void
match_write(FILE * fp, glist_t hyp, char *uttid, dict_t * dict, char *hdr)
{
gnode_t *gn;
srch_hyp_t *h;
int counter = 0;
if (fp == NULL)
return;
if (hyp == NULL) /* Following s3.0 convention */
fprintf(fp, "(null)");
fprintf(fp, "%s", (hdr ? hdr : ""));
/* for (gn = hyp; gn && (gnode_next(gn)); gn = gnode_next(gn)) { */
for (gn = hyp; gn; gn = gnode_next(gn)) {
h = (srch_hyp_t *) gnode_ptr(gn);
if (h->sf != h->ef) { /* FSG outputs zero-width hyps */
if ((!dict_filler_word(dict, h->id))
&& (h->id != dict_finishwid(dict))
&& (h->id != dict_startwid(dict)))
fprintf(fp, "%s ",
dict_wordstr(dict, dict_basewid(dict, h->id)));
counter++;
}
}
if (counter == 0)
fprintf(fp, " ");
fprintf(fp, "(%s)\n", uttid);
fflush(fp);
}
void
ps_search_init(ps_search_t *search, ps_searchfuncs_t *vt,
cmd_ln_t *config, acmod_t *acmod, dict_t *dict,
dict2pid_t *d2p)
{
search->vt = vt;
search->config = config;
search->acmod = acmod;
if (d2p)
search->d2p = dict2pid_retain(d2p);
else
search->d2p = NULL;
if (dict) {
search->dict = dict_retain(dict);
search->start_wid = dict_startwid(dict);
search->finish_wid = dict_finishwid(dict);
search->silence_wid = dict_silwid(dict);
search->n_words = dict_size(dict);
}
else {
search->dict = NULL;
search->start_wid = search->finish_wid = search->silence_wid = -1;
search->n_words = 0;
}
}
void
ps_search_base_reinit(ps_search_t *search, dict_t *dict,
dict2pid_t *d2p)
{
dict_free(search->dict);
dict2pid_free(search->d2p);
/* FIXME: _retain() should just return NULL if passed NULL. */
if (dict) {
search->dict = dict_retain(dict);
search->start_wid = dict_startwid(dict);
search->finish_wid = dict_finishwid(dict);
search->silence_wid = dict_silwid(dict);
search->n_words = dict_size(dict);
}
else {
search->dict = NULL;
search->start_wid = search->finish_wid = search->silence_wid = -1;
search->n_words = 0;
}
if (d2p)
search->d2p = dict2pid_retain(d2p);
else
search->d2p = NULL;
}
int
s3_decode_record_hyps(s3_decode_t * _decode, int _end_utt)
{
int32 i = 0;
glist_t hyp_list;
gnode_t *node;
srch_hyp_t *hyp;
char *hyp_strptr = 0;
char *hyp_str = 0;
srch_t *srch;
srch_hyp_t **hyp_segs = 0;
int hyp_seglen = 0;
int hyp_strlen = 0;
int finish_wid = 0;
kb_t *kb = 0;
dict_t *dict;
int rv;
if (_decode == NULL)
return S3_DECODE_ERROR_NULL_POINTER;
s3_decode_free_hyps(_decode);
kb = &_decode->kb;
dict = kbcore_dict(_decode->kbcore);
srch = (srch_t *) _decode->kb.srch;
hyp_list = srch_get_hyp(srch);
if (hyp_list == NULL) {
E_WARN("Failed to retrieve viterbi history.\n");
return S3_DECODE_ERROR_INTERNAL;
}
/** record the segment length and the overall string length */
finish_wid = dict_finishwid(dict);
for (node = hyp_list; node != NULL; node = gnode_next(node)) {
hyp = (srch_hyp_t *) gnode_ptr(node);
hyp_seglen++;
if (!dict_filler_word(dict, hyp->id) && hyp->id != finish_wid) {
hyp_strlen +=
strlen(dict_wordstr(dict, dict_basewid(dict, hyp->id))) +
1;
}
}
if (hyp_strlen == 0) {
hyp_strlen = 1;
}
/** allocate array to hold the segments and/or decoded string */
hyp_str = (char *) ckd_calloc(hyp_strlen, sizeof(char));
hyp_segs =
(srch_hyp_t **) ckd_calloc(hyp_seglen + 1, sizeof(srch_hyp_t *));
