/*
* Write exact hypothesis. Format:
* <id> T <scr> A <ascr> L <lscr> {<sf> <wascr> <wlscr> <word>}... <ef>
* where:
* scr = ascr + (lscr*lw+N*wip), where N = #words excluding <s>
* ascr = scaled acoustic score for entire utterance
* lscr = LM score (without lw or wip) for entire utterance
* sf = start frame for word
* wascr = scaled acoustic score for word
* wlscr = LM score (without lw or wip) for word
* ef = end frame for utterance.
*/
static void log_hypseg (char *uttid,
FILE *fp, /* Out: output file */
hyp_t *hypptr, /* In: Hypothesis */
int32 nfrm) /* In: #frames in utterance */
{
hyp_t *h;
int32 ascr, lscr, tscr;
ascr = lscr = tscr = 0;
for (h = hypptr; h; h = h->next) {
ascr += h->ascr;
if (dict_basewid(h->wid) != startwid) {
lscr += lm_rawscore (h->lscr, 1.0);
} else {
assert (h->lscr == 0);
}
tscr += h->ascr + h->lscr;
}
fprintf (fp, "%s T %d A %d L %d", uttid, tscr, ascr, lscr);
if (! hypptr) /* HACK!! */
fprintf (fp, " (null)\n");
else {
for (h = hypptr; h; h = h->next) {
lscr = (dict_basewid(h->wid) != startwid) ? lm_rawscore (h->lscr, 1.0) : 0;
fprintf (fp, " %d %d %d %s", h->sf, h->ascr, lscr, dict_wordstr (h->wid));
}
fprintf (fp, " %d\n", nfrm);
}
fflush (fp);
}
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);
}
main (int32 argc, char *argv[])
{
mdef_t *m;
dict_t *d;
char wd[1024];
s3wid_t wid;
int32 p;
if (argc < 3)
E_FATAL("Usage: %s {mdeffile | NULL} dict [fillerdict]\n", argv[0]);
m = (strcmp (argv[1], "NULL") != 0) ? mdef_init (argv[1]) : NULL;
d = dict_init (m, argv[2], ((argc > 3) ? argv[3] : NULL), '_');
for (;;) {
printf ("word> ");
scanf ("%s", wd);
wid = dict_wordid (d, wd);
if (NOT_WID(wid))
E_ERROR("Unknown word\n");
else {
for (wid = dict_basewid(d, wid); IS_WID(wid); wid = d->word[wid].alt) {
printf ("%s\t", dict_wordstr(d, wid));
for (p = 0; p < d->word[wid].pronlen; p++)
printf (" %s", dict_ciphone_str (d, wid, p));
printf ("\n");
}
}
}
}
/* Write hypothesis in old (pre-Nov95) NIST format */
void
log_hypstr(FILE * fp, srch_hyp_t * hypptr, char *uttid, int32 exact,
int32 scr, dict_t * dict)
{
srch_hyp_t *h;
s3wid_t w;
if (fp == NULL)
return;
if (!hypptr) /* HACK!! */
fprintf(fp, "(null)");
for (h = hypptr; h; h = h->next) {
if (h->sf != h->ef) { /* Take care of abnormality caused by
* FSG search or various different
* reasons */
w = h->id;
if (!exact) {
w = dict_basewid(dict, w);
if ((w != dict->startwid) && (w != dict->finishwid)
&& (!dict_filler_word(dict, w)))
fprintf(fp, "%s ", dict_wordstr(dict, w));
}
else
fprintf(fp, "%s ", dict_wordstr(dict, w));
}
}
if (scr != 0)
fprintf(fp, " (%s %d)\n", uttid, scr);
else
fprintf(fp, " (%s)\n", uttid);
fflush(fp);
}
static void
ngram_fwdflat_expand_all(ngram_search_t *ngs)
{
int n_words, i;
/* For all "real words" (not fillers or <s>/</s>) in the dictionary,
*
* 1) Add the ones which are in the LM to the fwdflat wordlist
* 2) And to the expansion list (since we are expanding all)
*/
ngs->n_expand_words = 0;
n_words = ps_search_n_words(ngs);
bitvec_clear_all(ngs->expand_word_flag, ps_search_n_words(ngs));
for (i = 0; i < n_words; ++i) {
if (!ngram_model_set_known_wid(ngs->lmset,
dict_basewid(ps_search_dict(ngs),i)))
continue;
ngs->fwdflat_wordlist[ngs->n_expand_words] = i;
ngs->expand_word_list[ngs->n_expand_words] = i;
bitvec_set(ngs->expand_word_flag, i);
