dict2pid_t *
dict2pid_build(bin_mdef_t * mdef, dict_t * dict)
{
dict2pid_t *dict2pid;
s3ssid_t ***rdiph_rc;
bitvec_t *ldiph, *rdiph, *single;
int32 pronlen;
int32 b, l, r, w, p;
E_INFO("Building PID tables for dictionary\n");
assert(mdef);
assert(dict);
dict2pid = (dict2pid_t *) ckd_calloc(1, sizeof(dict2pid_t));
dict2pid->refcount = 1;
dict2pid->mdef = bin_mdef_retain(mdef);
dict2pid->dict = dict_retain(dict);
E_INFO("Allocating %d^3 * %d bytes (%d KiB) for word-initial triphones\n",
mdef->n_ciphone, sizeof(s3ssid_t),
mdef->n_ciphone * mdef->n_ciphone * mdef->n_ciphone * sizeof(s3ssid_t) / 1024);
dict2pid->ldiph_lc =
(s3ssid_t ***) ckd_calloc_3d(mdef->n_ciphone, mdef->n_ciphone,
mdef->n_ciphone, sizeof(s3ssid_t));
/* Only used internally to generate rssid */
rdiph_rc =
(s3ssid_t ***) ckd_calloc_3d(mdef->n_ciphone, mdef->n_ciphone,
mdef->n_ciphone, sizeof(s3ssid_t));
dict2pid->lrdiph_rc = (s3ssid_t ***) ckd_calloc_3d(mdef->n_ciphone,
mdef->n_ciphone,
mdef->n_ciphone,
sizeof
(s3ssid_t));
/* Actually could use memset for this, if BAD_S3SSID is guaranteed
* to be 65535... */
for (b = 0; b < mdef->n_ciphone; ++b) {
for (r = 0; r < mdef->n_ciphone; ++r) {
for (l = 0; l < mdef->n_ciphone; ++l) {
dict2pid->ldiph_lc[b][r][l] = BAD_S3SSID;
dict2pid->lrdiph_rc[b][l][r] = BAD_S3SSID;
rdiph_rc[b][l][r] = BAD_S3SSID;
}
}
}
/* Track which diphones / ciphones have been seen. */
ldiph = bitvec_alloc(mdef->n_ciphone * mdef->n_ciphone);
rdiph = bitvec_alloc(mdef->n_ciphone * mdef->n_ciphone);
single = bitvec_alloc(mdef->n_ciphone);
for (w = 0; w < dict_size(dict2pid->dict); w++) {
pronlen = dict_pronlen(dict, w);
if (pronlen >= 2) {
b = dict_first_phone(dict, w);
r = dict_second_phone(dict, w);
/* Populate ldiph_lc */
if (bitvec_is_clear(ldiph, b * mdef->n_ciphone + r)) {
/* Mark this diphone as done */
bitvec_set(ldiph, b * mdef->n_ciphone + r);
/* Record all possible ssids for b(?,r) */
for (l = 0; l < bin_mdef_n_ciphone(mdef); l++) {
p = bin_mdef_phone_id_nearest(mdef, (s3cipid_t) b,
(s3cipid_t) l, (s3cipid_t) r,
WORD_POSN_BEGIN);
dict2pid->ldiph_lc[b][r][l] = bin_mdef_pid2ssid(mdef, p);
}
}
/* Populate rdiph_rc */
l = dict_second_last_phone(dict, w);
b = dict_last_phone(dict, w);
if (bitvec_is_clear(rdiph, b * mdef->n_ciphone + l)) {
/* Mark this diphone as done */
bitvec_set(rdiph, b * mdef->n_ciphone + l);
for (r = 0; r < bin_mdef_n_ciphone(mdef); r++) {
p = bin_mdef_phone_id_nearest(mdef, (s3cipid_t) b,
(s3cipid_t) l, (s3cipid_t) r,
WORD_POSN_END);
rdiph_rc[b][l][r] = bin_mdef_pid2ssid(mdef, p);
}
}
}
else if (pronlen == 1) {
b = dict_pron(dict, w, 0);
E_DEBUG(1,("Building tables for single phone word %s phone %d = %s\n",
dict_wordstr(dict, w), b, bin_mdef_ciphone_str(mdef, b)));
/* Populate lrdiph_rc (and also ldiph_lc, rdiph_rc if needed) */
if (bitvec_is_clear(single, b)) {
populate_lrdiph(dict2pid, rdiph_rc, b);