if (hyp_segs == NULL || hyp_str == NULL) {
E_WARN("Failed to allocate storage for hypothesis.\n");
rv = S3_DECODE_ERROR_OUT_OF_MEMORY;
goto s3_decode_record_hyps_cleanup;
}
/** iterate thru to fill in the array of segments and/or decoded string */
i = 0;
hyp_strptr = hyp_str;
for (node = hyp_list; node != NULL; node = gnode_next(node), i++) {
hyp = (srch_hyp_t *) gnode_ptr(node);
hyp_segs[i] = hyp;
hyp->word = dict_wordstr(dict, dict_basewid(dict, hyp->id));
if (!dict_filler_word(dict, hyp->id) && hyp->id != finish_wid) {
strcat(hyp_strptr,
dict_wordstr(dict, dict_basewid(dict, hyp->id)));
hyp_strptr += strlen(hyp_strptr);
*hyp_strptr = ' ';
hyp_strptr += 1;
}
}
glist_free(hyp_list);
hyp_str[hyp_strlen - 1] = '\0';
hyp_segs[hyp_seglen] = 0;
_decode->hyp_frame_num = _decode->num_frames_decoded;
_decode->hyp_segs = hyp_segs;
_decode->hyp_str = hyp_str;
return S3_DECODE_SUCCESS;
s3_decode_record_hyps_cleanup:
if (hyp_segs != NULL) {
ckd_free(hyp_segs);
}
if (hyp_str != NULL) {
ckd_free(hyp_str);
}
if (hyp_list != NULL) {
for (node = hyp_list; node != NULL; node = gnode_next(node)) {
if ((hyp = (srch_hyp_t *) gnode_ptr(node)) != NULL) {
ckd_free(hyp);
}
}
glist_free(hyp_list);
}
return rv;
}
int
ld_utt_hyps(live_decoder_t *decoder, char **hyp_str, hyp_t ***hyp_segs)
{
int32 id;
int32 i = 0;
glist_t hyp_list;
gnode_t *node;
hyp_t *hyp;
dict_t *dict;
char *hyp_strptr;
kb_t *kb = &decoder->kb;
if (decoder->ld_state == LD_STATE_ENDED) {
if (hyp_segs) {
*hyp_segs = kb->hyp_segs;
}
if (hyp_str) {
*hyp_str = kb->hyp_str;
}
return 0;
}
else {
kb_freehyps(kb);
}
dict = kbcore_dict (decoder->kbcore);
id = vithist_partialutt_end(kb->vithist, decoder->kbcore);
if (id >= 0) {
hyp_list = vithist_backtrace(kb->vithist, id);
/* record the segment length and the overall string length */
for (node = hyp_list; node; node = gnode_next(node)) {
hyp = (hyp_t *)gnode_ptr(node);
if (hyp_segs) {
kb->hyp_seglen++;
}
if (hyp_str) {
if (!dict_filler_word(dict, hyp->id) &&
hyp->id != dict_finishwid(dict)) {
kb->hyp_strlen +=
strlen(dict_wordstr(dict, dict_basewid(dict, hyp->id))) + 1;
}
}
}
/* allocate array to hold the segments and/or decoded string */
if (hyp_segs) {
kb->hyp_segs = (hyp_t **)ckd_calloc(kb->hyp_seglen, sizeof(hyp_t *));
}
if (hyp_str) {
kb->hyp_str = (char *)ckd_calloc(kb->hyp_strlen+1, sizeof(char));
}
/* iterate thru to fill in the array of segments and/or decoded string */
i = 0;
if (hyp_str) {
hyp_strptr = kb->hyp_str;
}
for (node = hyp_list; node; node = gnode_next(node), i++) {
hyp = (hyp_t *)gnode_ptr(node);
if (hyp_segs) {
kb->hyp_segs[i] = hyp;
}
if (hyp_str) {
strcat(hyp_strptr, dict_wordstr(dict, dict_basewid(dict, hyp->id)));
hyp_strptr += strlen(hyp_strptr);
strcat(hyp_strptr, " ");
hyp_strptr += 1;
}
}
glist_free(hyp_list);
if (hyp_str) {
kb->hyp_str[kb->hyp_strlen - 1] = '\0';
}
}
if (hyp_segs) {
*hyp_segs = kb->hyp_segs;
}
if (hyp_str) {
*hyp_str = kb->hyp_str;
}
return 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);
}
}
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);
}
}
/*
* Compute the left and right context CIphone sets for each state.
