static int
phone_loop_search_reinit(ps_search_t *search, dict_t *dict, dict2pid_t *d2p)
{
phone_loop_search_t *pls = (phone_loop_search_t *)search;
cmd_ln_t *config = ps_search_config(search);
acmod_t *acmod = ps_search_acmod(search);
int i;
/* Free old dict2pid, dict, if necessary. */
ps_search_base_reinit(search, dict, d2p);
/* Initialize HMM context. */
if (pls->hmmctx)
hmm_context_free(pls->hmmctx);
pls->hmmctx = hmm_context_init(bin_mdef_n_emit_state(acmod->mdef),
acmod->tmat->tp, NULL, acmod->mdef->sseq);
if (pls->hmmctx == NULL)
return -1;
/* Initialize penalty storage */
pls->n_phones = bin_mdef_n_ciphone(acmod->mdef);
pls->window = cmd_ln_int32_r(config, "-pl_window");
if (pls->penalties)
ckd_free(pls->penalties);
pls->penalties = (int32 *)ckd_calloc(pls->n_phones, sizeof(*pls->penalties));
if (pls->pen_buf)
ckd_free_2d(pls->pen_buf);
pls->pen_buf = (int32 **)ckd_calloc_2d(pls->window, pls->n_phones, sizeof(**pls->pen_buf));
/* Initialize phone HMMs. */
if (pls->hmms) {
for (i = 0; i < pls->n_phones; ++i)
hmm_deinit((hmm_t *)&pls->hmms[i]);
ckd_free(pls->hmms);
}
pls->hmms = (hmm_t *)ckd_calloc(pls->n_phones, sizeof(*pls->hmms));
for (i = 0; i < pls->n_phones; ++i) {
hmm_init(pls->hmmctx, (hmm_t *)&pls->hmms[i],
FALSE,
bin_mdef_pid2ssid(acmod->mdef, i),
bin_mdef_pid2tmatid(acmod->mdef, i));
}
pls->penalty_weight = cmd_ln_float64_r(config, "-pl_weight");
pls->beam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pl_beam")) >> SENSCR_SHIFT;
pls->pbeam = logmath_log(acmod->lmath, cmd_ln_float64_r(config, "-pl_pbeam")) >> SENSCR_SHIFT;
pls->pip = logmath_log(acmod->lmath, cmd_ln_float32_r(config, "-pl_pip")) >> SENSCR_SHIFT;
E_INFO("State beam %d Phone exit beam %d Insertion penalty %d\n",
pls->beam, pls->pbeam, pls->pip);
return 0;
}
static void
ngram_fwdflat_allocate_1ph(ngram_search_t *ngs)
{
dict_t *dict = ps_search_dict(ngs);
int n_words = ps_search_n_words(ngs);
int i, w;
/* Allocate single-phone words, since they won't have
* been allocated for us by fwdtree initialization. */
ngs->n_1ph_words = 0;
for (w = 0; w < n_words; w++) {
if (dict_is_single_phone(dict, w))
++ngs->n_1ph_words;
}
ngs->single_phone_wid = ckd_calloc(ngs->n_1ph_words,
sizeof(*ngs->single_phone_wid));
ngs->rhmm_1ph = ckd_calloc(ngs->n_1ph_words, sizeof(*ngs->rhmm_1ph));
i = 0;
for (w = 0; w < n_words; w++) {
if (!dict_is_single_phone(dict, w))
continue;
/* DICT2PID location */
ngs->rhmm_1ph[i].ciphone = dict_first_phone(dict, w);
ngs->rhmm_1ph[i].ci2phone = bin_mdef_silphone(ps_search_acmod(ngs)->mdef);
hmm_init(ngs->hmmctx, &ngs->rhmm_1ph[i].hmm, TRUE,
/* ssid */ bin_mdef_pid2ssid(ps_search_acmod(ngs)->mdef,
ngs->rhmm_1ph[i].ciphone),
/* tmatid */ bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef,
ngs->rhmm_1ph[i].ciphone));
ngs->rhmm_1ph[i].next = NULL;
ngs->word_chan[w] = (chan_t *) &(ngs->rhmm_1ph[i]);
ngs->single_phone_wid[i] = w;
i++;
}
}
/*
* Add the word emitted by the given transition (fsglink) to the given lextree
* (rooted at root), and return the new lextree root. (There may actually be
* several root nodes, maintained in a linked list via fsg_pnode_t.sibling.
* "root" is the head of this list.)
* lclist, rclist: sets of left and right context phones for this link.
* alloc_head: head of a linear list of all allocated pnodes for the parent
* FSG state, kept elsewhere and updated by this routine.
*/
static fsg_pnode_t *
psubtree_add_trans(fsg_lextree_t *lextree,
fsg_pnode_t * root,
fsg_glist_linklist_t **curglist,
fsg_link_t * fsglink,
int16 *lclist, int16 *rclist,
fsg_pnode_t ** alloc_head)
{
int32 silcipid; /* Silence CI phone ID */
int32 pronlen; /* Pronunciation length */
int32 wid; /* FSG (not dictionary!!) word ID */
int32 dictwid; /* Dictionary (not FSG!!) word ID */
int32 ssid; /* Senone Sequence ID */
int32 tmatid;
gnode_t *gn;
fsg_pnode_t *pnode, *pred, *head;
int32 n_ci, p, lc, rc;
glist_t lc_pnodelist; /* Temp pnodes list for different left contexts */
glist_t rc_pnodelist; /* Temp pnodes list for different right contexts */
int32 i, j;
int n_lc_alloc = 0, n_int_alloc = 0, n_rc_alloc = 0;
silcipid = bin_mdef_silphone(lextree->mdef);
n_ci = bin_mdef_n_ciphone(lextree->mdef);
wid = fsg_link_wid(fsglink);
assert(wid >= 0); /* Cannot be a null transition */
dictwid = dict_wordid(lextree->dict,
fsg_model_word_str(lextree->fsg, wid));
pronlen = dict_pronlen(lextree->dict, dictwid);
assert(pronlen >= 1);
assert(lclist[0] >= 0); /* At least one phonetic context provided */
assert(rclist[0] >= 0);
head = *alloc_head;
pred = NULL;
if (pronlen == 1) { /* Single-phone word */
int ci = dict_first_phone(lextree->dict, dictwid);
/* Only non-filler words are mpx */
if (dict_filler_word(lextree->dict, dictwid)) {
/*
* Left diphone ID for single-phone words already assumes SIL is right
* context; only left contexts need to be handled.
