2p;
    int32 cisencnt;

    /* STRUCTURE: Initialize the kb structure to zero, just in case */
    memset(kb, 0, sizeof(*kb));
    kb->kbcore = kbcore_init(config);
    if (kb->kbcore == NULL)
        E_FATAL("Initialization of kb failed\n");

    kbcore = kb->kbcore;
    mdef = kbcore_mdef(kbcore);
    dict = kbcore_dict(kbcore);
    d2p = kbcore_dict2pid(kbcore);

    err_set_debug_level(cmd_ln_int32_r(config, "-debug"));

    /* STRUCTURE INITIALIZATION: Initialize the beam data structure */
    if (cmd_ln_exists_r(config, "-ptranskip")) {
        kb->beam = beam_init(cmd_ln_float64_r(config, "-beam"),
                             cmd_ln_float64_r(config, "-pbeam"),
                             cmd_ln_float64_r(config, "-wbeam"),
                             cmd_ln_float64_r(config, "-wend_beam"),
                             cmd_ln_int32_r(config, "-ptranskip"), mdef_n_ciphone(mdef),
                             kbcore->logmath
            );

        /* REPORT : Report the parameters in the beam data structure */
        if (REPORT_KB)
                beam_report(kb->beam);
    }


    /* STRUCTURE INITIALIZATION: Initialize the fast GMM computation data structure */
    if (cmd_ln_exists_r(config, "-ci_pbeam")) {
        kb->fastgmm = fast_gmm_init(cmd_ln_int32_r(config, "-ds"),
                                    cmd_ln_int32_r(config, "-cond_ds"),
                                    cmd_ln_int32_r(config, "-dist_ds"),
                                    cmd_ln_int32_r(config, "-gs4gs"),
                                    cmd_ln_int32_r(config, "-svq4svq"),
                                    cmd_ln_float64_r(config, "-subvqbeam"),
                                    cmd_ln_float64_r(config, "-ci_pbeam"),
                                    cmd_ln_float64_r(config, "-tighten_factor"),
                                    cmd_ln_int32_r(config, "-maxcdsenpf"),
                                    mdef->n_ci_sen,
                                    kbcore->logmath);

        /* REPORT : Report the parameters in the fast_gmm_t data struture */
        if (REPORT_KB)
            fast_gmm_report(kb->fastgmm);
    }

    /* STRUCTURE INITIALIZATION: Initialize the phoneme lookahead data structure */
    if (cmd_ln_exists_r(config, "-pl_beam")) {
        kb->pl = pl_init(cmd_ln_int32_r(config, "-pheurtype"),
                         cmd_ln_float64_r(config, "-pl_beam"), mdef_n_ciphone(mdef),
                         kbcore->logmath
            );

        /* REPORT : Report the parameters in the pl_t data struture */
        if (REPORT_KB)
            pl_report(kb->pl);
    }

    /* STRUCTURE INITIALIZATION: Initialize the acoustic score data structure */
    {
        int32 pl_window = 1;

        if (cmd_ln_exists_r(config, "-pl_window"))
            pl_window = cmd_ln_int32_r(config, "-pl_window");

        for (cisencnt = 0; cisencnt == mdef->cd2cisen[cisencnt]; cisencnt++) ;
        kb->ascr = ascr_init(kbcore_n_mgau(kbcore),
                             kb->kbcore->dict2pid->n_comstate,
                             mdef_n_sseq(mdef),
                             dict2pid_n_comsseq(d2p),
                             pl_window, cisencnt);

        if (REPORT_KB)
            ascr_report(kb->ascr);
    }

    /* Initialize the front end if -adcin is specified */
    if (cmd_ln_exists_r(config, "-adcin") && cmd_ln_boolean_r(config, "-adcin")) {
        if ((kb->fe = fe_init_auto_r(config)) == NULL) {
            E_FATAL("fe_init_auto_r() failed\n");
        }
    }
    /* STRUCTURE INITIALIZATION : The feature vector */
    if ((kb->feat =
         feat_array_alloc(kbcore_fcb(kbcore), S3_MAX_FRAMES)) == NULL)
        E_FATAL("feat_array_alloc() failed\n");

    /* STRUCTURE INITIALIZATION : The statistics for the search */
    kb->stat = stat_init();

    /* STRUCTURE INITIALIZATION : The adaptation routines of the search */
    kb->adapt_am = adapt_am_init();

    if (cmd_ln_str_r(config, "-mllr")) {
        kb_setmllr(cmd_ln_str_r(config, "-mllr"), cmd_ln_str_r(config, "-cb2mllr"), kb);
    }

    /* CHECK: make sure when (-cond_ds) is specified, a Gaussian map is also specified */
    if (cmd_ln_int32_r(config, "-cond_ds") > 0 && kb->kbcore->gs == NULL)
        E_FATAL
            ("Conditional Down Sampling require the use of Gaussian Selection map\n");

    /* MEMORY ALLOCATION : Word best score and exit */
    /* Open hypseg file if specified */
    kb->matchsegfp = kb->matchfp = NULL;
    kb->matchsegfp = file_open(cmd_ln_str_r(config, "-hypseg"));
    kb->matchfp = file_open(cmd_ln_str_r(config, "-hyp"));

    if (cmd_ln_exists_r(config, "-hmmdump"))
        kb->hmmdumpfp = cmd_ln_int32_r(config, "-hmmdump") ? stderr : NULL;

    /* STRUCTURE INITIALIZATION : The search data structure, done only
       after kb is initialized kb is acted as a clipboard. */
    if (cmd_ln_exists_r(config, "-op_mode")) {
        /* -op_mode, if set (i.e. not -1), takes precedence over -mode. */
        if (cmd_ln_int32_r(config, "-op_mode") != -1)
            kb->op_mode = cmd_ln_int32_r(config, "-op_mode");
        else
            kb->op_mode = srch_mode_str_to_index(cmd_ln_str_r(config, "-mode"));
        E_INFO("SEARCH MODE INDEX %d\n", kb->op_mode);
        if ((kb->srch = (srch_t *) srch_init(kb, kb->op_mode)) == NULL) {
            E_FATAL("Search initialization failed. Forced exit\n");
        }
        if (REPORT_KB) {
            srch_report(kb->srch);
        }
    }
}
Exemplo n.º 5
0
/*ARCHAN, to allow backward compatibility -lm, -lmctlfn coexists. This makes the current implmentation more complicated than necessary. */
void kb_init (kb_t *kb)
{
    kbcore_t *kbcore;
    mdef_t *mdef;
    dict_t *dict;
    dict2pid_t *d2p;
    lm_t *lm;
    lmset_t *lmset;
    s3cipid_t sil, ci;
    s3wid_t w;
    int32 i, n, n_lc;
    wordprob_t *wp;
    s3cipid_t *lc;
    bitvec_t lc_active;
    char *str;
    int32 cisencnt;
    int32 j;
    
