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);
}
void
ngram_model_set_map_words(ngram_model_t * base,
const char **words, int32 n_words)
{
ngram_model_set_t *set = (ngram_model_set_t *) base;
int32 i;
/* Recreate the word mapping. */
if (base->writable) {
for (i = 0; i < base->n_words; ++i) {
ckd_free(base->word_str[i]);
}
}
ckd_free(base->word_str);
ckd_free_2d((void **) set->widmap);
base->writable = TRUE;
base->n_words = base->n_1g_alloc = n_words;
base->word_str = ckd_calloc(n_words, sizeof(*base->word_str));
set->widmap =
(int32 **) ckd_calloc_2d(n_words, set->n_models,
sizeof(**set->widmap));
hash_table_empty(base->wid);
for (i = 0; i < n_words; ++i) {
int32 j;
base->word_str[i] = ckd_salloc(words[i]);
(void) hash_table_enter_int32(base->wid, base->word_str[i], i);
for (j = 0; j < set->n_models; ++j) {
set->widmap[i][j] = ngram_wid(set->lms[j], base->word_str[i]);
}
}
}
int
ngram_model_casefold(ngram_model_t * model, int kase)
{
int writable, i;
hash_table_t *new_wid;
/* Were word strings already allocated? */
writable = model->writable;
/* Either way, we are going to allocate some word strings. */
model->writable = TRUE;
/* And, don't forget, we need to rebuild the word to unigram ID
* mapping. */
new_wid = hash_table_new(model->n_words, FALSE);
for (i = 0; i < model->n_words; ++i) {
char *outstr;
if (writable) {
outstr = model->word_str[i];
}
else {
outstr = ckd_salloc(model->word_str[i]);
}
/* Don't case-fold <tags> or [classes] */
if (outstr[0] == '<' || outstr[0] == '[') {
}
else {
switch (kase) {
case NGRAM_UPPER:
ucase(outstr);
break;
case NGRAM_LOWER:
lcase(outstr);
break;
default:
;
}
}
model->word_str[i] = outstr;
/* Now update the hash table. We might have terrible
* collisions here, so warn about them. */
if (hash_table_enter_int32(new_wid, model->word_str[i], i) != i) {
E_WARN("Duplicate word in dictionary after conversion: %s\n",
model->word_str[i]);
}
}
/* Swap out the hash table. */
hash_table_free(model->wid);
model->wid = new_wid;
return 0;
}
/**
* Add a word to the word string and ID mapping.
*/
int32
ngram_add_word_internal(ngram_model_t * model,
const char *word, int32 classid)
{
/* Check for hash collisions. */
int32 wid;
if (hash_table_lookup_int32(model->wid, word, &wid) == 0) {
E_WARN("Omit duplicate word '%s'\n", word);
return wid;
}
/* Take the next available word ID */
wid = model->n_words;
if (classid >= 0) {
wid = NGRAM_CLASSWID(wid, classid);
}
/* Reallocate word_str if necessary. */
if (model->n_words >= model->n_1g_alloc) {
model->n_1g_alloc += UG_ALLOC_STEP;
model->word_str = ckd_realloc(model->word_str,
sizeof(*model->word_str) *
model->n_1g_alloc);
}
/* Add the word string in the appropriate manner. */
/* Class words are always dynamically allocated. */
model->word_str[model->n_words] = ckd_salloc(word);
/* Now enter it into the hash table. */
if (hash_table_enter_int32
(model->wid, model->word_str[model->n_words], wid) != wid) {
E_ERROR
("Hash insertion failed for word %s => %p (should not happen)\n",
model->word_str[model->n_words], (void *) (long) (wid));
}
/* Increment number of words. */
++model->n_words;
return wid;
}
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;
}
s3wid_t
dict_add_word(dict_t * d, char const *word, s3cipid_t const * p, int32 np)
{
int32 len;
dictword_t *wordp;
s3wid_t newwid;
char *wword;
if (d->n_word >= d->max_words) {
E_INFO("Reallocating to %d KiB for word entries\n",
(d->max_words + S3DICT_INC_SZ) * sizeof(dictword_t) / 1024);
d->word =
(dictword_t *) ckd_realloc(d->word,
(d->max_words +
S3DICT_INC_SZ) * sizeof(dictword_t));
d->max_words = d->max_words + S3DICT_INC_SZ;
}
wordp = d->word + d->n_word;
wordp->word = (char *) ckd_salloc(word); /* Freed in dict_free */
/* Associate word string with d->n_word in hash table */
if (hash_table_enter_int32(d->ht, wordp->word, d->n_word) != d->n_word) {
ckd_free(wordp->word);
wordp->word = NULL;
return BAD_S3WID;
}
/* Fill in word entry, and set defaults */
if (p && (np > 0)) {
wordp->ciphone = (s3cipid_t *) ckd_malloc(np * sizeof(s3cipid_t)); /* Freed in dict_free */
memcpy(wordp->ciphone, p, np * sizeof(s3cipid_t));
wordp->pronlen = np;
}