* (Needed for building the phone HMM net using cross-word triphones. Invoke
* after computing null transitions closure.)
*/
static void
word_fsg_lc_rc(word_fsg_t * fsg)
{
int32 s, d, i, j;
int32 n_ci;
gnode_t *gn;
word_fsglink_t *l;
int32 silcipid;
int32 endwid;
int32 len;
dict_t *dict;
mdef_t *mdef;
dict = fsg->dict;
mdef = fsg->mdef;
assert(fsg);
assert(dict);
assert(mdef);
endwid = dict_basewid(dict, dict_finishwid(dict));
silcipid = mdef_silphone(mdef);
assert(silcipid >= 0);
E_INFO("Value of silcipid %d\n", silcipid);
n_ci = fsg->n_ciphone;
E_INFO("No of CI phones %d\n", n_ci);
if (n_ci > 127) {
E_FATAL
("#phones(%d) > 127; cannot use int8** for word_fsg_t.{lc,rc}\n",
n_ci);
}
/*
* fsg->lc[s] = set of left context CIphones for state s. Similarly, rc[s]
* for right context CIphones.
*/
fsg->lc =
(int8 **) ckd_calloc_2d(fsg->n_state, n_ci + 1, sizeof(int8));
fsg->rc =
(int8 **) ckd_calloc_2d(fsg->n_state, n_ci + 1, sizeof(int8));
for (s = 0; s < fsg->n_state; s++) {
for (d = 0; d < fsg->n_state; d++) {
for (gn = fsg->trans[s][d]; gn; gn = gnode_next(gn)) {
l = (word_fsglink_t *) gnode_ptr(gn);
assert(l->wid >= 0);
/*
* Add the first CIphone of l->wid to the rclist of state s, and
* the last CIphone to lclist of state d.
* (Filler phones are a pain to deal with. There is no direct
* marking of a filler phone; but only filler words are supposed to
* use such phones, so we use that fact. HACK!! FRAGILE!!)
*/
if (dict_filler_word(dict, l->wid) || (l->wid == endwid)) {
/* Filler phone; use silence phone as context */
fsg->rc[s][silcipid] = 1;
fsg->lc[d][silcipid] = 1;
}
else {
len = dict_pronlen(dict, l->wid);
fsg->rc[s][dict_pron(dict, l->wid, 0)] = 1;
fsg->lc[d][dict_pron(dict, l->wid, len - 1)] = 1;
}
}
}
/*
* Add SIL phone to the lclist and rclist of each state. Strictly
* speaking, only needed at start and final states, respectively, but
* all states considered since the user may change the start and final
* states. In any case, most applications would have a silence self
* loop at each state, hence these would be needed anyway.
*/
fsg->lc[s][silcipid] = 1;
fsg->rc[s][silcipid] = 1;
}
/*
* Propagate lc and rc lists past null transitions. (Since FSG contains
* null transitions closure, no need to worry about a chain of successive
* null transitions. Right??)
*/
for (s = 0; s < fsg->n_state; s++) {
for (d = 0; d < fsg->n_state; d++) {
l = fsg->null_trans[s][d];
if (l) {
/*
* lclist(d) |= lclist(s), because all the words ending up at s, can
* now also end at d, becoming the left context for words leaving d.