    /* Initialize the kb structure to zero, just in case */
    memset(kb, 0, sizeof(*kb));
    kb->kbcore = NULL;

    kb->kbcore = kbcore_init (cmd_ln_float32 ("-logbase"),
			      cmd_ln_str("-feat"),
			      cmd_ln_str("-cmn"),
			      cmd_ln_str("-varnorm"),
			      cmd_ln_str("-agc"),
			      cmd_ln_str("-mdef"),
			      cmd_ln_str("-dict"),
			      cmd_ln_str("-fdict"),
			      "",	/* Hack!! Hardwired constant 
						for -compsep argument */
			      cmd_ln_str("-lm"),
			      cmd_ln_str("-lmctlfn"),
			      cmd_ln_str("-lmdumpdir"),
			      cmd_ln_str("-fillpen"),
			      cmd_ln_str("-senmgau"),
			      cmd_ln_float32("-silprob"),
			      cmd_ln_float32("-fillprob"),
			      cmd_ln_float32("-lw"),
			      cmd_ln_float32("-wip"),
			      cmd_ln_float32("-uw"),
			      cmd_ln_str("-mean"),
			      cmd_ln_str("-var"),
			      cmd_ln_float32("-varfloor"),
			      cmd_ln_str("-mixw"),
			      cmd_ln_float32("-mixwfloor"),
			      cmd_ln_str("-subvq"),
			      cmd_ln_str("-gs"),
			      cmd_ln_str("-tmat"),
			      cmd_ln_float32("-tmatfloor"));
    if(kb->kbcore==NULL){
      E_FATAL("Initialization of kb failed\n");
    }

    kbcore = kb->kbcore;
    
    mdef = kbcore_mdef(kbcore);
    dict = kbcore_dict(kbcore);
    lm = kbcore_lm(kbcore);
    lmset=kbcore_lmset(kbcore);
    d2p = kbcore_dict2pid(kbcore);
    
    if (NOT_S3WID(dict_startwid(dict)) || NOT_S3WID(dict_finishwid(dict)))
	E_FATAL("%s or %s not in dictionary\n", S3_START_WORD, S3_FINISH_WORD);

    if(lmset){
      for(i=0;i<kbcore_nlm(kbcore);i++){
	if (NOT_S3LMWID(lm_startwid(lmset[i].lm)) || NOT_S3LMWID(lm_finishwid(lmset[i].lm)))
	E_FATAL("%s or %s not in LM %s\n", S3_START_WORD, S3_FINISH_WORD,lmset[i].name);
      }
    }else if(lm){
      if (NOT_S3LMWID(lm_startwid(lm)) || NOT_S3LMWID(lm_finishwid(lm)))
	E_FATAL("%s or %s not in LM\n", S3_START_WORD, S3_FINISH_WORD);
    }

    
    /* Check that HMM topology restrictions are not violated */
    if (tmat_chk_1skip (kbcore->tmat) < 0)
	E_FATAL("Tmat contains arcs skipping more than 1 state\n");
    
    /*
     * Unlink <s> and </s> between dictionary and LM, to prevent their 
     * recognition.  They are merely dummy words (anchors) at the beginning 
     * and end of each utterance.
     */
    if(lmset){
      for(i=0;i<kbcore_nlm(kbcore);i++){
	lm_lmwid2dictwid(lmset[i].lm, lm_startwid(lmset[i].lm)) = BAD_S3WID;
	lm_lmwid2dictwid(lmset[i].lm, lm_finishwid(lmset[i].lm)) = BAD_S3WID;

	for (w = dict_startwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
	  lmset[i].lm->dict2lmwid[w] = BAD_S3LMWID;
	for (w = dict_finishwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
	  lmset[i].lm->dict2lmwid[w] = BAD_S3LMWID;

      }
    }else if(lm){ /* No LM is set at this point*/
      lm_lmwid2dictwid(lm, lm_startwid(lm)) = BAD_S3WID;
      lm_lmwid2dictwid(lm, lm_finishwid(lm)) = BAD_S3WID;
      for (w = dict_startwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
	kbcore->dict2lmwid[w] = BAD_S3LMWID;
      for (w = dict_finishwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
	kbcore->dict2lmwid[w] = BAD_S3LMWID;

    }
    sil = mdef_silphone (kbcore_mdef (kbcore));
    if (NOT_S3CIPID(sil))
	E_FATAL("Silence phone '%s' not in mdef\n", S3_SILENCE_CIPHONE);
    
    
    kb->sen_active = (int32 *) ckd_calloc (mdef_n_sen(mdef), sizeof(int32));
    kb->rec_sen_active = (int32 *) ckd_calloc (mdef_n_sen(mdef), sizeof(int32));
    kb->ssid_active = (int32 *) ckd_calloc (mdef_n_sseq(mdef), sizeof(int32));
    kb->comssid_active = (int32 *) ckd_calloc (dict2pid_n_comsseq(d2p), sizeof(int32));
    
    /* Build set of all possible left contexts */
    lc = (s3cipid_t *) ckd_calloc (mdef_n_ciphone(mdef) + 1, sizeof(s3cipid_t));
    lc_active = bitvec_alloc (mdef_n_ciphone (mdef));
    for (w = 0; w < dict_size (dict); w++) {
	ci = dict_pron (dict, w, dict_pronlen(dict, w) - 1);
	if (! mdef_is_fillerphone (mdef, (int)ci))
	    bitvec_set (lc_active, ci);
    }
    ci = mdef_silphone(mdef);
    bitvec_set (lc_active, ci);
    for (ci = 0, n_lc = 0; ci < mdef_n_ciphone(mdef); ci++) {
	if (bitvec_is_set (lc_active, ci))
	    lc[n_lc++] = ci;
    }
    lc[n_lc] = BAD_S3CIPID;

    E_INFO("Building lextrees\n");
    /* Get the number of lexical tree*/
    kb->n_lextree = cmd_ln_int32 ("-Nlextree");
    if (kb->n_lextree < 1) {
	E_ERROR("No. of ugtrees specified: %d; will instantiate 1 ugtree\n", 
								kb->n_lextree);
	kb->n_lextree = 1;
    }

    /* ARCHAN: This code was rearranged in s3.4 implementation of dynamic LM */
    /* Build active word list */
    wp = (wordprob_t *) ckd_calloc (dict_size(dict), sizeof(wordprob_t));


    if(lmset){
      kb->ugtreeMulti = (lextree_t **) ckd_calloc (kbcore_nlm(kbcore)*kb->n_lextree, sizeof(lextree_t *));
      /* Just allocate pointers*/
      kb->ugtree = (lextree_t **) ckd_calloc (kb->n_lextree, sizeof(lextree_t *));

      for(i=0;i<kbcore_nlm(kbcore);i++){
	E_INFO("Creating Unigram Table for lm %d name %s\n",i,lmset[i].name);
	n=0;
	for(j=0;j<dict_size(dict);j++){ /*try to be very careful again */
	  wp[j].wid=-1;
	  wp[j].prob=-1;
	}
	n = lm_ug_wordprob (lmset[i].lm, dict,MAX_NEG_INT32, wp);
	E_INFO("Size of word table after unigram + words in class: %d.\n",n);
	if (n < 1)
	  E_FATAL("%d active words in %s\n", n,lmset[i].name);
	n = wid_wordprob2alt(dict,wp,n);
	E_INFO("Size of word table after adding alternative prons: %d.\n",n);
	if (cmd_ln_int32("-treeugprob") == 0) {
	  for (i = 0; i < n; i++)
	    wp[i].prob = -1;    	/* Flatten all initial probabilities */
	}

	for (j = 0; j < kb->n_lextree; j++) {
	  kb->ugtreeMulti[i*kb->n_lextree+j] = lextree_build (kbcore, wp, n, lc);
	  lextree_type (kb->ugtreeMulti[i*kb->n_lextree+j]) = 0;
	  E_INFO("Lextrees (%d) for lm %d name %s, %d nodes(ug)\n",
		 kb->n_lextree, i, lmset[i].name,lextree_n_node(kb->ugtreeMulti[i*kb->n_lextree+j]));
	}
      }

