int32 live_get_partialhyp(int32 endutt)
{
int32 id, nwds;
glist_t hyp;
gnode_t *gn;
hyp_t *h;
dict_t *dict;
dict = kbcore_dict (kb->kbcore);
if (endutt)
id = vithist_utt_end(kb->vithist, kb->kbcore);
else
id = vithist_partialutt_end(kb->vithist, kb->kbcore);
if (id > 0) {
hyp = vithist_backtrace(kb->vithist,id);
for (gn = hyp,nwds=0; gn; gn = gnode_next(gn),nwds++) {
h = (hyp_t *) gnode_ptr (gn);
if (parthyp[nwds].word != NULL) {
ckd_free(parthyp[nwds].word);
parthyp[nwds].word = NULL;
}
parthyp[nwds].word = strdup(dict_wordstr(dict, h->id));
parthyp[nwds].sf = h->sf;
parthyp[nwds].ef = h->ef;
parthyp[nwds].ascr = h->ascr;
parthyp[nwds].lscr = h->lscr;
}
if (parthyp[nwds].word != NULL){
ckd_free(parthyp[nwds].word);
parthyp[nwds].word = NULL;
}
/* Free hyplist */
for (gn = hyp; gn && (gnode_next(gn)); gn = gnode_next(gn)) {
h = (hyp_t *) gnode_ptr (gn);
ckd_free ((void *) h);
}
glist_free (hyp);
} else {
nwds = 0;
if (parthyp[nwds].word != NULL) {
ckd_free(parthyp[nwds].word);
parthyp[nwds].word = NULL;
}
}
return(nwds);
}
int32
vithist_utt_end(vithist_t * vh, ngram_model_t *lm, s3dict_t *dict,
dict2pid_t *dict2pid, fillpen_t *fp)
{
int32 f, i;
int32 sv, nsv, scr, bestscore, bestvh, vhid;
vithist_entry_t *ve, *bestve = 0;
int32 endwid = NGRAM_INVALID_WID;
bestscore = MAX_NEG_INT32;
bestvh = -1;
/* Find last frame with entries in vithist table */
/* by ARCHAN 20050525, it is possible that the last frame will not be reached in decoding */
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_WARN("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?) */
endwid = ngram_wid(lm, S3_FINISH_WORD);
for (i = sv; i < nsv; i++) {
int n_used;
ve = vithist_id2entry(vh, i);
scr = ve->path.score;
scr += ngram_tg_score(lm, endwid, ve->lmstate.lm3g.lwid[0],
ve->lmstate.lm3g.lwid[1], &n_used);
if (bestscore < scr) {
bestscore = scr;
bestvh = i;
bestve = ve;
}
}
assert(bestvh >= 0);
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, lm, dict, dict2pid, fp,
s3dict_silwid(dict), vh->n_frm,
bestve->path.score, bestvh, -1, -1);
vh->n_frm += 1;
vh->frame_start[vh->n_frm] = vh->n_entry;
return vithist_utt_end(vh, lm, dict, dict2pid, fp);
}
/* vithist_dump(vh,-1,kbc,stdout); */
/* Create an </s> entry */
ve = vithist_entry_alloc(vh);
ve->wid = s3dict_finishwid(dict);
ve->sf = (bestve->ef == BAD_S3FRMID) ? 0 : bestve->ef + 1;
ve->ef = vh->n_frm;
ve->ascr = 0;
ve->lscr = bestscore - bestve->path.score;
ve->path.score = bestscore;
ve->path.pred = bestvh;
ve->type = 0;
ve->valid = 1;
ve->lmstate.lm3g.lwid[0] = endwid;
ve->lmstate.lm3g.lwid[1] = ve->lmstate.lm3g.lwid[0];
vhid = vh->n_entry - 1;
/* vithist_dump(vh,-1,kbc,stdout); */
return vhid;
}
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;
}
int32 live_get_partialhyp(int32 endutt)
{
int32 id, nwds;
glist_t hyp;
gnode_t *gn;
hyp_t *h;
dict_t *dict;
dict = kbcore_dict (kb->kbcore);
if (endutt)
id = vithist_utt_end(kb->vithist, kb->kbcore);
else
id = vithist_partialutt_end(kb->vithist, kb->kbcore);
if (id > 0) {
hyp = vithist_backtrace(kb->vithist,id);
for (gn = hyp,nwds=0; gn; gn = gnode_next(gn),nwds++) {
h = (hyp_t *) gnode_ptr (gn);
if (parthyp[nwds].word != NULL) {
ckd_free(parthyp[nwds].word);
parthyp[nwds].word = NULL;
}
/* 20040905 L Galescu <*****@*****.**>
* Report noise? If not, replace with silence word. */
if ((cmd_ln_int32("-reportfill") == 0) &&
dict_filler_word(dict, h->id))
parthyp[nwds].word = strdup(dict_wordstr(dict, dict->silwid));
else
parthyp[nwds].word = strdup(dict_wordstr(dict, h->id));
/* 20040901 L Galescu <*****@*****.**>
* Choice to report alternative pronunciations or not. */
if (cmd_ln_int32("-reportpron") == 0)
dict_word2basestr(parthyp[nwds].word);
parthyp[nwds].sf = h->sf;
parthyp[nwds].ef = h->ef;
parthyp[nwds].ascr = h->ascr;
parthyp[nwds].lscr = h->lscr;
}
if (parthyp[nwds].word != NULL){
ckd_free(parthyp[nwds].word);
parthyp[nwds].word = NULL;
}
parthyplen = nwds;
/* Free hyplist */
for (gn = hyp; gn && (gnode_next(gn)); gn = gnode_next(gn)) {
h = (hyp_t *) gnode_ptr (gn);
ckd_free ((void *) h);
}
glist_free (hyp);
} else if (id == 0) {
nwds = 0;
if (parthyp[nwds].word != NULL) {
ckd_free(parthyp[nwds].word);
parthyp[nwds].word = NULL;
}
parthyplen = nwds;
} else {
/* lgalescu 2004/10/13: nothing changed; we return the same hyp that was there before */
}
return(parthyplen);
}
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;
}