ngs->n_expand_words++;
}
E_INFO("Utterance vocabulary contains %d words\n", ngs->n_expand_words);
ngs->expand_word_list[ngs->n_expand_words] = -1;
ngs->fwdflat_wordlist[ngs->n_expand_words] = -1;
}
static void pronerr_output (char *id, s3wid_t *ref, int32 nref,
wseg_t *wseg, s3cipid_t *ap, int8 *ap_err,
int32 ws, int32 we, int32 ps, int32 pe)
{
int32 j;
s3wid_t rcwid, lcwid;
char str[4096];
/* Word sequence for region in error */
sprintf (str, "%s", dict_wordstr (dict, dict_basewid(dict, ref[ws])));
for (j = ws+1; j <= we; j++) {
strcat (str, " ");
strcat (str, dict_wordstr (dict, dict_basewid(dict, ref[j])));
}
printf ("%-22s\t=>\t", str);
/* Print left context phone */
/*lcwid = ((wseg[ws].s < 0) && (ws > 0) && IS_WID(ref[ws-1])) ? ref[ws-1] : BAD_WID;*/
lcwid = (ws > 0) ? ref[ws-1] : BAD_WID;
if (IS_WID(lcwid)) {
j = dict->word[lcwid].pronlen - 1;
sprintf (str, "(%s)", mdef_ciphone_str (mdef, dict->word[lcwid].ciphone[j]));
} else
strcpy (str, "()");
printf ("%-5s", str);
/* Phone sequence for region in error */
for (j = ps; j <= pe; j++) {
strcpy (str, mdef_ciphone_str (mdef, ap[j]));
if (ap_err[j])
ucase (str);
else
lcase (str);
printf (" %s", str);
}
/* Right context if ending in error */
/* rcwid = ((wseg[we].e < 0) && IS_WID(ref[we+1])) ? ref[we+1] : BAD_WID; */
rcwid = ref[we+1];
if (IS_WID(rcwid))
printf ("\t(%s)", mdef_ciphone_str (mdef, dict->word[rcwid].ciphone[0]));
else
printf ("\t()");
printf (" ( %s )\n", id);
}
/*
* Write exact hypothesis. Format
* <id> S <scl> T <scr> A <ascr> L <lscr> {<sf> <wascr> <wlscr> <word>}... <ef>
* where:
* scl = acoustic score scaling for entire utterance
* scr = ascr + (lscr*lw+N*wip), where N = #words excluding <s>
* ascr = scaled acoustic score for entire utterance
* lscr = LM score (without lw or wip) for entire utterance
* sf = start frame for word
* wascr = scaled acoustic score for word
* wlscr = LM score (without lw or wip) for word
* ef = end frame for utterance.
*/
void
log_hypseg(char *uttid, FILE * fp, /* Out: output file */
srch_hyp_t * hypptr, /* In: Hypothesis */
int32 nfrm, /* In: #frames in utterance */
int32 scl, /* In: Acoustic scaling for entire utt */
float64 lwf, /* In: LM score scale-factor (in dagsearch) */
dict_t * dict, /* In: dictionary */
lm_t * lm, int32 unnorm
/**< Whether unscaled the score back */
)
{
srch_hyp_t *h;
int32 ascr, lscr, tscr;
if (fp == NULL)
return;
ascr = lscr = tscr = 0;
for (h = hypptr; h; h = h->next) {
ascr += h->ascr;
if (dict_basewid(dict, h->id) != dict->startwid) {
lscr += lm_rawscore(lm, h->lscr);
}
else {
assert(h->lscr == 0);
}
tscr += h->ascr + h->lscr;
}
fprintf(fp, "%s S %d T %d A %d L %d", uttid, scl, tscr, ascr, lscr);
if (!hypptr) /* HACK!! */
fprintf(fp, " (null)\n");
else {
for (h = hypptr; h; h = h->next) {
lscr =
(dict_basewid(dict, h->id) !=
dict->startwid) ? lm_rawscore(lm, h->lscr) : 0;
fprintf(fp, " %d %d %d %s", h->sf, h->ascr, lscr,
dict_wordstr(dict, h->id));
}
fprintf(fp, " %d\n", nfrm);
}
fflush(fp);
}
/*
* Scan the dictionary for compound words. This function should be called just after
* loading the dictionary. For the moment, compound words in a compound word are
* assumed to be separated by the given sep character, (underscore in the CMU dict).
* Return value: #compound words found in dictionary.