    }else if (lm){
      E_INFO("Creating Unigram Table\n");
      n=0;
      n = lm_ug_wordprob (lm, dict,MAX_NEG_INT32, wp);
      E_INFO("Size of word table after unigram + words in class: %d\n",n);
      if (n < 1)
	E_FATAL("%d active words\n", n);
      n = wid_wordprob2alt (dict, wp, n);	   /* Add alternative pronunciations */
      
      /* Retain or remove unigram probs from lextree, depending on option */
      if (cmd_ln_int32("-treeugprob") == 0) {
	for (i = 0; i < n; i++)
	  wp[i].prob = -1;    	/* Flatten all initial probabilities */
      }
      
      /* Create the desired no. of unigram lextrees */
      kb->ugtree = (lextree_t **) ckd_calloc (kb->n_lextree, sizeof(lextree_t *));
      for (i = 0; i < kb->n_lextree; i++) {
	kb->ugtree[i] = lextree_build (kbcore, wp, n, lc);
	lextree_type (kb->ugtree[i]) = 0;
      }
      E_INFO("Lextrees(%d), %d nodes(ug)\n",
	     kb->n_lextree, lextree_n_node(kb->ugtree[0]));
    }



    /* Create filler lextrees */
    /* ARCHAN : only one filler tree is supposed to be build even for dynamic LMs */
    n = 0;
    for (i = dict_filler_start(dict); i <= dict_filler_end(dict); i++) {
	if (dict_filler_word(dict, i)) {
	    wp[n].wid = i;
	    wp[n].prob = fillpen (kbcore->fillpen, i);
	    n++;
	}
    }


    kb->fillertree = (lextree_t **)ckd_calloc(kb->n_lextree,sizeof(lextree_t*));
    for (i = 0; i < kb->n_lextree; i++) {
	kb->fillertree[i] = lextree_build (kbcore, wp, n, NULL);
	lextree_type (kb->fillertree[i]) = -1;
    }
    ckd_free ((void *) wp);
    ckd_free ((void *) lc);
    bitvec_free (lc_active);


    E_INFO("Lextrees(%d), %d nodes(filler)\n",
	     kb->n_lextree, 
	     lextree_n_node(kb->fillertree[0]));
    

    if (cmd_ln_int32("-lextreedump")) {
      if(lmset){
	E_FATAL("Currently, doesn't support -lextreedump for multiple-LMs\n");
      }
      for (i = 0; i < kb->n_lextree; i++) {
	fprintf (stderr, "UGTREE %d\n", i);
	lextree_dump (kb->ugtree[i], dict, stderr);
      }
      for (i = 0; i < kb->n_lextree; i++) {
	fprintf (stderr, "FILLERTREE %d\n", i);
	lextree_dump (kb->fillertree[i], dict, stderr);
      }
      fflush (stderr);
    }
    
    kb->ascr = ascr_init (mgau_n_mgau(kbcore_mgau(kbcore)), 
				kbcore->dict2pid->n_comstate);
    kb->beam = beam_init (cmd_ln_float64("-subvqbeam"),
			  cmd_ln_float64("-beam"),
			  cmd_ln_float64("-pbeam"),
			  cmd_ln_float64("-wbeam"));
    E_INFO("Beam= %d, PBeam= %d, WBeam= %d, SVQBeam= %d\n",
	   kb->beam->hmm, kb->beam->ptrans, kb->beam->word, kb->beam->subvq);
    
    /*Sections of optimization related parameters*/
    kb->ds_ratio=cmd_ln_int32("-ds");
    E_INFO("Down Sampling Ratio = %d\n",kb->ds_ratio);
    
    kb->rec_bstcid=-1;
    kb->skip_count=0;
    
    kb->cond_ds=cmd_ln_int32("-cond_ds");
    E_INFO("Conditional Down Sampling Parameter = %d\n",kb->cond_ds);
    
    if(kb->cond_ds>0&&kb->kbcore->gs==NULL) E_FATAL("Conditional Down Sampling require the use of Gaussian Selection map\n");

    kb->gs4gs=cmd_ln_int32("-gs4gs");
    E_INFO("GS map would be used for Gaussian Selection? = %d\n",kb->gs4gs);

    kb->svq4svq=cmd_ln_int32("-svq4svq");
    E_INFO("SVQ would be used as Gaussian Score ?= %d\n",kb->svq4svq);

    kb->ci_pbeam=-1*logs3(cmd_ln_float32("-ci_pbeam"));
    E_INFO("CI phone beam to prune the number of parent CI phones in CI-base GMM Selection = %d\n",kb->ci_pbeam);
    if(kb->ci_pbeam>10000000){
      E_INFO("Virtually no CI phone beam is applied now. (ci_pbeam>1000000)\n");
    }
    
    kb->wend_beam=-1*logs3(cmd_ln_float32("-wend_beam"));
    E_INFO("Word-end pruning beam: %d\n",kb->wend_beam);

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

	kb->pl_window_start=0;

    kb->pl_beam=logs3(cmd_ln_float32("-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_window,cisencnt,sizeof(int32));
    kb->cache_best_list=(int32*)ckd_calloc(kb->pl_window,sizeof(int32));
    kb->phn_heur_list=(int32*)ckd_calloc(mdef_n_ciphone (mdef),sizeof(int32));
  

    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 SPEC_CPU_WINDOWS
	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 SPEC_CPU_WINDOWS
	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);
    }
}
Exemplo n.º 6
0
Arquivo: kb.c Projeto: 10v/cmusphinx
void kb_init (kb_t *kb)
{
    kbcore_t *kbcore;
    mdef_t *mdef;
    dict_t *dict;
    dict2pid_t *d2p;
    lm_t *lm;
    s3cipid_t sil, ci;
    s3wid_t w;
    int32 i, n, n_lc;
    wordprob_t *wp;
    s3cipid_t *lc;
    bitvec_t lc_active;
    char *str;
    
