int
main(int argc, char *argv[])
{
listelem_alloc_t *le;
struct bogus *bogus1, *bogus2;
int i;
TEST_ASSERT(le = listelem_alloc_init(sizeof(struct bogus)));
bogus1 = listelem_malloc(le);
bogus1->str = "hello";
bogus1->foobie = 42;
bogus2 = listelem_malloc(le);
bogus2->str = "goodbye";
bogus2->foobie = 69;
TEST_EQUAL(bogus1->foobie, 42);
TEST_EQUAL(0, strcmp(bogus1->str, "hello"));
listelem_free(le, bogus1);
listelem_free(le, bogus2);
listelem_alloc_free(le);
TEST_ASSERT(le = listelem_alloc_init(sizeof(struct bogus)));
listelem_stats(le);
for (i = 0; i < 60; ++i)
bogus1 = listelem_malloc(le);
listelem_stats(le);
listelem_alloc_free(le);
{
struct bogus *bogus[600];
int32 bogus_id[600];
le = listelem_alloc_init(sizeof(struct bogus));
for (i = 0; i < 600; ++i)
bogus[i] = listelem_malloc_id(le, bogus_id + i);
listelem_stats(le);
for (i = 0; i < 600; ++i) {
TEST_EQUAL(bogus[i], listelem_get_item(le, bogus_id[i]));
}
for (i = 0; i < 600; ++i)
listelem_free(le, bogus[i]);
listelem_stats(le);
for (i = 0; i < 600; ++i)
bogus[i] = listelem_malloc_id(le, bogus_id + i);
listelem_stats(le);
for (i = 0; i < 600; ++i)
TEST_EQUAL(bogus[i], listelem_get_item(le, bogus_id[i]));
listelem_alloc_free(le);
}
return 0;
}
void
fsg_model_trans_add(fsg_model_t * fsg,
int32 from, int32 to, int32 logp, int32 wid)
{
fsg_link_t *link;
glist_t gl;
gnode_t *gn;
if (fsg->trans[from].trans == NULL)
fsg->trans[from].trans = hash_table_new(5, HASH_CASE_YES);
/* Check for duplicate link (i.e., link already exists with label=wid) */
for (gn = gl = fsg_model_trans(fsg, from, to); gn; gn = gnode_next(gn)) {
link = (fsg_link_t *) gnode_ptr(gn);
if (link->wid == wid) {
if (link->logs2prob < logp)
link->logs2prob = logp;
return;
}
}
/* Create transition object */
link = listelem_malloc(fsg->link_alloc);
link->from_state = from;
link->to_state = to;
link->logs2prob = logp;
link->wid = wid;
/* Add it to the list of transitions and update the hash table */
gl = glist_add_ptr(gl, (void *) link);
hash_table_replace_bkey(fsg->trans[from].trans,
(char const *) &link->to_state,
sizeof(link->to_state), gl);
}
int
fsg_model_add_alt(fsg_model_t * fsg, char const *baseword,
char const *altword)
{
int i, basewid, altwid;
int ntrans;
/* FIXME: This will get slow, eventually... */
for (basewid = 0; basewid < fsg->n_word; ++basewid)
if (0 == strcmp(fsg->vocab[basewid], baseword))
break;
if (basewid == fsg->n_word) {
E_ERROR("Base word %s not present in FSG vocabulary!\n", baseword);
return -1;
}
altwid = fsg_model_word_add(fsg, altword);
if (fsg->altwords == NULL)
fsg->altwords = bitvec_alloc(fsg->n_word_alloc);
bitvec_set(fsg->altwords, altwid);
E_DEBUG(2,("Adding alternate word transitions (%s,%s) to FSG\n",
baseword, altword));
/* Look for all transitions involving baseword and duplicate them. */
/* FIXME: This will also get slow, eventually... */
ntrans = 0;
for (i = 0; i < fsg->n_state; ++i) {
hash_iter_t *itor;
if (fsg->trans[i].trans == NULL)
continue;
for (itor = hash_table_iter(fsg->trans[i].trans); itor;
itor = hash_table_iter_next(itor)) {
glist_t trans;
gnode_t *gn;
trans = hash_entry_val(itor->ent);
for (gn = trans; gn; gn = gnode_next(gn)) {
fsg_link_t *fl = gnode_ptr(gn);
if (fl->wid == basewid) {
fsg_link_t *link;
/* Create transition object */
link = listelem_malloc(fsg->link_alloc);
link->from_state = fl->from_state;
link->to_state = fl->to_state;
link->logs2prob = fl->logs2prob; /* FIXME!!!??? */
link->wid = altwid;
trans =
glist_add_ptr(trans, (void *) link);
++ntrans;
}
}
hash_entry_val(itor->ent) = trans;
}
}
E_DEBUG(2,("Added %d alternate word transitions\n", ntrans));
return ntrans;
}
int32
fsg_model_tag_trans_add(fsg_model_t * fsg, int32 from, int32 to,
int32 logp, int32 wid)
{
fsg_link_t *link, *link2;
/* Check for transition probability */
if (logp > 0) {
E_FATAL("Null transition prob must be <= 1.0 (state %d -> %d)\n",
from, to);
}
/* Self-loop null transitions (with prob <= 1.0) are redundant */
if (from == to)
return -1;
if (fsg->trans[from].null_trans == NULL)
fsg->trans[from].null_trans = hash_table_new(5, HASH_CASE_YES);
/* Check for a duplicate link; if found, keep the higher prob */
link = fsg_model_null_trans(fsg, from, to);
if (link) {
if (link->logs2prob < logp) {
link->logs2prob = logp;
return 0;
}
else
return -1;
}
/* Create null transition object */
link = listelem_malloc(fsg->link_alloc);
link->from_state = from;
link->to_state = to;
link->logs2prob = logp;
link->wid = -1;
link2 = (fsg_link_t *)
hash_table_enter_bkey(fsg->trans[from].null_trans,
(char const *) &link->to_state,
sizeof(link->to_state), link);
assert(link == link2);
return 1;
}
static void
load_tginfo(NGRAM_MODEL_TYPE *model, int32 lw1, int32 lw2)
{
int32 i, n, b, t;
bigram_t *bg;
tginfo_t *tginfo;
/* First allocate space for tg information for bg lw1,lw2 */
tginfo = (tginfo_t *) listelem_malloc(model->lm3g.le);
tginfo->w1 = lw1;
tginfo->tg = NULL;
tginfo->next = model->lm3g.tginfo[lw2];
model->lm3g.tginfo[lw2] = tginfo;
/* Locate bigram lw1,lw2 */
b = model->lm3g.unigrams[lw1].bigrams;
n = model->lm3g.unigrams[lw1 + 1].bigrams - b;
bg = model->lm3g.bigrams + b;
if ((n > 0) && ((i = find_bg(bg, n, lw2)) >= 0)) {
tginfo->bowt = model->lm3g.bo_wt2[bg[i].bo_wt2].l;
/* Find t = Absolute first trigram index for bigram lw1,lw2 */
b += i; /* b = Absolute index of bigram lw1,lw2 on disk */
t = FIRST_TG(model, b);
tginfo->tg = model->lm3g.trigrams + t;
/* Find #tg for bigram w1,w2 */
tginfo->n_tg = FIRST_TG(model, b + 1) - t;
}
else { /* No bigram w1,w2 */
tginfo->bowt = 0;
tginfo->n_tg = 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;
}
}
/**
* 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);
}
