jsgf_rule_t *
jsgf_define_rule(jsgf_t * jsgf, char *name, jsgf_rhs_t * rhs,
int is_public)
{
jsgf_rule_t *rule;
void *val;
if (name == NULL) {
name = ckd_malloc(strlen(jsgf->name) + 16);
sprintf(name, "<%s.g%05d>", jsgf->name,
hash_table_inuse(jsgf->rules));
}
else {
char *newname;
newname = jsgf_fullname(jsgf, name);
name = newname;
}
rule = ckd_calloc(1, sizeof(*rule));
rule->refcnt = 1;
rule->name = ckd_salloc(name);
rule->rhs = rhs;
rule->is_public = is_public;
E_INFO("Defined rule: %s%s\n",
rule->is_public ? "PUBLIC " : "", rule->name);
val = hash_table_enter(jsgf->rules, name, rule);
if (val != (void *) rule) {
E_WARN("Multiply defined symbol: %s\n", name);
}
return rule;
}
static void
build_widmap(ngram_model_t * base, logmath_t * lmath, int32 n)
{
ngram_model_set_t *set = (ngram_model_set_t *) base;
ngram_model_t **models = set->lms;
hash_table_t *vocab;
glist_t hlist;
gnode_t *gn;
int32 i;
/* Construct a merged vocabulary and a set of word-ID mappings. */
vocab = hash_table_new(models[0]->n_words, FALSE);
/* Create the set of merged words. */
for (i = 0; i < set->n_models; ++i) {
int32 j;
for (j = 0; j < models[i]->n_words; ++j) {
/* Ignore collisions. */
(void) hash_table_enter_int32(vocab, models[i]->word_str[j],
j);
}
}
/* Create the array of words, then sort it. */
if (hash_table_lookup(vocab, "<UNK>", NULL) != 0)
(void) hash_table_enter_int32(vocab, "<UNK>", 0);
/* Now we know the number of unigrams, initialize the base model. */
ngram_model_init(base, &ngram_model_set_funcs, lmath, n,
hash_table_inuse(vocab));
base->writable = FALSE; /* We will reuse the pointers from the submodels. */
i = 0;
hlist = hash_table_tolist(vocab, NULL);
for (gn = hlist; gn; gn = gnode_next(gn)) {
hash_entry_t *ent = gnode_ptr(gn);
base->word_str[i++] = (char *) ent->key;
}
glist_free(hlist);
qsort(base->word_str, base->n_words, sizeof(*base->word_str),
my_compare);
/* Now create the word ID mappings. */
if (set->widmap)
ckd_free_2d((void **) set->widmap);
set->widmap = (int32 **) ckd_calloc_2d(base->n_words, set->n_models,
sizeof(**set->widmap));
for (i = 0; i < base->n_words; ++i) {
int32 j;
/* Also create the master wid mapping. */
(void) hash_table_enter_int32(base->wid, base->word_str[i], i);
/* printf("%s: %d => ", base->word_str[i], i); */
for (j = 0; j < set->n_models; ++j) {
set->widmap[i][j] = ngram_wid(models[j], base->word_str[i]);
/* printf("%d ", set->widmap[i][j]); */
}
/* printf("\n"); */
}
hash_table_free(vocab);
}
fsg_model_t *
fsg_model_read(FILE * fp, logmath_t * lmath, float32 lw)
{
fsg_model_t *fsg;
hash_table_t *vocab;
hash_iter_t *itor;
int32 lastwid;
char **wordptr;
char *lineptr;
char *fsgname;
int32 lineno;
int32 n, i, j;
int n_state, n_trans, n_null_trans;
glist_t nulls;
float32 p;
lineno = 0;
vocab = hash_table_new(32, FALSE);
wordptr = NULL;
lineptr = NULL;
nulls = NULL;
fsgname = NULL;
fsg = NULL;
/* Scan upto FSG_BEGIN header */
for (;;) {
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if (n < 0) {
E_ERROR("%s declaration missing\n", FSG_MODEL_BEGIN_DECL);
goto parse_error;
}
if ((strcmp(wordptr[0], FSG_MODEL_BEGIN_DECL) == 0)) {
if (n > 2) {
E_ERROR("Line[%d]: malformed FSG_BEGIN declaration\n",
lineno);
goto parse_error;
}
break;
}
}
/* Save FSG name, or it will get clobbered below :(.
* If name is missing, try the default.