    /* Initialize the kb structure to zero, just in case */
    memset(kb, 0, sizeof(*kb));

    kb->kbcore = kbcore_init (cmd_ln_float32 ("-logbase"),
			      "1s_c_d_dd",    /* Hack!! Hardwired constant 
						for -feat argument */
			      cmd_ln_str("-cmn"),
			      cmd_ln_str("-varnorm"),
			      cmd_ln_str("-agc"),
			      cmd_ln_str("-mdef"),
			      cmd_ln_str("-dict"),
			      cmd_ln_str("-fdict"),
			      "",	/* Hack!! Hardwired constant 
						for -compsep argument */
			      cmd_ln_str("-lm"),
			      cmd_ln_str("-fillpen"),
			      cmd_ln_float32("-silprob"),
			      cmd_ln_float32("-fillprob"),
			      cmd_ln_float32("-lw"),
			      cmd_ln_float32("-wip"),
			      cmd_ln_float32("-uw"),
			      cmd_ln_str("-mean"),
			      cmd_ln_str("-var"),
			      cmd_ln_float32("-varfloor"),
			      cmd_ln_str("-mixw"),
			      cmd_ln_float32("-mixwfloor"),
			      cmd_ln_str("-subvq"),
			      cmd_ln_str("-tmat"),
			      cmd_ln_float32("-tmatfloor"));
    
    kbcore = kb->kbcore;
    
    mdef = kbcore_mdef(kbcore);
    dict = kbcore_dict(kbcore);
    lm = kbcore_lm(kbcore);
    d2p = kbcore_dict2pid(kbcore);
    
    if (NOT_S3WID(dict_startwid(dict)) || NOT_S3WID(dict_finishwid(dict)))
	E_FATAL("%s or %s not in dictionary\n", S3_START_WORD, S3_FINISH_WORD);
    if (NOT_S3LMWID(lm_startwid(lm)) || NOT_S3LMWID(lm_finishwid(lm)))
	E_FATAL("%s or %s not in LM\n", S3_START_WORD, S3_FINISH_WORD);
    
    /* Check that HMM topology restrictions are not violated */
    if (tmat_chk_1skip (kbcore->tmat) < 0)
	E_FATAL("Tmat contains arcs skipping more than 1 state\n");
    
    /*
     * Unlink <s> and </s> between dictionary and LM, to prevent their 
     * recognition.  They are merely dummy words (anchors) at the beginning 
     * and end of each utterance.
     */
    lm_lmwid2dictwid(lm, lm_startwid(lm)) = BAD_S3WID;
    lm_lmwid2dictwid(lm, lm_finishwid(lm)) = BAD_S3WID;
    for (w = dict_startwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
	kbcore->dict2lmwid[w] = BAD_S3LMWID;
    for (w = dict_finishwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
	kbcore->dict2lmwid[w] = BAD_S3LMWID;
    
    sil = mdef_silphone (kbcore_mdef (kbcore));
    if (NOT_S3CIPID(sil))
	E_FATAL("Silence phone '%s' not in mdef\n", S3_SILENCE_CIPHONE);
    
    E_INFO("Building lextrees\n");
    
    kb->sen_active = (int32 *) ckd_calloc (mdef_n_sen(mdef), sizeof(int32));
    kb->ssid_active = (int32 *) ckd_calloc (mdef_n_sseq(mdef), sizeof(int32));
    kb->comssid_active = (int32 *) ckd_calloc (dict2pid_n_comsseq(d2p), 
							sizeof(int32));
    /* Build active word list */
    wp = (wordprob_t *) ckd_calloc (dict_size(dict), sizeof(wordprob_t));
    n = lm_ug_wordprob (lm, MAX_NEG_INT32, wp);
    if (n < 1)
	E_FATAL("%d active words\n", n);
    n = wid_wordprob2alt (dict, wp, n);	   /* Add alternative pronunciations */
    
    /* Retain or remove unigram probs from lextree, depending on option */
    if (cmd_ln_int32("-treeugprob") == 0) {
	for (i = 0; i < n; i++)
	    wp[i].prob = -1;    	/* Flatten all initial probabilities */
    }
    
    /* Build set of all possible left contexts */
    lc = (s3cipid_t *) ckd_calloc (mdef_n_ciphone(mdef) + 1, sizeof(s3cipid_t));
    lc_active = bitvec_alloc (mdef_n_ciphone (mdef));
    for (w = 0; w < dict_size (dict); w++) {
	ci = dict_pron (dict, w, dict_pronlen(dict, w) - 1);
	if (! mdef_is_fillerphone (mdef, (int)ci))
	    bitvec_set (lc_active, ci);
    }
    ci = mdef_silphone(mdef);
    bitvec_set (lc_active, ci);
    for (ci = 0, n_lc = 0; ci < mdef_n_ciphone(mdef); ci++) {
	if (bitvec_is_set (lc_active, ci))
	    lc[n_lc++] = ci;
    }
    lc[n_lc] = BAD_S3CIPID;
    
    /* Create the desired no. of unigram lextrees */
    kb->n_lextree = cmd_ln_int32 ("-Nlextree");
    if (kb->n_lextree < 1) {
	E_ERROR("No. of ugtrees specified: %d; will instantiate 1 ugtree\n", 
								kb->n_lextree);
	kb->n_lextree = 1;
    }
    kb->ugtree = (lextree_t **) ckd_calloc (kb->n_lextree, sizeof(lextree_t *));
    for (i = 0; i < kb->n_lextree; i++) {
	kb->ugtree[i] = lextree_build (kbcore, wp, n, lc);
	lextree_type (kb->ugtree[i]) = 0;
    }
    bitvec_free (lc_active);
    ckd_free ((void *) lc);
    
    /* Create filler lextrees */
    n = 0;
    for (i = dict_filler_start(dict); i <= dict_filler_end(dict); i++) {
	if (dict_filler_word(dict, i)) {
	    wp[n].wid = i;
	    wp[n].prob = fillpen (kbcore->fillpen, i);
	    n++;
	}
    }
    kb->fillertree = (lextree_t **)ckd_calloc(kb->n_lextree,sizeof(lextree_t*));
    for (i = 0; i < kb->n_lextree; i++) {
	kb->fillertree[i] = lextree_build (kbcore, wp, n, NULL);
	lextree_type (kb->fillertree[i]) = -1;
    }
    ckd_free ((void *) wp);
    
    E_INFO("Lextrees(%d), %d nodes(ug), %d nodes(filler)\n",
	   kb->n_lextree, lextree_n_node(kb->ugtree[0]), 
			lextree_n_node(kb->fillertree[0]));
    
    if (cmd_ln_int32("-lextreedump")) {
	for (i = 0; i < kb->n_lextree; i++) {
	    fprintf (stderr, "UGTREE %d\n", i);
	    lextree_dump (kb->ugtree[i], dict, stderr);
	}
	for (i = 0; i < kb->n_lextree; i++) {
	    fprintf (stderr, "FILLERTREE %d\n", i);
	    lextree_dump (kb->fillertree[i], dict, stderr);
	}
	fflush (stderr);
    }
    
    kb->ascr = ascr_init (mgau_n_mgau(kbcore_mgau(kbcore)), 
				kbcore->dict2pid->n_comstate);
    kb->beam = beam_init (cmd_ln_float64("-subvqbeam"),
			  cmd_ln_float64("-beam"),
			  cmd_ln_float64("-pbeam"),
			  cmd_ln_float64("-wbeam"));
    E_INFO("Beam= %d, PBeam= %d, WBeam= %d, SVQBeam= %d\n",
	   kb->beam->hmm, kb->beam->ptrans, kb->beam->word, kb->beam->subvq);
    