*/
for (i = 0; i < n_ci; i++)
fsg->lc[d][i] |= fsg->lc[s][i];
/*
* Similarly, rclist(s) |= rclist(d), because all the words leaving d
* can equivalently leave s, becoming the right context for words
* ending up at s.
*/
for (i = 0; i < n_ci; i++)
fsg->rc[s][i] |= fsg->rc[d][i];
}
}
}
/* Convert the bit-vector representation into a list */
for (s = 0; s < fsg->n_state; s++) {
j = 0;
for (i = 0; i < n_ci; i++) {
if (fsg->lc[s][i]) {
fsg->lc[s][j] = i;
j++;
}
}
fsg->lc[s][j] = -1; /* Terminate the list */
j = 0;
for (i = 0; i < n_ci; i++) {
if (fsg->rc[s][i]) {
fsg->rc[s][j] = i;
j++;
}
}
fsg->rc[s][j] = -1; /* Terminate the list */
}
}
/* 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 */
}
/*
* Header written BEFORE this function is called.
*/
void vithist_dag_write (vithist_t *vh, glist_t hyp, dict_t *dict, int32 oldfmt, FILE *fp)
{
glist_t *sfwid; /* To maintain <start-frame, word-id> pair dagnodes */
vithist_entry_t *ve, *ve2;
gnode_t *gn, *gn2, *gn3;
dagnode_t *dn, *dn2;
int32 sf, ef, n_node;
int32 f, i;
hyp_t *h;
sfwid = (glist_t *) ckd_calloc (vh->n_frm+1, sizeof(glist_t));
n_node = 0;
for (i = 0; i < vh->n_entry; i++) { /* This range includes the dummy <s> and </s> entries */
ve = vh->entry[VITHIST_ID2BLK(i)] + VITHIST_ID2BLKOFFSET(i);
if (! ve->valid)
continue;
/*
* The initial <s> entry (at 0) is a dummy, with start/end frame = -1. But the old S3
* code treats it like a real word, so we have to reintroduce it in the dag file with
* a start time of 0. And shift the start time of words starting at frame 0 up by 1.
* MAJOR HACK!!
*/
if (ve->sf <= 0) {
assert (ve->sf >= -1);
assert ((ve->ef == -1) || (ve->ef > 1));
sf = ve->sf + 1;
ef = (ve->ef < 0) ? 0 : ve->ef;
} else {
sf = ve->sf;
ef = ve->ef;
}
for (gn = sfwid[sf]; gn; gn = gnode_next(gn)) {
dn = (dagnode_t *) gnode_ptr(gn);
if (dn->wid == ve->wid)
break;
}
if (! gn) {
dn = (dagnode_t *) ckd_calloc (1, sizeof(dagnode_t));
dn->wid = ve->wid;
dn->fef = ef;
dn->lef = ef;
dn->seqid = -1; /* Initially all invalid, selected ones validated below */
dn->velist = NULL;
n_node++;
sfwid[sf] = glist_add_ptr (sfwid[sf], (void *) dn);
} else {
dn->lef = ef;
}
/*
* Check if an entry already exists under dn->velist (generated by a different
* LM context; retain only the best scoring one.
*/
for (gn = dn->velist; gn; gn = gnode_next(gn)) {
ve2 = (vithist_entry_t *) gnode_ptr (gn);
if (ve2->ef == ve->ef)
break;
}
if (gn) {
if (ve->score > ve2->score)
gnode_ptr(gn) = (void *)ve;
} else
dn->velist = glist_add_ptr (dn->velist, (void *) ve);
}
/*
* Validate segments with >1 end times; if only 1 end time, can be pruned.
* But keep segments in the original hypothesis, regardless; mark them first.