*/
static int32 dict_build_comp (dict_t *d,
char sep) /* Separator character */
{
char wd[4096];
int32 w, cwid;
dictword_t *wordp;
int32 nc; /* # compound words in dictionary */
int32 i, j, l, n;
nc = 0;
for (w = 0; w < d->n_word; w++) {
wordp = d->word + dict_basewid(d, w);
strcpy (wd, wordp->word);
l = strlen(wd);
if ((wd[0] == sep) || (wd[l-1] == sep))
E_FATAL("Bad compound word %s: leading or trailing separator\n", wordp->word);
/* Count no. of components in this word */
n = 1;
for (i = 1; i < l-1; i++) /* 0 and l-1 already checked above */
if (wd[i] == sep)
n++;
if (n == 1)
continue; /* Not a compound word */
nc++;
if ((w == d->startwid) || (w == d->finishwid) || dict_filler_word (d, w))
E_FATAL("Compound special/filler word (%s) not allowed\n", wordp->word);
/* Allocate and fill in component word info */
wordp->n_comp = n;
wordp->comp = (s3wid_t *) ckd_calloc (n, sizeof(s3wid_t));
/* Parse word string into components */
n = 0;
for (i = 0; i < l; i++) {
for (j = i; (i < l) && (wd[i] != sep); i++);
if (j == i)
E_FATAL("Bad compound word %s: successive separators\n", wordp->word);
wd[i] = '\0';
cwid = dict_wordid (d, wd+j);
if (NOT_WID(cwid))
E_FATAL("Component word %s of %s not in dictionary\n", wd+j, wordp->word);
wordp->comp[n] = cwid;
n++;
}
}
if (nc > 0)
d->comp_head = dict_comp_head (d);
return nc;
}
int32 dict_filler_word (dict_t *d, s3wid_t w)
{
assert (d);
assert ((w >= 0) && (w < d->n_word));
w = dict_basewid(d, w);
if ((w == d->startwid) || (w == d->finishwid))
return 0;
if ((w >= d->filler_start) && (w <= d->filler_end))
return 1;
return 0;
}
main (int32 argc, char *argv[])
{
dict_t **d;
int32 i, k, p, wid;
char line[16384], *wp[1024];
if (argc < 2) {
E_INFO("Usage: %s dictfile [dictfile ...] < vocabfile\n", argv[0]);
exit(0);
}
d = (dict_t **) ckd_calloc (argc-1, sizeof(dict_t *));
for (i = 1; i < argc; i++)
d[i-1] = dict_init (NULL, argv[i], NULL, 0);
while (fgets (line, sizeof(line), stdin) != NULL) {
if ((k = str2words (line, wp, 1024)) < 0)
E_FATAL("Line too long: %s\n", line);
if (k > 2)
E_FATAL("Vocab entry contains too many words\n");
if (k == 0)
continue;
if (k == 1)
wp[1] = wp[0];
/* Look up word in each dictionary until found */
k = 0;
for (i = 0; (i < argc-1) && (k == 0); i++) {
wid = dict_wordid (d[i], wp[1]);
if (NOT_WID(wid))
continue;
for (wid = dict_basewid(d[i], wid);
IS_WID(wid);
wid = dict_nextalt(d[i], wid)) {
k++;
if (k == 1)
printf ("%s\t", wp[0]);
else
printf ("%s(%d)\t", wp[0], k);
for (p = 0; p < dict_pronlen(d[i], wid); p++)
printf (" %s", dict_ciphone_str (d[i], wid, p));
printf ("\n");
}
}
if (k == 0)
E_ERROR("No pronunciation for: '%s'\n", wp[0]);
}
}
int32 align_init ( void )
{
int32 k;
s3wid_t w;
float64 *f64arg;
mdef = mdef_getmdef ();
tmat = tmat_gettmat ();
dict = dict_getdict ();
assert (mdef && tmat && dict);
startwid = dict_wordid (START_WORD);
finishwid = dict_wordid (FINISH_WORD);
silwid = dict_wordid (SILENCE_WORD);
if ((NOT_WID(startwid)) || (NOT_WID(finishwid)))
E_FATAL("%s or %s not in dictionary\n", START_WORD, FINISH_WORD);
if (NOT_WID(silwid))
E_ERROR("%s not in dictionary; no optional silence inserted between words\n",
SILENCE_WORD);
/* Create list of optional filler words to be inserted between transcript words */
fillwid = (s3wid_t *) ckd_calloc ((dict->filler_end - dict->filler_start + 3),
sizeof(s3wid_t));
k = 0;
if (IS_WID(silwid))
fillwid[k++] = silwid;
for (w = dict->filler_start; w <= dict->filler_end; w++) {
if ((dict_basewid (w) == w) &&
(w != silwid) && (w != startwid) && (w != finishwid))
fillwid[k++] = w;
}
fillwid[k] = BAD_WID;
f64arg = (float64 *) cmd_ln_access ("-beam");
beam = logs3 (*f64arg);
E_INFO ("logs3(beam)= %d\n", beam);
score_scale = (int32 *) ckd_calloc (S3_MAX_FRAMES, sizeof(int32));
hist_head = NULL;
align_stseg = NULL;
align_phseg = NULL;
align_wdseg = NULL;
ctr_nstate = counter_new ("NS");
return 0;
}
int
dict_altid(dict_t *d, int32 wid)
{
char const *basestr, *wordstr, *c;
if (wid < 0 || wid >= d->n_word)
return 0;
if (wid == dict_basewid(d, wid))
return 1;