    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));
    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");
    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);
    }
}
Exemplo n.º 7
0
word_fsg_t *
word_fsg_load(s2_fsg_t * fsg,
              int use_altpron, int use_filler,
	      kbcore_t *kbc)
{
    float32 silprob = kbc->fillpen->silprob;
    float32 fillprob = kbc->fillpen->fillerprob;
    float32 lw = kbc->fillpen->lw;
    word_fsg_t *word_fsg;
    s2_fsg_trans_t *trans;
    int32 n_trans, n_null_trans, n_alt_trans, n_filler_trans, n_unk;
    int32 wid;
    int32 logp;
    glist_t nulls;
    int32 i, j;

    assert(fsg);

    /* Some error checking */
    if (lw <= 0.0)
        E_WARN("Unusual language-weight value: %.3e\n", lw);
    if (use_filler && ((silprob < 0.0) || (fillprob < 0.0))) {
        E_ERROR("silprob/fillprob must be >= 0\n");
        return NULL;
    }
    if ((fsg->n_state <= 0)
        || ((fsg->start_state < 0) || (fsg->start_state >= fsg->n_state))
        || ((fsg->final_state < 0) || (fsg->final_state >= fsg->n_state))) {
        E_ERROR("Bad #states/start_state/final_state values: %d/%d/%d\n",
                fsg->n_state, fsg->start_state, fsg->final_state);
        return NULL;
    }
    for (trans = fsg->trans_list; trans; trans = trans->next) {
        if ((trans->from_state < 0) || (trans->from_state >= fsg->n_state)
            || (trans->to_state < 0) || (trans->to_state >= fsg->n_state)
            || (trans->prob <= 0) || (trans->prob > 1.0)) {
            E_ERROR("Bad transition: P(%d -> %d) = %e\n",
                    trans->from_state, trans->to_state, trans->prob);
            return NULL;
        }
    }


    word_fsg = (word_fsg_t *) ckd_calloc(1, sizeof(word_fsg_t));
    word_fsg->name = ckd_salloc(fsg->name ? fsg->name : "");
    word_fsg->n_state = fsg->n_state;
    word_fsg->start_state = fsg->start_state;
    word_fsg->final_state = fsg->final_state;
    word_fsg->use_altpron = use_altpron;
    word_fsg->use_filler = use_filler;
    word_fsg->lw = lw;
    word_fsg->lc = NULL;
    word_fsg->rc = NULL;
    word_fsg->dict = kbc->dict;
    word_fsg->mdef = kbc->mdef;
    word_fsg->tmat = kbc->tmat;
    word_fsg->n_ciphone = mdef_n_ciphone(kbc->mdef);


    /* Allocate non-epsilon transition matrix array */
    word_fsg->trans = (glist_t **) ckd_calloc_2d(word_fsg->n_state,
                                                 word_fsg->n_state,
                                                 sizeof(glist_t));
    /* Allocate epsilon transition matrix array */
    word_fsg->null_trans = (word_fsglink_t ***)
        ckd_calloc_2d(word_fsg->n_state, word_fsg->n_state,
                      sizeof(word_fsglink_t *));

    /* Process transitions */
    n_null_trans = 0;
    n_alt_trans = 0;
    n_filler_trans = 0;
    n_unk = 0;
    nulls = NULL;

    for (trans = fsg->trans_list, n_trans = 0;
         trans; trans = trans->next, n_trans++) {
        /* Convert prob to logs2prob and apply language weight */
        logp = (int32) (logs3(kbcore_logmath(kbc), trans->prob) * lw);

        /* Check if word is in dictionary */
        if (trans->word) {
            wid = dict_wordid(kbc->dict, trans->word);
            if (wid < 0) {
                E_ERROR("Unknown word '%s'; ignored\n", trans->word);
                n_unk++;
            }
            else if (use_altpron) {
                wid = dict_basewid(kbc->dict, wid);
                assert(wid >= 0);
            }
        }
        else
            wid = -1;           /* Null transition */

        /* Add transition to word_fsg structure */
        i = trans->from_state;
        j = trans->to_state;
        if (wid < 0) {
            if (word_fsg_null_trans_add(word_fsg, i, j, logp) == 1) {
                n_null_trans++;
                nulls =
                    glist_add_ptr(nulls,
                                  (void *) word_fsg->null_trans[i][j]);
            }
        }
        else {
            word_fsg_trans_add(word_fsg, i, j, logp, wid);

            /* Add transitions for alternative pronunciations, if any */
            if (use_altpron) {
                for (wid = dict_nextalt(kbc->dict, wid);
                     wid >= 0; wid = dict_nextalt(kbc->dict, wid)) {
                    word_fsg_trans_add(word_fsg, i, j, logp, wid);
                    n_alt_trans++;
                    n_trans++;
                }
            }
        }
    }

    /* Add silence and noise filler word transitions if specified */
    if (use_filler) {
        n_filler_trans = word_fsg_add_filler(word_fsg, silprob, fillprob, kbcore_logmath(kbc));
        n_trans += n_filler_trans;
    }

    E_INFO
        ("FSG: %d states, %d transitions (%d null, %d alt, %d filler,  %d unknown)\n",
         word_fsg->n_state, n_trans, n_null_trans, n_alt_trans,
         n_filler_trans, n_unk);

#if __FSG_DBG__
    E_INFO("FSG before NULL closure:\n");
    word_fsg_write(word_fsg, stdout);
#endif

    /* Null transitions closure */
    nulls = word_fsg_null_trans_closure(word_fsg, nulls);
    glist_free(nulls);

#if __FSG_DBG__
    E_INFO("FSG after NULL closure:\n");
    word_fsg_write(word_fsg, stdout);
#endif

    /* Compute left and right context CIphone lists for each state */
    word_fsg_lc_rc(word_fsg);

#if __FSG_DBG__
    E_INFO("FSG after lc/rc:\n");
    word_fsg_write(word_fsg, stdout);
#endif

    return word_fsg;
}
Exemplo n.º 8
0
static int32
dict_read(FILE * fp, dict_t * d)
{
    char line[16384], **wptr;
    s3cipid_t p[4096];
    int32 lineno, nwd;
    s3wid_t w;
    int32 i, maxwd;
    s3cipid_t ci;
    int32 ph;

    maxwd = 4092;
    wptr = (char **) ckd_calloc(maxwd, sizeof(char *)); /* Freed below */

    lineno = 0;
    while (fgets(line, sizeof(line), fp) != NULL) {
        lineno++;
        if (line[0] == '#')     /* Comment line */
            continue;

        if ((nwd = str2words(line, wptr, maxwd)) < 0)
            E_FATAL("str2words(%s) failed; Increase maxwd from %d\n", line,
                    maxwd);

        if (nwd == 0)           /* Empty line */
            continue;
        /* wptr[0] is the word-string and wptr[1..nwd-1] the pronunciation sequence */
        if (nwd == 1) {
            E_ERROR("Line %d: No pronunciation for word %s; ignored\n",
                    lineno, wptr[0]);
            continue;
        }
        {char * fin;
	  float proba=0.0;
	  int deca=0;
	  proba=strtod(wptr[1],&fin);
	  if (fin !=wptr[1]) 
	    deca=1;
	  else
	    proba=0.0;
	  