*/
for (gn = hyp; gn; gn = gnode_next(gn)) {
h = (hyp_t *) gnode_ptr (gn);
for (gn2 = sfwid[h->sf]; gn2; gn2 = gnode_next(gn2)) {
dn = (dagnode_t *) gnode_ptr (gn2);
if (h->id == dn->wid)
dn->seqid = 0; /* Do not discard (prune) this dagnode */
}
}
/* Validate startwid and finishwid nodes */
dn = (dagnode_t *) gnode_ptr(sfwid[0]);
assert (dn->wid == dict_startwid(dict));
dn->seqid = 0;
dn = (dagnode_t *) gnode_ptr(sfwid[vh->n_frm]);
assert (dn->wid == dict_finishwid(dict));
dn->seqid = 0;
/* Now prune dagnodes with only 1 end frame if not validated above */
i = 0;
for (f = vh->n_frm; f >= 0; --f) {
for (gn = sfwid[f]; gn; gn = gnode_next(gn)) {
dn = (dagnode_t *) gnode_ptr(gn);
if ((dn->lef > dn->fef) || (dn->seqid >= 0))
dn->seqid = i++;
else
dn->seqid = -1; /* Flag: discard */
}
}
n_node = i;
/* Write nodes info; the header should have been written before this function is called */
fprintf (fp, "Nodes %d (NODEID WORD STARTFRAME FIRST-ENDFRAME LAST-ENDFRAME)\n", n_node);
for (f = vh->n_frm; f >= 0; --f) {
for (gn = sfwid[f]; gn; gn = gnode_next(gn)) {
dn = (dagnode_t *) gnode_ptr(gn);
if (dn->seqid >= 0) {
fprintf (fp, "%d %s %d %d %d\n",
dn->seqid, dict_wordstr(dict, dn->wid), f, dn->fef, dn->lef);
}
}
}
fprintf (fp, "#\n");
fprintf (fp, "Initial %d\nFinal %d\n", n_node-1, 0);
fprintf (fp, "#\n");
fprintf (fp, "BestSegAscr 0 (NODEID ENDFRAME ASCORE)\n");
fprintf (fp, "#\n");
/* Edges */
if (oldfmt)
fprintf (fp, "Edges (FROM-NODEID TO-NODEID ASCORE)\n");
else
fprintf (fp, "Edges (FROM-NODEID ENDFRAME ASCORE)\n");
for (f = vh->n_frm-1; f >= 0; --f) {
for (gn = sfwid[f]; gn; gn = gnode_next(gn)) {
dn = (dagnode_t *) gnode_ptr(gn);
/* Look for transitions from this dagnode to later ones, if not discarded */
if (dn->seqid < 0)
continue;
for (gn2 = dn->velist; gn2; gn2 = gnode_next(gn2)) {
ve = (vithist_entry_t *) gnode_ptr (gn2);
sf = (ve->ef < 0) ? 1 : (ve->ef + 1);
if (oldfmt) {
for (gn3 = sfwid[sf]; gn3; gn3 = gnode_next(gn3)) {
dn2 = (dagnode_t *) gnode_ptr(gn3);
if (dn2->seqid >= 0)
fprintf (fp, "%d %d %d\n", dn->seqid, dn2->seqid, ve->ascr);
}
} else {
for (gn3 = sfwid[sf]; gn3; gn3 = gnode_next(gn3)) {
dn2 = (dagnode_t *) gnode_ptr(gn3);
if (dn2->seqid >= 0) {
fprintf (fp, "%d %d %d\n", dn->seqid, sf-1, ve->ascr);
break;
}
}
}
}
}
}
fprintf (fp, "End\n");
/* Free dagnodes structure */
for (f = 0; f <= vh->n_frm; f++) {
for (gn = sfwid[f]; gn; gn = gnode_next(gn)) {
dn = (dagnode_t *) gnode_ptr(gn);