/* FIXME: the order in which they appear in nextalt is unrelated
* to the actual alterate indices in the words... */
wordstr = dict_wordstr(d, wid);
basestr = dict_basestr(d, wid);
assert(strlen(wordstr) > strlen(basestr));
c = wordstr + strlen(basestr);
assert(*c == '(');
++c;
return atoi(c);
}
/* Write hypothesis in old (pre-Nov95) NIST format */
static void log_hypstr (FILE *fp, hyp_t *hypptr, char *uttid, int32 scr)
{
hyp_t *h;
s3wid_t w;
if (! hypptr) /* HACK!! */
fprintf (fp, "(null)");
for (h = hypptr; h; h = h->next) {
w = dict_basewid (h->wid);
if ((w != startwid) && (w != finishwid) && (! dict_filler_word (w)))
fprintf (fp, "%s ", dict_wordstr(w));
}
if (scr != 0)
fprintf (fp, " (%s %d)\n", uttid, scr);
else
fprintf (fp, " (%s)\n", uttid);
fflush (fp);
}
s3wid_t dict_wids2compwid (dict_t *d, s3wid_t *wid, int32 len)
{
s3wid_t w;
int32 i;
if (! d->comp_head)
return BAD_WID;
assert (len > 1);
for (w = d->comp_head[wid[0]]; IS_WID(w); w = d->comp_head[w]) {
/* w is a compound word beginning with wid[0]; check if rest matches */
assert (d->word[w].n_comp > 1);
assert (d->word[w].comp[0] == wid[0]);
if (d->word[w].n_comp == len) {
for (i = 0; (i < len) && (d->word[w].comp[i] == wid[i]); i++);
if (i == len)
return (dict_basewid(d, w));
}
}
return BAD_WID;
}
word_fsg_t *
word_fsg_load(s2_fsg_t * fsg,
int use_altpron, int use_filler,
kbcore_t *kbc)
{
float32 silprob = kbc->fillpen->silprob;
float32 fillprob = kbc->fillpen->fillerprob;
float32 lw = kbc->fillpen->lw;
word_fsg_t *word_fsg;
s2_fsg_trans_t *trans;
int32 n_trans, n_null_trans, n_alt_trans, n_filler_trans, n_unk;
int32 wid;
int32 logp;
glist_t nulls;
int32 i, j;
assert(fsg);
/* Some error checking */
if (lw <= 0.0)
E_WARN("Unusual language-weight value: %.3e\n", lw);
if (use_filler && ((silprob < 0.0) || (fillprob < 0.0))) {
E_ERROR("silprob/fillprob must be >= 0\n");
return NULL;
}
if ((fsg->n_state <= 0)
|| ((fsg->start_state < 0) || (fsg->start_state >= fsg->n_state))
|| ((fsg->final_state < 0) || (fsg->final_state >= fsg->n_state))) {
E_ERROR("Bad #states/start_state/final_state values: %d/%d/%d\n",
fsg->n_state, fsg->start_state, fsg->final_state);
return NULL;
}
for (trans = fsg->trans_list; trans; trans = trans->next) {
if ((trans->from_state < 0) || (trans->from_state >= fsg->n_state)
|| (trans->to_state < 0) || (trans->to_state >= fsg->n_state)
|| (trans->prob <= 0) || (trans->prob > 1.0)) {
E_ERROR("Bad transition: P(%d -> %d) = %e\n",
trans->from_state, trans->to_state, trans->prob);
return NULL;
}
}
word_fsg = (word_fsg_t *) ckd_calloc(1, sizeof(word_fsg_t));
word_fsg->name = ckd_salloc(fsg->name ? fsg->name : "");
word_fsg->n_state = fsg->n_state;
word_fsg->start_state = fsg->start_state;
word_fsg->final_state = fsg->final_state;
word_fsg->use_altpron = use_altpron;
word_fsg->use_filler = use_filler;
word_fsg->lw = lw;
word_fsg->lc = NULL;
word_fsg->rc = NULL;
word_fsg->dict = kbc->dict;
word_fsg->mdef = kbc->mdef;
word_fsg->tmat = kbc->tmat;
word_fsg->n_ciphone = mdef_n_ciphone(kbc->mdef);
/* Allocate non-epsilon transition matrix array */
word_fsg->trans = (glist_t **) ckd_calloc_2d(word_fsg->n_state,
word_fsg->n_state,
sizeof(glist_t));
/* Allocate epsilon transition matrix array */
word_fsg->null_trans = (word_fsglink_t ***)
ckd_calloc_2d(word_fsg->n_state, word_fsg->n_state,
sizeof(word_fsglink_t *));
/* Process transitions */
n_null_trans = 0;
n_alt_trans = 0;
n_filler_trans = 0;
n_unk = 0;
nulls = NULL;
for (trans = fsg->trans_list, n_trans = 0;
trans; trans = trans->next, n_trans++) {
/* Convert prob to logs2prob and apply language weight */
logp = (int32) (logs3(kbcore_logmath(kbc), trans->prob) * lw);
/* Check if word is in dictionary */
if (trans->word) {
wid = dict_wordid(kbc->dict, trans->word);
if (wid < 0) {
E_ERROR("Unknown word '%s'; ignored\n", trans->word);
n_unk++;
}
else if (use_altpron) {
wid = dict_basewid(kbc->dict, wid);
assert(wid >= 0);
}
}
else
wid = -1; /* Null transition */
/* Add transition to word_fsg structure */
i = trans->from_state;