        /* Convert pronunciation string to CI-phone-ids */
        for (i = 1; i < nwd-deca; i++) {
            p[i - 1] = dict_ciphone_id(d, wptr[i+deca]);
            if (NOT_S3CIPID(p[i - 1])) {
                E_ERROR("Line %d: Bad ciphone: %s; word %s ignored\n",
                        lineno, wptr[i], wptr[0]);
                break;
            }
        }

        if (i == nwd-deca) {         /* All CI-phones successfully converted to IDs */
            w = dict_add_word(d, wptr[0], p, nwd - 1-deca);
            if (NOT_S3WID(w))
                E_ERROR
                    ("Line %d: dict_add_word (%s) failed (duplicate?); ignored\n",
                     lineno, wptr[0]);
	    d->word[w].proba=proba;
        }
	}
    }
    

    if (d->lts_rules) {

#if 1                           /* Until we allow user to put in a mapping of the phoneset from LTS to the phoneset from mdef, 
                                   The checking will intrusively stop the recognizer.  */

        for (ci = 0; ci < mdef_n_ciphone(d->mdef); ci++) {

            if (!mdef_is_fillerphone(d->mdef, ci)) {
                for (ph = 0; cmu6_lts_phone_table[ph] != NULL; ph++) {

                    /*        E_INFO("%s %s\n",cmu6_lts_phone_table[ph],mdef_ciphone_str(d->mdef,ci)); */
                    if (!strcmp
                        (cmu6_lts_phone_table[ph],
                         mdef_ciphone_str(d->mdef, ci)))
                        break;
                }
                if (cmu6_lts_phone_table[ph] == NULL) {
                    E_FATAL
                        ("A phone in the model definition doesn't appear in the letter to sound ",
                         "rules. \n This is case we don't recommend user to ",
                         "use the built-in LTS. \n Please kindly turn off ",
                         "-lts_mismatch\n");
                }
            }
        }
#endif
    }
Exemplo n.º 9
0
Arquivo: kb.c Projeto: wdebeaum/cabot
/* Update kb w/ new dictionary and new LM.
 * assumes: single-LM kbcore (before & after)
 * requires: updating kbcore
 * Lucian Galescu, 08/11/2005
 */
void kb_update_lm(kb_t *kb, char *dictfile, char *lmfile)
{
  kbcore_t *kbcore;
  mdef_t *mdef;
  dict_t *dict;
  dict2pid_t *d2p;
  lm_t *lm;
  s3cipid_t ci;
  s3wid_t w;
  int32 i, n, n_lc;
  wordprob_t *wp;
  s3cipid_t *lc;
  bitvec_t lc_active;
  
  /*** clean up ***/
  vithist_t *vithist = kb->vithist;

  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);
  }
  
  /*** re-initialize ***/
  
  kb->kbcore = kbcore_update_lm(kb->kbcore, 
                                dictfile, 
                                cmd_ln_str("-fdict"),
                                "",	/* Hack!! Hardwired constant for -compsep argument */
                                lmfile,
                                cmd_ln_str("-fillpen"),
                                cmd_ln_float32("-silprob"),
                                cmd_ln_float32("-fillprob"),
                                cmd_ln_float32("-lw"),
                                cmd_ln_float32("-wip"),
                                cmd_ln_float32("-uw"));
  if(kb->kbcore==NULL){
    E_FATAL("Updating kbcore failed\n");
  }
  
  kbcore = kb->kbcore;
  
  mdef = kbcore_mdef(kbcore);
  dict = kbcore_dict(kbcore);
  lm = kbcore_lm(kbcore);
  d2p = kbcore_dict2pid(kbcore);
  
  if (NOT_S3WID(dict_startwid(dict)) || NOT_S3WID(dict_finishwid(dict)))
    E_FATAL("%s or %s not in dictionary\n", S3_START_WORD, S3_FINISH_WORD);
  
  if(lm){
    if (NOT_S3LMWID(lm_startwid(lm)) || NOT_S3LMWID(lm_finishwid(lm)))
      E_FATAL("%s or %s not in LM\n", S3_START_WORD, S3_FINISH_WORD);
  }
  
  /*
   * Unlink <s> and </s> between dictionary and LM, to prevent their 
   * recognition.  They are merely dummy words (anchors) at the beginning 
   * and end of each utterance.
   */
  if(lm){
    lm_lmwid2dictwid(lm, lm_startwid(lm)) = BAD_S3WID;
    lm_lmwid2dictwid(lm, lm_finishwid(lm)) = BAD_S3WID;
    for (w = dict_startwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
      kbcore->dict2lmwid[w] = BAD_S3LMWID;
    for (w = dict_finishwid(dict); IS_S3WID(w); w = dict_nextalt(dict, w))
      kbcore->dict2lmwid[w] = BAD_S3LMWID;
  }
  
  /* Build set of all possible left contexts */
  lc = (s3cipid_t *) ckd_calloc (mdef_n_ciphone(mdef) + 1, sizeof(s3cipid_t));
  lc_active = bitvec_alloc (mdef_n_ciphone (mdef));
  for (w = 0; w < dict_size (dict); w++) {
    ci = dict_pron (dict, w, dict_pronlen(dict, w) - 1);
    if (! mdef_is_fillerphone (mdef, (int)ci))
	    bitvec_set (lc_active, ci);
  }
  ci = mdef_silphone(mdef);
  bitvec_set (lc_active, ci);
  for (ci = 0, n_lc = 0; ci < mdef_n_ciphone(mdef); ci++) {
    if (bitvec_is_set (lc_active, ci))
	    lc[n_lc++] = ci;
  }
  lc[n_lc] = BAD_S3CIPID;
  
  E_INFO("Building lextrees\n");
  /* Get the number of lexical tree*/
  kb->n_lextree = cmd_ln_int32 ("-Nlextree");
  if (kb->n_lextree < 1) {
    E_ERROR("No. of ugtrees specified: %d; will instantiate 1 ugtree\n", 
            kb->n_lextree);
    kb->n_lextree = 1;
  }
  
  /* ARCHAN: This code was rearranged in s3.4 implementation of dynamic LM */
  /* Build active word list */
  wp = (wordprob_t *) ckd_calloc (dict_size(dict), sizeof(wordprob_t));
  
  
  if (lm) {
    E_INFO("Creating Unigram Table\n");
    n=0;
    n = lm_ug_wordprob (lm, dict, MAX_NEG_INT32, wp);
    E_INFO("Size of word table after unigram + words in class: %d\n",n);
    if (n < 1)
      E_FATAL("%d active words\n", n);
    n = wid_wordprob2alt (dict, wp, n);	   /* Add alternative pronunciations */
    
    /* Retain or remove unigram probs from lextree, depending on option */
    if (cmd_ln_int32("-treeugprob") == 0) {
      for (i = 0; i < n; i++)
        wp[i].prob = -1;    	/* Flatten all initial probabilities */
    }
    
