/*
* 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
matchseg_write(FILE * fp, glist_t hyp, char *uttid, char *hdr,
lm_t * lm, dict_t * dict, int32 num_frm, int32 * ascale,
int32 unnorm)
{
gnode_t *gn;
srch_hyp_t *h;
int32 ascr, lscr, scl, hypscale, global_hypscale;
int32 i;
if (fp == NULL)
return;
ascr = 0;
lscr = 0;
scl = 0;
hypscale = 0;
global_hypscale = 0;
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 */
ascr += h->ascr;
lscr += lm ? lm_rawscore(lm, h->lscr) : h->lscr;
if (unnorm)
global_hypscale += compute_scale(h->sf, h->ef, ascale);
}
}
for (i = 0; i < num_frm; i++)
scl += ascale[i];
fprintf(fp, "%s%s S %d T %d A %d L %d", (hdr ? hdr : ""), uttid,
scl, ascr + lscr + global_hypscale, ascr + global_hypscale,
lscr);
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 */
hypscale = 0;
if (unnorm)
hypscale += compute_scale(h->sf, h->ef, ascale);
fprintf(fp, " %d %d %d %s", h->sf, h->ascr + hypscale,
lm ? lm_rawscore(lm, h->lscr) : h->lscr,
dict_wordstr(dict, h->id));
}
}
fprintf(fp, " %d\n", num_frm);
fflush(fp);
}
/*
* 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);
}