bitvec_set(single, b);
}
}
}
bitvec_free(ldiph);
bitvec_free(rdiph);
bitvec_free(single);
/* Try to compress rdiph_rc into rdiph_rc_compressed */
compress_right_context_tree(dict2pid, rdiph_rc);
compress_left_right_context_tree(dict2pid);
ckd_free_3d(rdiph_rc);
dict2pid_report(dict2pid);
return dict2pid;
}
int
ps_alignment_populate(ps_alignment_t *al)
{
dict2pid_t *d2p;
dict_t *dict;
bin_mdef_t *mdef;
int i, lc;
/* Clear phone and state sequences. */
ps_alignment_vector_empty(&al->sseq);
ps_alignment_vector_empty(&al->state);
/* For each word, expand to phones/senone sequences. */
d2p = al->d2p;
dict = d2p->dict;
mdef = d2p->mdef;
lc = bin_mdef_silphone(mdef);
for (i = 0; i < al->word.n_ent; ++i) {
ps_alignment_entry_t *went = al->word.seq + i;
ps_alignment_entry_t *sent;
int wid = went->id.wid;
int len = dict_pronlen(dict, wid);
int j, rc;
if (i < al->word.n_ent - 1)
rc = dict_first_phone(dict, al->word.seq[i+1].id.wid);
else
rc = bin_mdef_silphone(mdef);
/* First phone. */
if ((sent = ps_alignment_vector_grow_one(&al->sseq)) == NULL) {
E_ERROR("Failed to add phone entry!\n");
return -1;
}
sent->id.pid.cipid = dict_first_phone(dict, wid);
sent->id.pid.tmatid = bin_mdef_pid2tmatid(mdef, sent->id.pid.cipid);
sent->start = went->start;
sent->duration = went->duration;
sent->parent = i;
went->child = (uint16)(sent - al->sseq.seq);
if (len == 1)
sent->id.pid.ssid
= dict2pid_lrdiph_rc(d2p, sent->id.pid.cipid, lc, rc);
else
sent->id.pid.ssid
= dict2pid_ldiph_lc(d2p, sent->id.pid.cipid,
dict_second_phone(dict, wid), lc);
oe_assert(sent->id.pid.ssid != BAD_SSID);
/* Internal phones. */
for (j = 1; j < len - 1; ++j) {
if ((sent = ps_alignment_vector_grow_one(&al->sseq)) == NULL) {
E_ERROR("Failed to add phone entry!\n");
return -1;
}
sent->id.pid.cipid = dict_pron(dict, wid, j);
sent->id.pid.tmatid = bin_mdef_pid2tmatid(mdef, sent->id.pid.cipid);
sent->id.pid.ssid = dict2pid_internal(d2p, wid, j);
oe_assert(sent->id.pid.ssid != BAD_SSID);
sent->start = went->start;
sent->duration = went->duration;
sent->parent = i;
}
/* Last phone. */
if (j < len) {
xwdssid_t *rssid;
oe_assert(j == len - 1);
if ((sent = ps_alignment_vector_grow_one(&al->sseq)) == NULL) {
E_ERROR("Failed to add phone entry!\n");
return -1;
}
sent->id.pid.cipid = dict_last_phone(dict, wid);
sent->id.pid.tmatid = bin_mdef_pid2tmatid(mdef, sent->id.pid.cipid);
rssid = dict2pid_rssid(d2p, sent->id.pid.cipid,
dict_second_last_phone(dict, wid));
sent->id.pid.ssid = rssid->ssid[rssid->cimap[rc]];
oe_assert(sent->id.pid.ssid != BAD_SSID);
sent->start = went->start;
sent->duration = went->duration;
sent->parent = i;
}
/* Update lc. Could just use sent->id.pid.cipid here but that
* seems needlessly obscure. */
lc = dict_last_phone(dict, wid);
}
/* For each senone sequence, expand to senones. (we could do this
* nested above but this makes it more clear and easier to