    /* Create the desired no. of unigram lextrees */
    kb->ugtree = (lextree_t **) ckd_calloc (kb->n_lextree, sizeof(lextree_t *));
    for (i = 0; i < kb->n_lextree; i++) {
      kb->ugtree[i] = lextree_build (kbcore, wp, n, lc);
      lextree_type (kb->ugtree[i]) = 0;
    }
    E_INFO("Lextrees(%d), %d nodes(ug)\n",
           kb->n_lextree, lextree_n_node(kb->ugtree[0]));
  }

  /* Create filler lextrees */
  /* ARCHAN : only one filler tree is supposed to be build even for dynamic LMs */
  n = 0;
  for (i = dict_filler_start(dict); i <= dict_filler_end(dict); i++) {
    if (dict_filler_word(dict, i)) {
      wp[n].wid = i;
      wp[n].prob = fillpen (kbcore->fillpen, i);
      n++;
    }
  }


  kb->fillertree = (lextree_t **)ckd_calloc(kb->n_lextree,sizeof(lextree_t*));
  for (i = 0; i < kb->n_lextree; i++) {
    kb->fillertree[i] = lextree_build (kbcore, wp, n, NULL);
    lextree_type (kb->fillertree[i]) = -1;
  }
  ckd_free ((void *) wp);
  ckd_free ((void *) lc);
  bitvec_free (lc_active);


  E_INFO("Lextrees(%d), %d nodes(filler)\n",
         kb->n_lextree, 
         lextree_n_node(kb->fillertree[0]));

  if (cmd_ln_int32("-lextreedump")) {
    for (i = 0; i < kb->n_lextree; i++) {
      fprintf (stderr, "UGTREE %d\n", i);
      lextree_dump (kb->ugtree[i], dict, stderr);
    }
    for (i = 0; i < kb->n_lextree; i++) {
      fprintf (stderr, "FILLERTREE %d\n", i);
      lextree_dump (kb->fillertree[i], dict, stderr);
    }
    fflush (stderr);
  }

  kb->ascr = ascr_init (mgau_n_mgau(kbcore_mgau(kbcore)), 
                        kbcore->dict2pid->n_comstate);

  kb->vithist = vithist_init(kbcore, kb->beam->word, cmd_ln_int32("-bghist"));

  kb->hmm_hist_binsize = cmd_ln_int32("-hmmhistbinsize");

  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 */

}
void dict2pid_dump (FILE *fp, dict2pid_t *d2p, mdef_t *mdef, dict_t *dict)
{
    int32 w, p, pronlen;
    int32 i, j, b, l, r;
    
    fprintf (fp, "# INTERNAL (wd comssid ssid ssid ... ssid comssid)\n");
    for (w = 0; w < dict_size(dict); w++) {
	fprintf (fp, "%30s ", dict_wordstr(dict, w));
	
	pronlen = dict_pronlen(dict, w);
	for (p = 0; p < pronlen; p++)
	    fprintf (fp, " %5d", d2p->internal[w][p]);
	fprintf (fp, "\n");
    }
    fprintf (fp, "#\n");
    
    fprintf (fp, "# LDIPH_LC (b r l ssid)\n");
    for (b = 0; b < mdef_n_ciphone(mdef); b++) {
	for (r = 0; r < mdef_n_ciphone(mdef); r++) {
	    for (l = 0; l < mdef_n_ciphone(mdef); l++) {
		if (IS_S3SSID(d2p->ldiph_lc[b][r][l]))
		    fprintf (fp, "%6s %6s %6s %5d\n",
			     mdef_ciphone_str (mdef, (s3cipid_t)b),
			     mdef_ciphone_str (mdef, (s3cipid_t)r),
			     mdef_ciphone_str (mdef, (s3cipid_t)l),
			     d2p->ldiph_lc[b][r][l]); /* RAH, ldiph_lc is returning an int32, %d expects an int16 */
	    }
	}
    }
    fprintf (fp, "#\n");
    
    fprintf (fp, "# SINGLE_LC (b l comssid)\n");
    for (b = 0; b < mdef_n_ciphone(mdef); b++) {
	for (l = 0; l < mdef_n_ciphone(mdef); l++) {
	    if (IS_S3SSID(d2p->single_lc[b][l]))
		fprintf (fp, "%6s %6s %5d\n",
			 mdef_ciphone_str (mdef, (s3cipid_t)b),
			 mdef_ciphone_str (mdef, (s3cipid_t)l),
			 d2p->single_lc[b][l]);	/* RAH, single_lc is returning an int32, %d expects an int16 */
	}
    }
    fprintf (fp, "#\n");
    
    fprintf (fp, "# SSEQ %d (senid senid ...)\n", mdef->n_sseq);
    for (i = 0; i < mdef->n_sseq; i++) {
	fprintf (fp, "%5d ", i);
	for (j = 0; j < mdef_n_emit_state(mdef); j++)
	    fprintf (fp, " %5d", mdef->sseq[i][j]);
	fprintf (fp, "\n");
    }
    fprintf (fp, "#\n");
    
    fprintf (fp, "# COMSSEQ %d (comstate comstate ...)\n", d2p->n_comsseq);
    for (i = 0; i < d2p->n_comsseq; i++) {
	fprintf (fp, "%5d ", i);
	for (j = 0; j < mdef_n_emit_state(mdef); j++)
	    fprintf (fp, " %5d", d2p->comsseq[i][j]);
	fprintf (fp, "\n");
    }
    fprintf (fp, "#\n");
    
    fprintf (fp, "# COMSTATE %d (senid senid ...)\n", d2p->n_comstate);
    for (i = 0; i < d2p->n_comstate; i++) {
	fprintf (fp, "%5d ", i);
	for (j = 0; IS_S3SENID(d2p->comstate[i][j]); j++)
	    fprintf (fp, " %5d", d2p->comstate[i][j]);
	fprintf (fp, "\n");
    }
    fprintf (fp, "#\n");
    fprintf (fp, "# END\n");
    
    fflush (fp);
}
/* RAH 4.16.01 This code has several leaks that must be fixed */
dict2pid_t *dict2pid_build (mdef_t *mdef, dict_t *dict)
{
    dict2pid_t *dict2pid;
    s3ssid_t *internal, **ldiph, **rdiph, *single;
    int32 pronlen;
    hash_table_t *hs, *hp;
    glist_t g;
    gnode_t *gn;
    s3senid_t *sen;
    hash_entry_t *he;
    int32 *cslen;
    int32 i, j, b, l, r, w, n, p;
    
    E_INFO("Building PID tables for dictionary\n");

    dict2pid = (dict2pid_t *) ckd_calloc (1, sizeof(dict2pid_t));
    dict2pid->internal = (s3ssid_t **) ckd_calloc (dict_size(dict), sizeof(s3ssid_t *));
    dict2pid->ldiph_lc = (s3ssid_t ***) ckd_calloc_3d (mdef->n_ciphone,
						       mdef->n_ciphone,
						       mdef->n_ciphone,
						       sizeof(s3ssid_t));
    dict2pid->single_lc = (s3ssid_t **) ckd_calloc_2d (mdef->n_ciphone,
						       mdef->n_ciphone,
						       sizeof(s3ssid_t));
    dict2pid->n_comstate = 0;
    dict2pid->n_comsseq = 0;
    
    hs = hash_new (mdef->n_ciphone * mdef->n_ciphone * mdef->n_emit_state, HASH_CASE_YES);
    hp = hash_new (mdef->n_ciphone * mdef->n_ciphone, HASH_CASE_YES);
    
    for (w = 0, n = 0; w < dict_size(dict); w++) {
	pronlen = dict_pronlen(dict, w);
	if (pronlen < 0)
	    E_FATAL("Pronunciation-length(%s)= %d\n", dict_wordstr(dict, w), pronlen);
	n += pronlen;
    }

    internal = (s3ssid_t *) ckd_calloc (n, sizeof(s3ssid_t));
    