j = trans->to_state;
if (wid < 0) {
if (word_fsg_null_trans_add(word_fsg, i, j, logp) == 1) {
n_null_trans++;
nulls =
glist_add_ptr(nulls,
(void *) word_fsg->null_trans[i][j]);
}
}
else {
word_fsg_trans_add(word_fsg, i, j, logp, wid);
/* Add transitions for alternative pronunciations, if any */
if (use_altpron) {
for (wid = dict_nextalt(kbc->dict, wid);
wid >= 0; wid = dict_nextalt(kbc->dict, wid)) {
word_fsg_trans_add(word_fsg, i, j, logp, wid);
n_alt_trans++;
n_trans++;
}
}
}
}
/* Add silence and noise filler word transitions if specified */
if (use_filler) {
n_filler_trans = word_fsg_add_filler(word_fsg, silprob, fillprob, kbcore_logmath(kbc));
n_trans += n_filler_trans;
}
E_INFO
("FSG: %d states, %d transitions (%d null, %d alt, %d filler, %d unknown)\n",
word_fsg->n_state, n_trans, n_null_trans, n_alt_trans,
n_filler_trans, n_unk);
#if __FSG_DBG__
E_INFO("FSG before NULL closure:\n");
word_fsg_write(word_fsg, stdout);
#endif
/* Null transitions closure */
nulls = word_fsg_null_trans_closure(word_fsg, nulls);
glist_free(nulls);
#if __FSG_DBG__
E_INFO("FSG after NULL closure:\n");
word_fsg_write(word_fsg, stdout);
#endif
/* Compute left and right context CIphone lists for each state */
word_fsg_lc_rc(word_fsg);
#if __FSG_DBG__
E_INFO("FSG after lc/rc:\n");
word_fsg_write(word_fsg, stdout);
#endif
return word_fsg;
}
/*
* 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 */
}
}
fillpen_t *fillpen_init (dict_t *dict, char *file, float64 silprob, float64 fillprob,
float64 lw, float64 wip)
{
s3wid_t w, bw;
float64 prob;
FILE *fp;
char line[1024], wd[1024];
int32 k;
fillpen_t *_fillpen;
_fillpen = (fillpen_t *) ckd_calloc (1, sizeof(fillpen_t));
_fillpen->dict = dict;
_fillpen->lw = lw;
_fillpen->wip = wip;
if (dict->filler_end >= dict->filler_start)
_fillpen->prob = (int32 *) ckd_calloc (dict->filler_end - dict->filler_start + 1,
sizeof(int32));
else
_fillpen->prob = NULL;
/* Initialize all words with filler penalty (HACK!! backward compatibility) */
prob = fillprob;
for (w = dict->filler_start; w <= dict->filler_end; w++)
_fillpen->prob[w - dict->filler_start] = (int32) ((logs3(prob) + logs3(wip)) * lw);
/* Overwrite silence penalty (HACK!! backward compatibility) */
w = dict_wordid (dict, S3_SILENCE_WORD);
if (NOT_S3WID(w) || (w < dict->filler_start) || (w > dict->filler_end))
E_FATAL("%s not a filler word in the given dictionary\n", S3_SILENCE_WORD);
prob = silprob;
_fillpen->prob[w - dict->filler_start] = (int32) ((logs3(prob) + logs3(wip)) * lw);
/* Overwrite with filler prob input file, if specified */
if (! file)
return _fillpen;
E_INFO("Reading filler penalty file: %s\n", file);
if ((fp = fopen (file, "r")) == NULL)
E_FATAL("fopen(%s,r) failed\n", file);
while (fgets (line, sizeof(line), fp) != NULL) {
if (line[0] == '#') /* Skip comment lines */
continue;
k = sscanf (line, "%s %lf", wd, &prob);
if ((k != 0) && (k != 2))
E_FATAL("Bad input line: %s\n", line);
w = dict_wordid(dict, wd);
if (NOT_S3WID(w) || (w < dict->filler_start) || (w > dict->filler_end))
E_FATAL("%s not a filler word in the given dictionary\n", S3_SILENCE_WORD);
_fillpen->prob[w - dict->filler_start] = (int32) ((logs3(prob) + logs3(wip)) * lw);
}
fclose (fp);
/* Replicate fillpen values for alternative pronunciations */
for (w = dict->filler_start; w <= dict->filler_end; w++) {
bw = dict_basewid (dict, w);
if (bw != w)
_fillpen->prob[w-dict->filler_start] = _fillpen->prob[bw-dict->filler_start];
}
return _fillpen;
}
static void
fwdflat_word_transition(ngram_search_t *ngs, int frame_idx)
{
int32 cf, nf, b, thresh, pip, i, w, newscore;
int32 best_silrc_score = 0, best_silrc_bp = 0; /* FIXME: good defaults? */
bptbl_t *bp;
int32 *rcss;
root_chan_t *rhmm;
int32 *awl;
float32 lwf;
dict_t *dict = ps_search_dict(ngs);
dict2pid_t *d2p = ps_search_dict2pid(ngs);
cf = frame_idx;
nf = cf + 1;
thresh = ngs->best_score + ngs->fwdflatbeam;
pip = ngs->pip;
best_silrc_score = WORST_SCORE;
lwf = ngs->fwdflat_fwdtree_lw_ratio;
/* Search for all words starting within a window of this frame.