glist_free (dn->velist);
ckd_free ((void *) dn);
}
glist_free (sfwid[f]);
}
ckd_free ((void *) sfwid);
}
int32 vithist_utt_end (vithist_t *vh, kbcore_t *kbc)
{
int32 f, i, b, l;
int32 sv, nsv, scr, bestscore, bestvh, vhid;
vithist_entry_t *ve, *bestve=0;
s3lmwid_t endwid;
lm_t *lm;
dict_t *dict;
/* Find last frame with entries in vithist table */
for (f = vh->n_frm-1; f >= 0; --f) {
sv = vh->frame_start[f]; /* First vithist entry in frame f */
nsv = vh->frame_start[f+1]; /* First vithist entry in next frame (f+1) */
if (sv < nsv)
break;
}
if (f < 0)
return -1;
if (f != vh->n_frm-1)
E_ERROR("No word exit in frame %d, using exits from frame %d\n", vh->n_frm-1, f);
/* Terminate in a final </s> node (make this optional?) */
lm = kbcore_lm (kbc);
endwid = lm_finishwid (lm);
bestscore = MAX_NEG_INT32;
bestvh = -1;
for (i = sv; i < nsv; i++) {
b = VITHIST_ID2BLK (i);
l = VITHIST_ID2BLKOFFSET (i);
ve = vh->entry[b] + l;
scr = ve->score;
scr += lm_tg_score (lm, ve->lmstate.lm3g.lwid[1], ve->lmstate.lm3g.lwid[0], endwid);
if (bestscore < scr) {
bestscore = scr;
bestvh = i;
bestve = ve;
}
}
assert (bestvh >= 0);
dict = kbcore_dict (kbc);
if (f != vh->n_frm-1) {
E_ERROR("No word exit in frame %d, using exits from frame %d\n", vh->n_frm-1, f);
/* Add a dummy silwid covering the remainder of the utterance */
assert (vh->frame_start[vh->n_frm-1] == vh->frame_start[vh->n_frm]);
vh->n_frm -= 1;
vithist_rescore (vh, kbc, dict_silwid (dict), vh->n_frm, bestve->score, bestvh, -1);
vh->n_frm += 1;
vh->frame_start[vh->n_frm] = vh->n_entry;
return vithist_utt_end (vh, kbc);
}
/* Create an </s> entry */
vhid = vh->n_entry;
ve = vithist_entry_alloc (vh);
ve->wid = dict_finishwid (dict);
ve->sf = (bestve->ef == BAD_S3FRMID) ? 0 : bestve->ef + 1;
ve->ef = vh->n_frm;
ve->ascr = 0;
ve->lscr = bestscore - bestve->score;
ve->score = bestscore;
ve->pred = bestvh;
ve->type = 0;
ve->valid = 1;
ve->lmstate.lm3g.lwid[0] = endwid;
ve->lmstate.lm3g.lwid[1] = ve->lmstate.lm3g.lwid[0];
return vhid;
}
int
ld_record_hyps(live_decoder_t * _decoder, int _end_utt)
{
int32 id;
int32 i = 0;
glist_t hyp_list;
gnode_t *node;
srch_hyp_t *hyp;
char *hyp_strptr = 0;
char *hyp_str = 0;
srch_hyp_t **hyp_segs = 0;
int hyp_seglen = 0;
int hyp_strlen = 0;
int finish_wid = 0;
kb_t *kb = 0;
dict_t *dict;
int rv;
assert(_decoder != NULL);
ld_free_hyps(_decoder);
kb = &_decoder->kb;
dict = kbcore_dict(_decoder->kbcore);
id = _end_utt ?
