int32
s3_cd_gmm_compute_sen_comp(void *srch, float32 ** feat, int32 wav_idx)
{
int32 flag;
srch_t *s;
ascr_t *ascr;
kbcore_t *kbcore;
s = (srch_t *) srch;
kbcore = s->kbc;
ascr = s->ascr;
assert(kbcore->ms_mgau || kbcore->mgau || kbcore->s2_mgau);
assert(!(kbcore->ms_mgau && kbcore->mgau && kbcore->s2_mgau));
flag = s3_cd_gmm_compute_sen(srch, feat, wav_idx);
if (flag != SRCH_SUCCESS) {
E_INFO("Computation of senone failed\n");
return flag;
}
/* Evaluate composite senone scores from senone scores */
dict2pid_comsenscr(kbcore_dict2pid(kbcore), ascr->senscr,
ascr->comsen);
return SRCH_SUCCESS;
}
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);
}
}
}
/*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 */
}
void
gmm_compute(void *data, utt_res_t * ur, int32 sf, int32 ef, char *uttid)
{
kb_t *kb;
kbcore_t *kbcore;
mdef_t *mdef;
dict_t *dict;
dict2pid_t *d2p;
mgau_model_t *mgau;
subvq_t *svq;
gs_t *gs;
int32 ptranskip;
int32 s, f, t;
int32 single_el_list[2];
stats_t cur_ci_st;
stats_t cur_cd_st;
stats_t cur_cd_Nbest_st;
stats_t *stptr;
char str[100];
int32 *idx;
int32 *cur_bstidx;
int32 *last_bstidx;
int32 *cur_scr;
int32 *last_scr;
int32 tmpint;
s3senid_t *cd2cisen;
int32 pheurtype;
E_INFO("Processing: %s\n", uttid);
kb = (kb_t *) data;
kbcore = kb->kbcore;
mdef = kbcore_mdef(kbcore);
dict = kbcore_dict(kbcore);
d2p = kbcore_dict2pid(kbcore);
mgau = kbcore_mgau(kbcore);
svq = kbcore_svq(kbcore);
gs = kbcore_gs(kbcore);
kb->uttid = uttid;
ptranskip = kb->beam->ptranskip;
pheurtype = kb->pl->pheurtype;
single_el_list[0] = -1;
single_el_list[1] = -1;
/* Read mfc file and build feature vectors for entire utterance */
kb->stat->nfr =
feat_s2mfc2feat(kbcore_fcb(kbcore), ur->uttfile,
cmd_ln_str("-cepdir"), ".mfc", sf, ef, kb->feat,
S3_MAX_FRAMES);
cd2cisen = mdef_cd2cisen(mdef);
/*This should be a procedure instead of just logic */
init_stat(&cur_cd_st, "Current CD Senone");
init_stat(&cur_ci_st, "Current CI Senone");
init_stat(&cur_cd_Nbest_st, "Current CD NBest Senone");
for (s = 0; s < mdef->n_ci_sen; s++) {
sprintf(str, "Cur Senone %d", s);
init_stat(&cur_sen_st[s], str);
}
for (t = 0; t < (int32) (mdef->n_sen - mdef->n_ci_sen) / NBEST_STEP;
t++) {
sprintf(str, " %d -Cur Best Senone", t * NBEST_STEP);
init_stat(&cur_sen_Nbest_st[t], str);
}
idx = ckd_calloc(mdef->n_sen - mdef->n_ci_sen, sizeof(int32));
/* Allocate temporary array for CurScr and Curbst indx and Lat index */
cur_bstidx = ckd_calloc(mdef->n_sen, sizeof(int32));
last_bstidx = ckd_calloc(mdef->n_sen, sizeof(int32));
cur_scr = ckd_calloc(mdef->n_sen, sizeof(int32));
last_scr = ckd_calloc(mdef->n_sen, sizeof(int32));
for (f = 0; f < kb->stat->nfr; f++) {
for (s = 0; s < mgau->n_mgau; s++) {
/*1, Compute the approximate scores with the last best index. */
if (mgau->mgau[s].bstidx != NO_BSTIDX) {
single_el_list[0] = mgau->mgau[s].bstidx;
last_bstidx[s] = mgau->mgau[s].bstidx;
last_scr[s] =