* These are the successors for words exiting now. */
get_expand_wordlist(ngs, cf, ngs->max_sf_win);
/* Scan words exited in current frame */
for (b = ngs->bp_table_idx[cf]; b < ngs->bpidx; b++) {
xwdssid_t *rssid;
int32 silscore;
bp = ngs->bp_table + b;
ngs->word_lat_idx[bp->wid] = NO_BP;
if (bp->wid == ps_search_finish_wid(ngs))
continue;
/* DICT2PID location */
/* Get the mapping from right context phone ID to index in the
* right context table and the bscore_stack. */
rcss = ngs->bscore_stack + bp->s_idx;
if (bp->last2_phone == -1)
rssid = NULL;
else
rssid = dict2pid_rssid(d2p, bp->last_phone, bp->last2_phone);
/* Transition to all successor words. */
for (i = 0; ngs->expand_word_list[i] >= 0; i++) {
int32 n_used;
w = ngs->expand_word_list[i];
/* Get the exit score we recorded in save_bwd_ptr(), or
* something approximating it. */
if (rssid)
newscore = rcss[rssid->cimap[dict_first_phone(dict, w)]];
else
newscore = bp->score;
if (newscore == WORST_SCORE)
continue;
/* FIXME: Floating point... */
newscore += lwf
* (ngram_tg_score(ngs->lmset,
dict_basewid(dict, w),
bp->real_wid,
bp->prev_real_wid,
&n_used) >> SENSCR_SHIFT);
newscore += pip;
/* Enter the next word */
if (newscore BETTER_THAN thresh) {
rhmm = (root_chan_t *) ngs->word_chan[w];
if ((hmm_frame(&rhmm->hmm) < cf)
|| (newscore BETTER_THAN hmm_in_score(&rhmm->hmm))) {
hmm_enter(&rhmm->hmm, newscore, b, nf);
/* DICT2PID: This is where mpx ssids get introduced. */
/* Look up the ssid to use when entering this mpx triphone. */
hmm_mpx_ssid(&rhmm->hmm, 0) =
dict2pid_ldiph_lc(d2p, rhmm->ciphone, rhmm->ci2phone,
dict_last_phone(dict, bp->wid));
assert(IS_S3SSID(hmm_mpx_ssid(&rhmm->hmm, 0)));
E_DEBUG(6,("ssid %d(%d,%d) = %d\n",
rhmm->ciphone, dict_last_phone(dict, bp->wid), rhmm->ci2phone,
hmm_mpx_ssid(&rhmm->hmm, 0)));
bitvec_set(ngs->word_active, w);
}
}
}
/* Get the best exit into silence. */
if (rssid)
silscore = rcss[rssid->cimap[ps_search_acmod(ngs)->mdef->sil]];
else
silscore = bp->score;
if (silscore BETTER_THAN best_silrc_score) {
best_silrc_score = silscore;
best_silrc_bp = b;
}
}
/* Transition to <sil> */
newscore = best_silrc_score + ngs->silpen + pip;
if ((newscore BETTER_THAN thresh) && (newscore BETTER_THAN WORST_SCORE)) {
w = ps_search_silence_wid(ngs);
rhmm = (root_chan_t *) ngs->word_chan[w];
if ((hmm_frame(&rhmm->hmm) < cf)
|| (newscore BETTER_THAN hmm_in_score(&rhmm->hmm))) {
hmm_enter(&rhmm->hmm, newscore,
best_silrc_bp, nf);
bitvec_set(ngs->word_active, w);
}
}
/* Transition to noise words */
newscore = best_silrc_score + ngs->fillpen + pip;
if ((newscore BETTER_THAN thresh) && (newscore BETTER_THAN WORST_SCORE)) {
for (w = ps_search_silence_wid(ngs) + 1; w < ps_search_n_words(ngs); w++) {
rhmm = (root_chan_t *) ngs->word_chan[w];
/* Noise words that aren't a single phone will have NULL here. */
if (rhmm == NULL)
continue;
if ((hmm_frame(&rhmm->hmm) < cf)
|| (newscore BETTER_THAN hmm_in_score(&rhmm->hmm))) {
hmm_enter(&rhmm->hmm, newscore,
best_silrc_bp, nf);
bitvec_set(ngs->word_active, w);
}
}
}
/* Reset initial channels of words that have become inactive even after word trans. */
i = ngs->n_active_word[cf & 0x1];
awl = ngs->active_word_list[cf & 0x1];
for (w = *(awl++); i > 0; --i, w = *(awl++)) {
rhmm = (root_chan_t *) ngs->word_chan[w];
if (hmm_frame(&rhmm->hmm) == cf) {
hmm_clear_scores(&rhmm->hmm);
}
}
}
/*
* Build a sentence HMM for the given transcription (wordstr). A two-level DAG is
* built: phone-level and state-level.