    /* Temporary */
    ldiph = (s3ssid_t **) ckd_calloc_2d (mdef->n_ciphone, mdef->n_ciphone, sizeof(s3ssid_t));
    rdiph = (s3ssid_t **) ckd_calloc_2d (mdef->n_ciphone, mdef->n_ciphone, sizeof(s3ssid_t));
    single = (s3ssid_t *) ckd_calloc (mdef->n_ciphone, sizeof(s3ssid_t));
    for (b = 0; b < mdef->n_ciphone; b++) {
	for (l = 0; l < mdef->n_ciphone; l++) {
	    for (r = 0; r < mdef->n_ciphone; r++)
		dict2pid->ldiph_lc[b][r][l] = BAD_S3SSID;
	    
	    dict2pid->single_lc[b][l] = BAD_S3SSID;
	    
	    ldiph[b][l] = BAD_S3SSID;
	    rdiph[b][l] = BAD_S3SSID;
	}
	single[b] = BAD_S3SSID;
    }
    
    for (w = 0; w < dict_size(dict); w++) {
	dict2pid->internal[w] = internal;
	pronlen = dict_pronlen(dict,w);
	
	if (pronlen >= 2) {
	    b = dict_pron(dict, w, 0);
	    r = dict_pron(dict, w, 1);
	    if (NOT_S3SSID(ldiph[b][r])) {
		g = ldiph_comsseq(mdef, b, r);
		ldiph[b][r] = ssidlist2comsseq (g, mdef, dict2pid, hs, hp);
		glist_free (g);
		
		for (l = 0; l < mdef_n_ciphone(mdef); l++) {
		    p = mdef_phone_id_nearest (mdef, (s3cipid_t)b, (s3cipid_t)l, (s3cipid_t)r, WORD_POSN_BEGIN);
		    dict2pid->ldiph_lc[b][r][l] = mdef_pid2ssid(mdef, p);
		}
	    }
	    internal[0] = ldiph[b][r];
	    
	    for (i = 1; i < pronlen-1; i++) {
		l = b;
		b = r;
		r = dict_pron(dict, w, i+1);
		
		p = mdef_phone_id_nearest(mdef, (s3cipid_t)b, (s3cipid_t)l, (s3cipid_t)r, WORD_POSN_INTERNAL);
		internal[i] = mdef_pid2ssid(mdef, p);
	    }
	    
	    l = b;
	    b = r;
	    if (NOT_S3SSID(rdiph[b][l])) {
		g = rdiph_comsseq(mdef, b, l);
		rdiph[b][l] = ssidlist2comsseq (g, mdef, dict2pid, hs, hp);
		glist_free (g);
	    }
	    internal[pronlen-1] = rdiph[b][l];
	} else if (pronlen == 1) {
	    b = dict_pron(dict, w, 0);
	    if (NOT_S3SSID(single[b])) {
		g = single_comsseq(mdef, b);
		single[b] = ssidlist2comsseq (g, mdef, dict2pid, hs, hp);
		glist_free (g);
		
		for (l = 0; l < mdef_n_ciphone(mdef); l++) {
		    g = single_lc_comsseq(mdef, b, l);
		    dict2pid->single_lc[b][l] = ssidlist2comsseq (g, mdef, dict2pid, hs, hp);
		    glist_free (g);
		}
	    }
	    internal[0] = single[b];
	}
	
	internal += pronlen;
    }
    
    ckd_free_2d ((void **) ldiph);
    ckd_free_2d ((void **) rdiph);
    ckd_free ((void *) single);
    
    /* Allocate space for composite state table */
    cslen = (int32 *) ckd_calloc (dict2pid->n_comstate, sizeof(int32));
    g = hash_tolist(hs, &n);
    assert (n == dict2pid->n_comstate);
    n = 0;
    for (gn = g; gn; gn = gnode_next(gn)) {
	he = (hash_entry_t *) gnode_ptr (gn);
	sen = (s3senid_t *) hash_entry_key(he);
	for (i = 0; IS_S3SENID(sen[i]); i++);
	
	cslen[hash_entry_val(he)] = i+1;	/* +1 for terminating sentinel */
	
	n += (i+1);
    }
    dict2pid->comstate = (s3senid_t **) ckd_calloc (dict2pid->n_comstate, sizeof(s3senid_t *));
    sen = (s3senid_t *) ckd_calloc (n, sizeof(s3senid_t));
    for (i = 0; i < dict2pid->n_comstate; i++) {
	dict2pid->comstate[i] = sen;
	sen += cslen[i];
    }
    
    /* Build composite state table from hash table hs */
    for (gn = g; gn; gn = gnode_next(gn)) {
	he = (hash_entry_t *) gnode_ptr (gn);
	sen = (s3senid_t *) hash_entry_key(he);
	i = hash_entry_val(he);
	
	for (j = 0; j < cslen[i]; j++)
	    dict2pid->comstate[i][j] = sen[j];
	assert (sen[j-1] == BAD_S3SENID);

	ckd_free ((void *)sen);
    }
    ckd_free (cslen);
    glist_free (g);
    hash_free (hs);
    
    /* Allocate space for composite sseq table */
    dict2pid->comsseq = (s3senid_t **) ckd_calloc (dict2pid->n_comsseq, sizeof(s3senid_t *));
    g = hash_tolist (hp, &n);
    assert (n == dict2pid->n_comsseq);
    
    /* Build composite sseq table */
    for (gn = g; gn; gn = gnode_next(gn)) {
	he = (hash_entry_t *) gnode_ptr (gn);
	i = hash_entry_val(he);
	dict2pid->comsseq[i] = (s3senid_t *) hash_entry_key(he);
    }
    glist_free (g);
    hash_free (hp);
    
    /* Weight for each composite state */
    dict2pid->comwt = (int32 *) ckd_calloc (dict2pid->n_comstate, sizeof(int32));
    for (i = 0; i < dict2pid->n_comstate; i++) {
	sen = dict2pid->comstate[i];
	
	for (j = 0; IS_S3SENID(sen[j]); j++);
#if 0
	/* if comstate i has N states, its weight= (1/N^2) (Major Hack!!) */
	dict2pid->comwt[i] = - (logs3 ((float64)j) << 1);
#else
	/* if comstate i has N states, its weight= 1/N */
	dict2pid->comwt[i] = - logs3 ((float64)j);
#endif
    }
    
    E_INFO("%d composite states; %d composite sseq\n",
	   dict2pid->n_comstate, dict2pid->n_comsseq);
    
    return dict2pid;
}