mgau_eval(mgau, s, single_el_list, kb->feat[f][0], f,
0);
}
else {
last_bstidx[s] = NO_BSTIDX;
}
/*2, Compute the exact scores and sort them and get the ranking. */
kb->ascr->senscr[s] =
mgau_eval(mgau, s, NULL, kb->feat[f][0], f, 1);
/*3, Compute the approximate scores with the current best index */
if (mgau->mgau[s].bstidx != NO_BSTIDX) {
single_el_list[0] = mgau->mgau[s].bstidx;
cur_bstidx[s] = mgau->mgau[s].bstidx;
cur_scr[s] =
mgau_eval(mgau, s, single_el_list, kb->feat[f][0], f,
0);
}
else {
cur_bstidx[s] = NO_BSTIDX;
}
/* Only test for CD senones, test for best index hit */
/*Update either CI senone and CD senone) */
if (!mdef_is_cisenone(mdef, s))
stptr = &cur_cd_st;
else
stptr = &cur_ci_st;
increment_stat(stptr,
abs(last_scr[s] - kb->ascr->senscr[s]),
abs(cur_scr[s] - kb->ascr->senscr[s]),
abs(kb->ascr->senscr[cd2cisen[s]] -
kb->ascr->senscr[s]),
(cur_bstidx[s] == last_bstidx[s]));
if (!mdef_is_cisenone(mdef, s)) {
stptr = &cur_sen_st[cd2cisen[s]];
increment_stat(stptr,
abs(last_scr[s] - kb->ascr->senscr[s]),
abs(cur_scr[s] - kb->ascr->senscr[s]),
abs(kb->ascr->senscr[cd2cisen[s]] -
kb->ascr->senscr[s]),
(cur_bstidx[s] == last_bstidx[s]));
stptr->total_senone += 1;
}
}
cur_cd_st.total_fr++;
cur_cd_st.total_senone += mdef->n_sen - mdef->n_ci_sen;
cur_ci_st.total_fr++;
cur_ci_st.total_senone += mdef->n_ci_sen;
for (s = 0; s < mdef->n_ci_sen; s++) {
cur_sen_st[s].total_fr++;
}
/*This is the part we need to do sorting */
/*1, sort the scores in the current frames */
/*E_INFO("At frame %d\n",f); */
/*Pointer trick at here. */
for (s = 0; s < mdef->n_sen - mdef->n_ci_sen; s++) {
idx[s] = s;
}
cd = &(kb->ascr->senscr[mdef->n_ci_sen]);
qsort(idx, mdef->n_sen - mdef->n_ci_sen, sizeof(int32),
intcmp_gmm_compute);
/*This loop is stupid and it is just a hack. */
for (s = 0; s < mdef->n_sen - mdef->n_ci_sen; s++) {
tmpint = idx[s] + mdef->n_ci_sen;
for (t = 0;
t <
(int32) ((float) (mdef->n_sen - mdef->n_ci_sen) /
(float) NBEST_STEP); t++) {
if (s < t * NBEST_STEP) {
increment_stat(&cur_sen_Nbest_st[t],
abs(last_scr[tmpint] -
kb->ascr->senscr[tmpint]),
abs(cur_scr[tmpint] -
kb->ascr->senscr[tmpint]),
abs(kb->ascr->senscr[cd2cisen[tmpint]] -
kb->ascr->senscr[tmpint]),
(cur_bstidx[tmpint] ==
last_bstidx[tmpint]));
cur_sen_Nbest_st[t].total_senone += 1;
}
}
}
for (t = 0;
t <
(int32) ((float) (mdef->n_sen - mdef->n_ci_sen) /
(float) NBEST_STEP); t++) {
cur_sen_Nbest_st[t].total_fr++;
}
}
print_stat(&cur_cd_st);
print_stat(&cur_ci_st);
print_stat(&cur_sen_Nbest_st[1]); /*Only show the first NBEST_STEP best */
add_stat(&cd_st, &cur_cd_st);
add_stat(&ci_st, &cur_ci_st);
for (s = 0; s < mdef->n_ci_sen; s++) {
add_stat(&sen_st[s], &cur_sen_st[s]);
}
for (s = 0; s < (int32) (mdef->n_sen - mdef->n_ci_sen) / NBEST_STEP;
s++) {
add_stat(&sen_Nbest_st[s], &cur_sen_Nbest_st[s]);
}
ckd_free(idx);
ckd_free(cur_bstidx);
ckd_free(last_bstidx);
ckd_free(cur_scr);
ckd_free(last_scr);
}