* refactor) */
for (i = 0; i < al->sseq.n_ent; ++i) {
ps_alignment_entry_t *pent = al->sseq.seq + i;
ps_alignment_entry_t *sent;
int j;
for (j = 0; j < bin_mdef_n_emit_state(mdef); ++j) {
if ((sent = ps_alignment_vector_grow_one(&al->state)) == NULL) {
E_ERROR("Failed to add state entry!\n");
return -1;
}
sent->id.senid = bin_mdef_sseq2sen(mdef, pent->id.pid.ssid, j);
oe_assert(sent->id.senid != BAD_SENID);
sent->start = pent->start;
sent->duration = pent->duration;
sent->parent = i;
if (j == 0)
pent->child = (uint16)(sent - al->state.seq);
}
}
return 0;
}
int
dict2pid_add_word(dict2pid_t *d2p,
int32 wid)
{
bin_mdef_t *mdef = d2p->mdef;
dict_t *d = d2p->dict;
if (dict_pronlen(d, wid) > 1) {
s3cipid_t l;
/* Make sure we have left and right context diphones for this
* word. */
if (d2p->ldiph_lc[dict_first_phone(d, wid)][dict_second_phone(d, wid)][0]
== BAD_S3SSID) {
E_INFO("Filling in left-context diphones for %s(?,%s)\n",
bin_mdef_ciphone_str(mdef, dict_first_phone(d, wid)),
bin_mdef_ciphone_str(mdef, dict_second_phone(d, wid)));
for (l = 0; l < bin_mdef_n_ciphone(mdef); l++) {
s3ssid_t p
= bin_mdef_phone_id_nearest(mdef,
dict_first_phone(d, wid), l,
dict_second_phone(d, wid),
WORD_POSN_BEGIN);
d2p->ldiph_lc[dict_first_phone(d, wid)][dict_second_phone(d, wid)][l]
= bin_mdef_pid2ssid(mdef, p);
}
}
if (d2p->rssid[dict_last_phone(d, wid)][dict_second_last_phone(d, wid)].n_ssid
== 0) {
s3ssid_t *rmap;
s3ssid_t *tmpssid;
s3cipid_t *tmpcimap;
s3cipid_t r;
E_INFO("Filling in right-context diphones for %s(%s,?)\n",
bin_mdef_ciphone_str(mdef, dict_last_phone(d, wid)),
bin_mdef_ciphone_str(mdef, dict_second_last_phone(d, wid)));
rmap = ckd_calloc(bin_mdef_n_ciphone(mdef), sizeof(*rmap));
for (r = 0; r < bin_mdef_n_ciphone(mdef); r++) {
s3ssid_t p
= bin_mdef_phone_id_nearest(mdef,
dict_last_phone(d, wid),
dict_second_last_phone(d, wid), r,
WORD_POSN_END);
rmap[r] = bin_mdef_pid2ssid(mdef, p);
}
tmpssid = ckd_calloc(bin_mdef_n_ciphone(mdef), sizeof(*tmpssid));
tmpcimap = ckd_calloc(bin_mdef_n_ciphone(mdef), sizeof(*tmpcimap));
compress_table(rmap, tmpssid, tmpcimap, bin_mdef_n_ciphone(mdef));
for (r = 0; r < mdef->n_ciphone && tmpssid[r] != BAD_S3SSID; r++)
;
d2p->rssid[dict_last_phone(d, wid)][dict_second_last_phone(d, wid)].ssid = tmpssid;
d2p->rssid[dict_last_phone(d, wid)][dict_second_last_phone(d, wid)].cimap = tmpcimap;
d2p->rssid[dict_last_phone(d, wid)][dict_second_last_phone(d, wid)].n_ssid = r;
ckd_free(rmap);
}
}
else {
/* Make sure we have a left-right context triphone entry for
* this word. */
E_INFO("Filling in context triphones for %s(?,?)\n",
bin_mdef_ciphone_str(mdef, dict_first_phone(d, wid)));
if (d2p->lrdiph_rc[dict_first_phone(d, wid)][0][0] == BAD_S3SSID) {
populate_lrdiph(d2p, NULL, dict_first_phone(d, wid));
}
}
return 0;
}
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);
}
}
}