* - <s> and </s> always added at the beginning and end of sentence to form an
* augmented transcription.
* - Optional <sil> and noise words added between words in the augmented
* transcription.
* wordstr must contain only the transcript; no extraneous stuff such as utterance-id.
* Phone-level HMM structure has replicated nodes to allow for different left and right
* context CI phones; hence, each pnode corresponds to a unique triphone in the sentence
* HMM.
* Return 0 if successful, <0 if any error (eg, OOV word encountered).
*/
int32 align_build_sent_hmm (char *wordstr)
{
s3wid_t w, nextw;
int32 k, oov;
pnode_t *word_end, *node;
char *wd, delim, *wdcopy;
/* HACK HACKA HACK BHIKSHA */
int32 firsttime = 1;
/* END HACK HACKA HACK */
/* Initialize dummy head and tail entries of sent hmm */
phead.wid = BAD_WID;
phead.ci = BAD_CIPID;
phead.lc = BAD_CIPID; /* No predecessor */
phead.rc = BAD_CIPID; /* Any phone can follow head */
phead.pid = BAD_PID;
phead.succlist = NULL;
phead.predlist = NULL;
phead.next = NULL; /* Will ultimately be the head of list of all pnodes */
phead.id = -1; /* Hardwired */
phead.startstate = NULL;
ptail.wid = BAD_WID;
ptail.ci = BAD_CIPID;
ptail.lc = BAD_CIPID; /* Any phone can precede tail */
ptail.rc = BAD_CIPID; /* No successor */
ptail.pid = BAD_PID;
ptail.succlist = NULL;
ptail.predlist = NULL;
ptail.next = NULL;
ptail.id = -2; /* Hardwired */
ptail.startstate = NULL;
n_pnode = 0;
pnode_list = NULL;
oov = 0;
/* State-level DAG initialization should be here in case the build is aborted */
shead.pnode = &phead;
shead.succlist = NULL;
shead.predlist = NULL;
shead.sen = BAD_SENID;
shead.state = mdef->n_emit_state;
shead.hist = NULL;
stail.pnode = &ptail;
stail.succlist = NULL;
stail.predlist = NULL;
stail.sen = BAD_SENID;
stail.state = 0;
stail.hist = NULL;
/* Obtain the first transcript word */
k = nextword (wordstr, " \t\n", &wd, &delim);
if (k < 0)
nextw = finishwid;
else {
wordstr = wd + k;
wdcopy = ckd_salloc (wd);
*wordstr = delim;
nextw = dict_wordid (wdcopy);
if (IS_WID(nextw))
nextw = dict_basewid (nextw);
}
/* Create node(s) for <s> before any transcript word */
/* HACK HACKA HACK BHIKSHA
word_end = append_transcript_word (startwid, &phead, nextw, 0, 1);
END HACK HACKA HACK BHIKSHA */
/* Append each word in transcription to partial sent HMM created so far */
while (k >= 0) {
w = nextw;
if (NOT_WID(w)) {
E_ERROR("%s not in dictionary\n", wdcopy);
oov = 1;
/* Hack!! Temporarily set w to some dummy just to run through sentence */
w = finishwid;
}
ckd_free (wdcopy);
k = nextword (wordstr, " \t\n", &wd, &delim);
if (k < 0)
nextw = finishwid;
else {
wordstr = wd + k;
wdcopy = ckd_salloc (wd);
*wordstr = delim;
nextw = dict_wordid (wdcopy);
if (IS_WID(nextw))
nextw = dict_basewid (nextw);
}
/* HACK HACKA HACK BHIKSHA */
if (firsttime){
word_end = append_transcript_word (w, &phead, nextw, 0, 1);
firsttime = 0;
}
else
if (nextw == finishwid)
word_end = append_transcript_word (w, word_end, BAD_WID, 1, 0);
else
word_end = append_transcript_word (w, word_end, nextw, 1, 1);
/* END HACK HACKA HACK BHIKSHA */
}
if (oov)
return -1;
/* Append phone HMMs for </s> at the end; link to tail node */
/* HACK HACKA HACK BHIKSHA
word_end = append_transcript_word (finishwid, word_end, BAD_WID, 1, 0);
END HACK HACKA HACK BHIKSHA */
for (node = word_end; node; node = node->next)
link_pnodes (node, &ptail);
/* Build state-level DAG from the phone-level one */
build_state_dag ();
/* Dag must begin and end at shead and stail, respectively */
assert (shead.succlist);
assert (stail.predlist);
assert (! shead.predlist);
assert (! stail.succlist);
#if _DEBUG_ALIGN_
dump_sent_hmm (); /* For debugging */
#endif
k = n_pnode * mdef->n_emit_state;
if (k > active_list_size) { /* Need to grow active list arrays */
if (active_list_size > 0) {
ckd_free (cur_active);
ckd_free (next_active);
}
for (; active_list_size <= k; active_list_size += ACTIVE_LIST_SIZE_INCR);