else {
wordp->ciphone = NULL;
wordp->pronlen = 0;
}
wordp->alt = BAD_S3WID;
wordp->basewid = d->n_word;
/* Determine base/alt wids */
wword = ckd_salloc(word);
if ((len = dict_word2basestr(wword)) > 0) {
int32 w;
/* Truncated to a baseword string; find its ID */
if (hash_table_lookup_int32(d->ht, wword, &w) < 0) {
E_ERROR("Missing base word for: %s\n", word);
ckd_free(wword);
ckd_free(wordp->word);
wordp->word = NULL;
return BAD_S3WID;
}
/* Link into alt list */
wordp->basewid = w;
wordp->alt = d->word[w].alt;
d->word[w].alt = d->n_word;
}
ckd_free(wword);
newwid = d->n_word++;
return newwid;
}
int
ngram_model_recode(ngram_model_t *model, const char *from, const char *to)
{
iconv_t ic;
char *outbuf;
size_t maxlen;
int i, writable;
hash_table_t *new_wid;
/* FIXME: Need to do a special case thing for the GB-HEX encoding
* used in Sphinx3 Mandarin models. */
if ((ic = iconv_open(to, from)) == (iconv_t)-1) {
E_ERROR_SYSTEM("iconv_open() failed");
return -1;
}
/* iconv(3) is a piece of crap and won't accept a NULL out buffer,
* unlike wcstombs(3). So we have to either call it over and over
* again until our buffer is big enough, or call it with a huge
* buffer and then copy things back to the output. We will use a
* mix of these two approaches here. We'll keep a single big
* buffer around, and expand it as necessary.
*/
maxlen = 0;
for (i = 0; i < model->n_words; ++i) {
if (strlen(model->word_str[i]) > maxlen)
maxlen = strlen(model->word_str[i]);
}
/* Were word strings already allocated? */
writable = model->writable;
/* Either way, we are going to allocate some word strings. */
model->writable = TRUE;
/* Really should be big enough except for pathological cases. */
maxlen = maxlen * sizeof(int) + 15;
outbuf = ckd_calloc(maxlen, 1);
/* And, don't forget, we need to rebuild the word to unigram ID
* mapping. */
new_wid = hash_table_new(model->n_words, FALSE);
for (i = 0; i < model->n_words; ++i) {
ICONV_CONST char *in;
char *out;
size_t inleft, outleft, result;
start_conversion:
in = (ICONV_CONST char *)model->word_str[i];
/* Yes, this assumes that we don't have any NUL bytes. */
inleft = strlen(in);
out = outbuf;
outleft = maxlen;
while ((result = iconv(ic, &in, &inleft, &out, &outleft)) == (size_t)-1) {
if (errno != E2BIG) {
/* FIXME: if we already converted any words, then they
* are going to be in an inconsistent state. */
E_ERROR_SYSTEM("iconv() failed");
ckd_free(outbuf);
hash_table_free(new_wid);
return -1;
}
/* Reset the internal state of conversion. */
iconv(ic, NULL, NULL, NULL, NULL);
/* Make everything bigger. */
maxlen *= 2;
out = outbuf = ckd_realloc(outbuf, maxlen);
/* Reset the input pointers. */
in = (ICONV_CONST char *)model->word_str[i];
inleft = strlen(in);
}
/* Now flush a shift-out sequence, if any. */
if ((result = iconv(ic, NULL, NULL, &out, &outleft)) == (size_t)-1) {
if (errno != E2BIG) {
/* FIXME: if we already converted any words, then they
* are going to be in an inconsistent state. */
E_ERROR_SYSTEM("iconv() failed (state reset sequence)");
ckd_free(outbuf);
hash_table_free(new_wid);
return -1;
}
/* Reset the internal state of conversion. */
iconv(ic, NULL, NULL, NULL, NULL);
/* Make everything bigger. */
maxlen *= 2;
outbuf = ckd_realloc(outbuf, maxlen);
/* Be very evil. */
goto start_conversion;
}
result = maxlen - outleft;
/* Okay, that was hard, now let's go shopping. */
if (writable) {
/* Grow or shrink the output string as necessary. */
model->word_str[i] = ckd_realloc(model->word_str[i], result + 1);
model->word_str[i][result] = '\0';
}
else {
/* It actually was not allocated previously, so do that now. */
model->word_str[i] = ckd_calloc(result + 1, 1);
}
/* Copy the new thing in. */
memcpy(model->word_str[i], outbuf, result);
/* Now update the hash table. We might have terrible
* collisions if a non-reversible conversion was requested.,
* so warn about them. */
if (hash_table_enter_int32(new_wid, model->word_str[i], i) != i) {
E_WARN("Duplicate word in dictionary after conversion: %s\n",
model->word_str[i]);
}
}
ckd_free(outbuf);
iconv_close(ic);
/* Swap out the hash table. */
hash_table_free(model->wid);
model->wid = new_wid;
return 0;
}