vithist_utt_end(kb->vithist, _decoder->kbcore) :
vithist_partialutt_end(kb->vithist, _decoder->kbcore);
if (id < 0) {
E_WARN("Failed to retrieve viterbi history.\n");
return LD_ERROR_INTERNAL;
}
/** record the segment length and the overall string length */
hyp_list = vithist_backtrace(kb->vithist, id, dict);
finish_wid = dict_finishwid(dict);
for (node = hyp_list; node != NULL; node = gnode_next(node)) {
hyp = (srch_hyp_t *) gnode_ptr(node);
hyp_seglen++;
if (!dict_filler_word(dict, hyp->id) && hyp->id != finish_wid) {
hyp_strlen +=
strlen(dict_wordstr(dict, dict_basewid(dict, hyp->id))) +
1;
}
}
if (hyp_strlen == 0) {
hyp_strlen = 1;
}
/** allocate array to hold the segments and/or decoded string */
hyp_str = (char *) ckd_calloc(hyp_strlen, sizeof(char));
hyp_segs =
(srch_hyp_t **) ckd_calloc(hyp_seglen + 1, sizeof(srch_hyp_t *));
if (hyp_segs == NULL || hyp_str == NULL) {
E_WARN("Failed to allocate storage for hypothesis.\n");
rv = LD_ERROR_OUT_OF_MEMORY;
goto ld_record_hyps_cleanup;
}
/** iterate thru to fill in the array of segments and/or decoded string */
i = 0;
hyp_strptr = hyp_str;
for (node = hyp_list; node != NULL; node = gnode_next(node), i++) {
hyp = (srch_hyp_t *) gnode_ptr(node);
hyp_segs[i] = hyp;
hyp->word = dict_wordstr(dict, dict_basewid(dict, hyp->id));
if (!dict_filler_word(dict, hyp->id) && hyp->id != finish_wid) {
strcat(hyp_strptr,
dict_wordstr(dict, dict_basewid(dict, hyp->id)));
hyp_strptr += strlen(hyp_strptr);
*hyp_strptr = ' ';
hyp_strptr += 1;
}
}
glist_free(hyp_list);
hyp_str[hyp_strlen - 1] = '\0';
hyp_segs[hyp_seglen] = 0;
_decoder->hyp_frame_num = _decoder->num_frames_decoded;
_decoder->hyp_segs = hyp_segs;
_decoder->hyp_str = hyp_str;
return LD_SUCCESS;
ld_record_hyps_cleanup:
if (hyp_segs != NULL) {
ckd_free(hyp_segs);
}
if (hyp_str != NULL) {
ckd_free(hyp_segs);
}
if (hyp_list != NULL) {
for (node = hyp_list; node != NULL; node = gnode_next(node)) {
if ((hyp = (srch_hyp_t *) gnode_ptr(node)) != NULL) {
ckd_free(hyp);
}
}
}
return rv;
}
void
match_detailed(FILE * fp, glist_t hyp, char *uttid, char *LBL, char *lbl,
int32 * senscale, dict_t * dict)
{
int32 ascr, lscr;
int32 scl;
gnode_t *gn;
srch_hyp_t *h;
if (fp == NULL)
return;
ascr = lscr = scl = 0;
assert(dict);
if (senscale) {
fprintf(fp, "%s:%s> %20s %5s %5s %12s %10s %10s %10s\n", LBL,
uttid, "WORD", "SFrm", "EFrm", "AScr(UnNorm)", "LMScore",
"AScr+LScr", "AScale");
}
else {
fprintf(fp, "%s:%s> %20s %5s %5s %12s %10s %10s %10s\n", LBL,
uttid, "WORD", "SFrm", "EFrm", "AScr(Norm)", "LMScore",
"AScr+LScr", "AScale");
}
for (gn = hyp; gn; gn = gnode_next(gn)) {
h = (srch_hyp_t *) gnode_ptr(gn);
if (h->id != dict_finishwid(dict) && h->id != dict_startwid(dict)) {
scl = 0;
if (h->id < 0 || (h->sf == h->ef))
continue;
scl += compute_scale(h->sf, h->ef, senscale);
if (senscale) {
fprintf(fp, "%s:%s> %20s %5d %5d %12d %10d %10d %10d \n",
lbl, uttid, dict_wordstr(dict, h->id), h->sf,
h->ef, h->ascr + scl, h->lscr,
h->ascr + scl + h->lscr, scl);
}
else {
fprintf(fp, "%s:%s> %20s %5d %5d %12d %10d %10d %10d\n",
lbl, uttid, dict_wordstr(dict, h->id), h->sf,
h->ef, h->ascr, h->lscr, h->ascr + h->lscr, scl);
}
ascr += h->ascr;
if (senscale)
ascr += scl;
lscr += h->lscr;
}
}
fprintf(fp, "%s:%s> %20s %5s %5s %12d %10d\n", LBL, uttid,
"TOTAL", "", "", ascr, lscr);
}