cur_active = (snode_t **) ckd_calloc (active_list_size, sizeof(snode_t *));
next_active = (snode_t **) ckd_calloc (active_list_size, sizeof(snode_t *));
}
return 0;
}
void lmcontext_load (corpus_t *corp, char *uttid, s3wid_t *pred, s3wid_t *succ)
{
char *str, wd[4096], *strp;
s3wid_t w[3];
int32 i, n;
dict_t *dict;
s3lmwid_t lwid;
if ((str = corpus_lookup (corp, uttid)) == NULL)
E_FATAL("Couldn't find LM context for %s\n", uttid);
dict = dict_getdict ();
strp = str;
for (i = 0; i < 4; i++) {
if (sscanf (strp, "%s%n", wd, &n) != 1) {
if (i < 3)
E_FATAL("Bad LM context spec for %s: %s\n", uttid, str);
else
break;
}
strp += n;
if (strcmp (wd, "-") == 0)
w[i] = BAD_WID;
else {
w[i] = dict_wordid (wd);
if (NOT_WID(w[i]))
E_FATAL("LM context word (%s) for %s not in dictionary\n", wd, uttid);
w[i] = dict_basewid(w[i]);
switch (i) {
case 0:
if ((n = dict->word[w[0]].n_comp) > 0)
w[0] = dict->word[w[0]].comp[n-1].wid;
break;
case 1:
if ((n = dict->word[w[1]].n_comp) > 0) {
w[0] = dict->word[w[1]].comp[n-2].wid;
w[1] = dict->word[w[1]].comp[n-1].wid;
}
break;
case 2:
if (w[2] != dict_wordid(FINISH_WORD))
E_FATAL("Illegal successor LM context for %s: %s\n", uttid, str);
break;
default:
assert (0); /* Should never get here */
break;
}
}
}
if (IS_WID(w[0]) && NOT_WID(w[1]))
E_FATAL("Bad LM context spec for %s: %s\n", uttid, str);
for (i = 0; i < 3; i++) {
if (IS_WID(w[i])) {
lwid = lm_lmwid (w[i]);
if (NOT_LMWID(lwid))
E_FATAL("LM context word (%s) for %s not in LM\n", wd, uttid);
}
}
pred[0] = w[0];
pred[1] = w[1];
*succ = w[2];
}
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;
}
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;
}
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;
}
/**
* Find all active words in backpointer table and sort by frame.
*/
static void
build_fwdflat_wordlist(ngram_search_t *ngs)
{
int32 i, f, sf, ef, wid, nwd;
bptbl_t *bp;
ps_latnode_t *node, *prevnode, *nextnode;
/* No tree-search, use statically allocated wordlist. */
if (!ngs->fwdtree)
return;
memset(ngs->frm_wordlist, 0, ngs->n_frame_alloc * sizeof(*ngs->frm_wordlist));
/* Scan the backpointer table for all active words and record
* their exit frames. */
for (i = 0, bp = ngs->bp_table; i < ngs->bpidx; i++, bp++) {
sf = (bp->bp < 0) ? 0 : ngs->bp_table[bp->bp].frame + 1;
ef = bp->frame;
wid = bp->wid;
/* Anything that can be transitioned to in the LM can go in
* the word list. */
if (!ngram_model_set_known_wid(ngs->lmset,
dict_basewid(ps_search_dict(ngs), wid)))
continue;
/* Look for it in the wordlist. */
for (node = ngs->frm_wordlist[sf]; node && (node->wid != wid);
node = node->next);
/* Update last end frame. */
if (node)
node->lef = ef;
else {
/* New node; link to head of list */
node = listelem_malloc(ngs->latnode_alloc);
node->wid = wid;
node->fef = node->lef = ef;
node->next = ngs->frm_wordlist[sf];
ngs->frm_wordlist[sf] = node;
}
}
/* Eliminate "unlikely" words, for which there are too few end points */
for (f = 0; f < ngs->n_frame; f++) {
prevnode = NULL;
for (node = ngs->frm_wordlist[f]; node; node = nextnode) {
nextnode = node->next;
/* Word has too few endpoints */
if ((node->lef - node->fef < ngs->min_ef_width) ||
/* Word is </s> and doesn't actually end in last frame */
((node->wid == ps_search_finish_wid(ngs)) && (node->lef < ngs->n_frame - 1))) {
if (!prevnode)
ngs->frm_wordlist[f] = nextnode;
else
prevnode->next = nextnode;
listelem_free(ngs->latnode_alloc, node);
}
else
prevnode = node;
}
}
/* Form overall wordlist for 2nd pass */
nwd = 0;
bitvec_clear_all(ngs->word_active, ps_search_n_words(ngs));
for (f = 0; f < ngs->n_frame; f++) {
for (node = ngs->frm_wordlist[f]; node; node = node->next) {
if (!bitvec_is_set(ngs->word_active, node->wid)) {
bitvec_set(ngs->word_active, node->wid);
ngs->fwdflat_wordlist[nwd++] = node->wid;
}
}
}
ngs->fwdflat_wordlist[nwd] = -1;
E_INFO("Utterance vocabulary contains %d words\n", nwd);
}
