/*
* Load and cross-check all models (acoustic/lexical/linguistic).
*/
static void models_init ( void )
{
dict_t *dict;
/* HMM model definition */
mdef = mdef_init ((char *) cmd_ln_access("-mdeffn"));
/* Dictionary */
dict = dict_init ((char *) cmd_ln_access("-dictfn"),
(char *) cmd_ln_access("-fdictfn"));
/* HACK!! Make sure SILENCE_WORD, START_WORD and FINISH_WORD are in dictionary */
silwid = dict_wordid (SILENCE_WORD);
startwid = dict_wordid (START_WORD);
finishwid = dict_wordid (FINISH_WORD);
if (NOT_WID(silwid) || NOT_WID(startwid) || NOT_WID(finishwid)) {
E_FATAL("%s, %s, or %s missing from dictionary\n",
SILENCE_WORD, START_WORD, FINISH_WORD);
}
if ((dict->filler_start > dict->filler_end) || (! dict_filler_word (silwid)))
E_FATAL("%s must occur (only) in filler dictionary\n", SILENCE_WORD);
/* No check that alternative pronunciations for filler words are in filler range!! */
/* LM */
lm_read ((char *) cmd_ln_access("-lmfn"), "");
/* Filler penalties */
fillpen_init ((char *) cmd_ln_access("-fillpenfn"),
dict->filler_start, dict->filler_end);
}
int32 align_init ( void )
{
int32 k;
s3wid_t w;
float64 *f64arg;
mdef = mdef_getmdef ();
tmat = tmat_gettmat ();
dict = dict_getdict ();
assert (mdef && tmat && dict);
startwid = dict_wordid (START_WORD);
finishwid = dict_wordid (FINISH_WORD);
silwid = dict_wordid (SILENCE_WORD);
if ((NOT_WID(startwid)) || (NOT_WID(finishwid)))
E_FATAL("%s or %s not in dictionary\n", START_WORD, FINISH_WORD);
if (NOT_WID(silwid))
E_ERROR("%s not in dictionary; no optional silence inserted between words\n",
SILENCE_WORD);
/* Create list of optional filler words to be inserted between transcript words */
fillwid = (s3wid_t *) ckd_calloc ((dict->filler_end - dict->filler_start + 3),
sizeof(s3wid_t));
k = 0;
if (IS_WID(silwid))
fillwid[k++] = silwid;
for (w = dict->filler_start; w <= dict->filler_end; w++) {
if ((dict_basewid (w) == w) &&
(w != silwid) && (w != startwid) && (w != finishwid))
fillwid[k++] = w;
}
fillwid[k] = BAD_WID;
f64arg = (float64 *) cmd_ln_access ("-beam");
beam = logs3 (*f64arg);
E_INFO ("logs3(beam)= %d\n", beam);
score_scale = (int32 *) ckd_calloc (S3_MAX_FRAMES, sizeof(int32));
hist_head = NULL;
align_stseg = NULL;
align_phseg = NULL;
align_wdseg = NULL;
ctr_nstate = counter_new ("NS");
return 0;
}
main (int32 argc, char *argv[])
{
mdef_t *m;
dict_t *d;
char wd[1024];
s3wid_t wid;
int32 p;
if (argc < 3)
E_FATAL("Usage: %s {mdeffile | NULL} dict [fillerdict]\n", argv[0]);
m = (strcmp (argv[1], "NULL") != 0) ? mdef_init (argv[1]) : NULL;
d = dict_init (m, argv[2], ((argc > 3) ? argv[3] : NULL), '_');
for (;;) {
printf ("word> ");
scanf ("%s", wd);
wid = dict_wordid (d, wd);
if (NOT_WID(wid))
E_ERROR("Unknown word\n");
else {
for (wid = dict_basewid(d, wid); IS_WID(wid); wid = d->word[wid].alt) {
printf ("%s\t", dict_wordstr(d, wid));
for (p = 0; p < d->word[wid].pronlen; p++)
printf (" %s", dict_ciphone_str (d, wid, p));
printf ("\n");
}
}
}
}
static void homfile_load (char *file)
{
FILE *fp;
char line[16380], w1[4096], w2[4096];
int32 k, n;
s3wid_t wid1, wid2;
s3cipid_t ci[1];
hom_t *h;
E_INFO("Reading homophones file %s\n", file);
if ((fp = fopen(file, "r")) == NULL)
E_FATAL("fopen(%s,r) failed\n", file);