*/
lc_pnodelist = NULL;
for (i = 0; lclist[i] >= 0; i++) {
lc = lclist[i];
ssid = dict2pid_lrdiph_rc(lextree->d2p, ci, lc, silcipid);
tmatid = bin_mdef_pid2tmatid(lextree->mdef, dict_first_phone(lextree->dict, dictwid));
/* Check if this ssid already allocated for some other context */
for (gn = lc_pnodelist; gn; gn = gnode_next(gn)) {
pnode = (fsg_pnode_t *) gnode_ptr(gn);
if (hmm_nonmpx_ssid(&pnode->hmm) == ssid) {
/* already allocated; share it for this context phone */
fsg_pnode_add_ctxt(pnode, lc);
break;
}
}
if (!gn) { /* ssid not already allocated */
pnode =
(fsg_pnode_t *) ckd_calloc(1, sizeof(fsg_pnode_t));
pnode->ctx = lextree->ctx;
pnode->next.fsglink = fsglink;
pnode->logs2prob =
(fsg_link_logs2prob(fsglink) >> SENSCR_SHIFT)
+ lextree->wip + lextree->pip;
pnode->ci_ext = dict_first_phone(lextree->dict, dictwid);
pnode->ppos = 0;
pnode->leaf = TRUE;
pnode->sibling = root; /* All root nodes linked together */
fsg_pnode_add_ctxt(pnode, lc); /* Initially zeroed by calloc above */
pnode->alloc_next = head;
head = pnode;
root = pnode;
++n_lc_alloc;
hmm_init(lextree->ctx, &pnode->hmm, FALSE, ssid, tmatid);
lc_pnodelist =
glist_add_ptr(lc_pnodelist, (void *) pnode);
}
}
/* FIXME: Somewhat the same as the above function, needs refactoring */
int
ps_alignment_populate_ci(ps_alignment_t *al)
{
dict2pid_t *d2p;
dict_t *dict;
bin_mdef_t *mdef;
int i;
/* 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;
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;
for (j = 0; j < len; ++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 = bin_mdef_pid2ssid(mdef, sent->id.pid.cipid);
oe_assert(sent->id.pid.ssid != BAD_SSID);
sent->start = went->start;
sent->duration = went->duration;
sent->parent = i;
}
}
/* 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
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;
}
/**
* Build HMM network for one utterance of fwdflat search.
*/
static void
build_fwdflat_chan(ngram_search_t *ngs)
{
int32 i, wid, p;
root_chan_t *rhmm;
chan_t *hmm, *prevhmm;
dict_t *dict;
dict2pid_t *d2p;
dict = ps_search_dict(ngs);
d2p = ps_search_dict2pid(ngs);
/* Build word HMMs for each word in the lattice. */
for (i = 0; ngs->fwdflat_wordlist[i] >= 0; i++) {
wid = ngs->fwdflat_wordlist[i];
/* Single-phone words are permanently allocated */
if (dict_is_single_phone(dict, wid))
continue;
assert(ngs->word_chan[wid] == NULL);
/* Multiplex root HMM for first phone (one root per word, flat
* lexicon). diphone is irrelevant here, for the time being,
* at least. */
rhmm = listelem_malloc(ngs->root_chan_alloc);
rhmm->ci2phone = dict_second_phone(dict, wid);
rhmm->ciphone = dict_first_phone(dict, wid);
rhmm->next = NULL;
hmm_init(ngs->hmmctx, &rhmm->hmm, TRUE,
bin_mdef_pid2ssid(ps_search_acmod(ngs)->mdef, rhmm->ciphone),
bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef, rhmm->ciphone));
/* HMMs for word-internal phones */
prevhmm = NULL;
for (p = 1; p < dict_pronlen(dict, wid) - 1; p++) {
hmm = listelem_malloc(ngs->chan_alloc);
hmm->ciphone = dict_pron(dict, wid, p);
hmm->info.rc_id = (p == dict_pronlen(dict, wid) - 1) ? 0 : -1;
hmm->next = NULL;
hmm_init(ngs->hmmctx, &hmm->hmm, FALSE,
dict2pid_internal(d2p,wid,p),
bin_mdef_pid2tmatid(ps_search_acmod(ngs)->mdef, hmm->ciphone));
if (prevhmm)
prevhmm->next = hmm;
else
rhmm->next = hmm;
prevhmm = hmm;
}
/* Right-context phones */
ngram_search_alloc_all_rc(ngs, wid);
/* Link in just allocated right-context phones */
if (prevhmm)
prevhmm->next = ngs->word_chan[wid];
else
rhmm->next = ngs->word_chan[wid];
ngs->word_chan[wid] = (chan_t *) rhmm;
}
}