*/
if (n == 2) {
fsgname = ckd_salloc(wordptr[1]);
}
else {
E_WARN("FSG name is missing\n");
fsgname = ckd_salloc("unknown");
}
/* Read #states */
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if ((n != 2)
|| ((strcmp(wordptr[0], FSG_MODEL_N_DECL) != 0)
&& (strcmp(wordptr[0], FSG_MODEL_NUM_STATES_DECL) != 0))
|| (sscanf(wordptr[1], "%d", &n_state) != 1)
|| (n_state <= 0)) {
E_ERROR
("Line[%d]: #states declaration line missing or malformed\n",
lineno);
goto parse_error;
}
/* Now create the FSG. */
fsg = fsg_model_init(fsgname, lmath, lw, n_state);
ckd_free(fsgname);
fsgname = NULL;
/* Read start state */
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if ((n != 2)
|| ((strcmp(wordptr[0], FSG_MODEL_S_DECL) != 0)
&& (strcmp(wordptr[0], FSG_MODEL_START_STATE_DECL) != 0))
|| (sscanf(wordptr[1], "%d", &(fsg->start_state)) != 1)
|| (fsg->start_state < 0)
|| (fsg->start_state >= fsg->n_state)) {
E_ERROR
("Line[%d]: start state declaration line missing or malformed\n",
lineno);
goto parse_error;
}
/* Read final state */
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if ((n != 2)
|| ((strcmp(wordptr[0], FSG_MODEL_F_DECL) != 0)
&& (strcmp(wordptr[0], FSG_MODEL_FINAL_STATE_DECL) != 0))
|| (sscanf(wordptr[1], "%d", &(fsg->final_state)) != 1)
|| (fsg->final_state < 0)
|| (fsg->final_state >= fsg->n_state)) {
E_ERROR
("Line[%d]: final state declaration line missing or malformed\n",
lineno);
goto parse_error;
}
/* Read transitions */
lastwid = 0;
n_trans = n_null_trans = 0;
for (;;) {
int32 wid, tprob;
n = nextline_str2words(fp, &lineno, &lineptr, &wordptr);
if (n <= 0) {
E_ERROR("Line[%d]: transition or FSG_END statement expected\n",
lineno);
goto parse_error;
}
if ((strcmp(wordptr[0], FSG_MODEL_END_DECL) == 0)) {
break;
}
if ((strcmp(wordptr[0], FSG_MODEL_T_DECL) == 0)
|| (strcmp(wordptr[0], FSG_MODEL_TRANSITION_DECL) == 0)) {
if (((n != 4) && (n != 5))
|| (sscanf(wordptr[1], "%d", &i) != 1)
|| (sscanf(wordptr[2], "%d", &j) != 1)
|| (i < 0) || (i >= fsg->n_state)
|| (j < 0) || (j >= fsg->n_state)) {
E_ERROR
("Line[%d]: transition spec malformed; Expecting: from-state to-state trans-prob [word]\n",
lineno);
goto parse_error;
}
p = atof_c(wordptr[3]);
if ((p <= 0.0) || (p > 1.0)) {
E_ERROR
("Line[%d]: transition spec malformed; Expecting float as transition probability\n",
lineno);
goto parse_error;
}
}
else {
E_ERROR("Line[%d]: transition or FSG_END statement expected\n",
lineno);
goto parse_error;
}
tprob = (int32) (logmath_log(lmath, p) * fsg->lw);
/* Add word to "dictionary". */
if (n > 4) {
if (hash_table_lookup_int32(vocab, wordptr[4], &wid) < 0) {
(void) hash_table_enter_int32(vocab,
ckd_salloc(wordptr[4]),
lastwid);
wid = lastwid;
++lastwid;
}
fsg_model_trans_add(fsg, i, j, tprob, wid);
++n_trans;
}
else {
if (fsg_model_null_trans_add(fsg, i, j, tprob) == 1) {
++n_null_trans;
nulls =
glist_add_ptr(nulls, fsg_model_null_trans(fsg, i, j));
}
}
}
E_INFO("FSG: %d states, %d unique words, %d transitions (%d null)\n",
fsg->n_state, hash_table_inuse(vocab), n_trans, n_null_trans);
/* Now create a string table from the "dictionary" */
fsg->n_word = hash_table_inuse(vocab);
fsg->n_word_alloc = fsg->n_word + 10; /* Pad it a bit. */
fsg->vocab = ckd_calloc(fsg->n_word_alloc, sizeof(*fsg->vocab));
for (itor = hash_table_iter(vocab); itor;
itor = hash_table_iter_next(itor)) {
char const *word = hash_entry_key(itor->ent);
int32 wid = (int32) (long) hash_entry_val(itor->ent);
fsg->vocab[wid] = (char *) word;
}
hash_table_free(vocab);
/* Do transitive closure on null transitions */
nulls = fsg_model_null_trans_closure(fsg, nulls);
glist_free(nulls);
ckd_free(lineptr);
ckd_free(wordptr);
return fsg;
parse_error:
for (itor = hash_table_iter(vocab); itor;
itor = hash_table_iter_next(itor))
ckd_free((char *) hash_entry_key(itor->ent));
glist_free(nulls);
hash_table_free(vocab);
ckd_free(fsgname);
ckd_free(lineptr);
ckd_free(wordptr);
fsg_model_free(fsg);
return NULL;
}