ci[0] = (s3cipid_t) 0; /* Dummy */
n = 0;
while (fgets (line, sizeof(line), fp) != NULL) {
if ((k = sscanf (line, "%s %s", w1, w2)) == 2) {
wid1 = dict_wordid (dict, w1);
if (NOT_WID(wid1)) {
E_INFO("Adding %s to dictionary\n", w1);
wid1 = dict_add_word (dict, w1, ci, 1);
if (NOT_WID(wid1))
E_FATAL("dict_add_word(%s) failed\n", w1);
}
wid2 = dict_wordid (dict, w2);
if ((NOT_WID(wid2)) || (wid2 >= oovbegin))
E_FATAL("%s not in dictionary\n", w2);
h = (hom_t *) listelem_alloc (sizeof(hom_t));
h->w1 = wid1;
h->w2 = wid2;
h->next = homlist;
homlist = h;
n++;
} else
E_FATAL("Bad homophones line: %s\n", line);
}
E_INFO("%d homophone pairs read\n", n);
fclose (fp);
}
/*
* Create a degenerate DAG (linear sequence of nodes) for the given hyp line.
* The DAG contains a terminal sentinel silwid node.
*/
static dag_t *hypline2dag (char *ref_uttid, char *line)
{
char junk1[4096], junk2[4096], uttid[4096];
s3wid_t wid[MAX_UTT_LEN];
int32 i, n;
dag_t *dag;
dagnode_t *d;
if ((n = line2wid (dict, line, wid, MAX_UTT_LEN-1, 0, uttid)) < 0)
E_FATAL("Error in parsing hyp line: %s\n", line);
/* Verify uttid with ref_uttid */
if (strcmp (uttid, ref_uttid) != 0) {
strcpy (junk1, uttid);
ucase (junk1);
strcpy (junk2, ref_uttid);
ucase (junk2);
if (strcmp (junk1, junk2) != 0)
E_FATAL("Uttid mismatch: %s(ref), %s(hyp)\n", ref_uttid, uttid);
}
/* Build DAG from word sequence */
dag = ckd_calloc (1, sizeof(dag_t));
dag->node_sf = (dagnode_t **) ckd_calloc (S3_MAX_FRAMES, sizeof(dagnode_t *));
dag->nnode = 0;
dag->nfrm = 0;
dag->nlink = 0;
for (i = 0; i < n; i++) {
if ((NOT_WID(wid[i])) || (wid[i] >= oovbegin))
E_FATAL("%s: Unknown word in line: %s\n", uttid, line);
/* Create DAG node for word */
d = (dagnode_t *) listelem_alloc (sizeof(dagnode_t));
wid2dagnode (d, i, wid[i]);
dag->node_sf[i] = d;
if (i > 0) {
dag_link (dag->node_sf[i-1], d);
dag->nlink++;
}
dag->nnode++;
}
dag->nfrm = dag->nnode;
dag->entry.src = NULL;
dag->entry.dst = dag->node_sf[0];
dag->entry.next = NULL;
dag->exit.src = NULL;
dag->exit.dst = dag->node_sf[dag->nnode - 1];
dag->exit.next = NULL;
return dag;
}
int32 line2wid (dict_t *dict, char *line, s3wid_t *wid, int32 max_n_wid, int32 add_oov,
char *uttid)
{
char *lp, word[1024];
int32 n, k;
s3wid_t w;
s3cipid_t ci[1];
uttid[0] = '\0';
ci[0] = (s3cipid_t) 0;
lp = line;
n = 0;
while (sscanf (lp, "%s%n", word, &k) == 1) {
lp += k;
if (n >= max_n_wid)
return -n;
if (is_uttid (word, uttid))
break;
wid[n] = dict_wordid (dict, word); /* Up to caller to handle BAD_WIDs */
if (NOT_WID(wid[n])) {
/* OOV word */
if (add_oov) {
E_INFO("Adding %s to dictionary\n", word);
wid[n] = dict_add_word (dict, word, NULL, 0);
if (NOT_WID(wid[n]))
E_FATAL("dict_add_word(%s) failed for line: %s\n", word, line);
} else
E_FATAL("Unknown word (%s) in line: %s\n", word, line);
}
n++;
}
if (sscanf (lp, "%s", word) == 1) /* Check that line really ended */
E_WARN("Nonempty data ignored after uttid(%s) in line: %s\n", uttid, line);
return n;
}
/*
* Scan the dictionary for compound words. This function should be called just after
* loading the dictionary. For the moment, compound words in a compound word are
* assumed to be separated by the given sep character, (underscore in the CMU dict).
