void vithist_rescore (vithist_t *vh, kbcore_t *kbc, s3wid_t wid, int32 ef, int32 score, int32 pred, int32 type) { vithist_entry_t *pve, tve; s3lmwid_t lwid; int32 se, fe; int32 i; assert (vh->n_frm == ef); pve = vh->entry[VITHIST_ID2BLK(pred)] + VITHIST_ID2BLKOFFSET(pred); /* Create a temporary entry with all the info currently available */ tve.wid = wid; tve.sf = pve->ef + 1; tve.ef = ef; tve.type = type; tve.valid = 1; tve.ascr = score - pve->score; if (pred == 0) { /* Special case for the initial <s> entry */ se = 0; fe = 1; } else { se = vh->frame_start[pve->ef]; fe = vh->frame_start[pve->ef + 1]; } if (dict_filler_word (kbcore_dict(kbc), wid)) { tve.lscr = fillpen (kbcore_fillpen(kbc), wid); tve.score = score + tve.lscr; tve.pred = pred; tve.lmstate.lm3g = pve->lmstate.lm3g; vithist_enter (vh, kbc, &tve); } else { lwid = kbcore_dict2lmwid (kbc, wid); tve.lmstate.lm3g.lwid[0] = lwid; for (i = se; i < fe; i++) { pve = vh->entry[VITHIST_ID2BLK(i)] + VITHIST_ID2BLKOFFSET(i); if (pve->valid) { tve.lscr = lm_tg_score (kbcore_lm(kbc), pve->lmstate.lm3g.lwid[1], pve->lmstate.lm3g.lwid[0], lwid); tve.score = pve->score + tve.ascr + tve.lscr; if ((tve.score - vh->wbeam) >= vh->bestscore[vh->n_frm]) { tve.pred = i; tve.lmstate.lm3g.lwid[1] = pve->lmstate.lm3g.lwid[0]; vithist_enter (vh, kbc, &tve); } } } } }
void vithist_rescore(vithist_t * vh, ngram_model_t *lm, s3dict_t *dict, dict2pid_t *dict2pid, fillpen_t *fp, s3wid_t wid, int32 ef, int32 score, int32 pred, int32 type, int32 rc) { vithist_entry_t *pve, tve; int32 lwid; int32 se, fe; int32 i; assert(vh->n_frm == ef); if (pred == -1) { /* Always do E_FATAL assuming upper level function take care of error checking. */ E_FATAL ("Hmm->out.history equals to -1 with score %d, some active phone was not computed?\n", score); } /* pve is the previous entry before word with wid or, se an fe is the first to the last entry before pve. So pve is w_{n-1} */ pve = vithist_id2entry(vh, pred); /* Create a temporary entry with all the info currently available */ tve.wid = wid; tve.sf = pve->ef + 1; tve.ef = ef; tve.type = type; tve.valid = 1; tve.ascr = score - pve->path.score; tve.lscr = 0; tve.rc = NULL; tve.n_rc = 0; /* Filler words only have unigram language model scores, so not * much special needs to be done for them. vithist_prune() is * going to prune out most of these later on, anyway. */ if (s3dict_filler_word(dict, wid)) { tve.path.score = score; tve.lscr = fillpen(fp, wid); tve.path.score += tve.lscr; if ((tve.path.score - vh->wbeam) >= vh->bestscore[vh->n_frm]) { tve.path.pred = pred; /* Note that they just propagate the same LM state since * they are not in the LM. */ tve.lmstate.lm3g = pve->lmstate.lm3g; vithist_enter(vh, dict, dict2pid, &tve, rc); } } else { if (pred == 0) { /* Special case for the initial <s> entry */ se = 0; fe = 1; } else { se = vh->frame_start[pve->ef]; fe = vh->frame_start[pve->ef + 1]; } /* Now if it is a word, backtrack again to get all possible previous word So pve becomes the w_{n-2}. */ lwid = ngram_wid(lm, s3dict_wordstr(dict, s3dict_basewid(dict, wid))); tve.lmstate.lm3g.lwid[0] = lwid; /* FIXME: This loop is completely awful. For each entry in * this frame, we scan every entry in the previous frame, * potentially creating a new history entry. This means that * without pruning, the size of the vithist table (and thus * the time taken here) is exponential in the number of * frames! */ for (i = se; i < fe; i++) { pve = vithist_id2entry(vh, i); if (pve->valid) { int n_used; tve.path.score = pve->path.score + tve.ascr; /* Try at all costs to avoid calling ngram_tg_score() * because it is the main time consuming part here * (but as noted above... ugh...) See below as well. */ if ((tve.path.score - vh->wbeam) < vh->bestscore[vh->n_frm]) continue; /* The trigram cache is supposed to make this fast, * but due to the crazy number of times this could be * called, it's still slow compared to a hash * table. */ tve.lscr = ngram_tg_score(lm, lwid, pve->lmstate.lm3g.lwid[0], pve->lmstate.lm3g.lwid[1], &n_used); tve.path.score += tve.lscr; /* A different word exit threshold - we would have to * be inside the general word beam in order to get * here, now we apply a second beam to the *vithist * entries* in this frame. There can be an ungodly * number of them for reasons that aren't entirely * clear to me, so this is kind of a pre-pruning. * NOTE: the "backwards" math here is because * vh->bestscore is frequently MAX_NEG_INT32. ALSO * NOTE: We can't precompute the threshold since the * best score will be updated by vithist_enter(). */ if ((tve.path.score - vh->wbeam) >= vh->bestscore[vh->n_frm]) { tve.path.pred = i; tve.lmstate.lm3g.lwid[1] = pve->lmstate.lm3g.lwid[0]; vithist_enter(vh, dict, dict2pid, &tve, rc); } } } } }
/*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); } }
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); } }
/* 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 */ }