* Return value: #compound words found in dictionary.
*/
static int32 dict_build_comp (dict_t *d,
char sep) /* Separator character */
{
char wd[4096];
int32 w, cwid;
dictword_t *wordp;
int32 nc; /* # compound words in dictionary */
int32 i, j, l, n;
nc = 0;
for (w = 0; w < d->n_word; w++) {
wordp = d->word + dict_basewid(d, w);
strcpy (wd, wordp->word);
l = strlen(wd);
if ((wd[0] == sep) || (wd[l-1] == sep))
E_FATAL("Bad compound word %s: leading or trailing separator\n", wordp->word);
/* Count no. of components in this word */
n = 1;
for (i = 1; i < l-1; i++) /* 0 and l-1 already checked above */
if (wd[i] == sep)
n++;
if (n == 1)
continue; /* Not a compound word */
nc++;
if ((w == d->startwid) || (w == d->finishwid) || dict_filler_word (d, w))
E_FATAL("Compound special/filler word (%s) not allowed\n", wordp->word);
/* Allocate and fill in component word info */
wordp->n_comp = n;
wordp->comp = (s3wid_t *) ckd_calloc (n, sizeof(s3wid_t));
/* Parse word string into components */
n = 0;
for (i = 0; i < l; i++) {
for (j = i; (i < l) && (wd[i] != sep); i++);
if (j == i)
E_FATAL("Bad compound word %s: successive separators\n", wordp->word);
wd[i] = '\0';
cwid = dict_wordid (d, wd+j);
if (NOT_WID(cwid))
E_FATAL("Component word %s of %s not in dictionary\n", wd+j, wordp->word);
wordp->comp[n] = cwid;
n++;
}
}
if (nc > 0)
d->comp_head = dict_comp_head (d);
return nc;
}
static void dump_pnode_info (pnode_t *p)
{
if (NOT_WID(p->wid))
printf ("%s", (p->id == -1) ? "<head>" : "<tail>");
else
printf ("%s.%d.",
dict_wordstr(p->wid), p->pos, mdef_ciphone_str (mdef, p->ci));
printf ("%s", IS_CIPID(p->lc) ? mdef_ciphone_str (mdef, p->lc) : "-");
printf ("(%s)", IS_CIPID(p->ci) ? mdef_ciphone_str (mdef, p->ci) : "-");
printf ("%s", IS_CIPID(p->rc) ? mdef_ciphone_str (mdef, p->rc) : "-");
}
main (int32 argc, char *argv[])
{
dict_t **d;
int32 i, k, p, wid;
char line[16384], *wp[1024];
if (argc < 2) {
E_INFO("Usage: %s dictfile [dictfile ...] < vocabfile\n", argv[0]);
exit(0);
}
d = (dict_t **) ckd_calloc (argc-1, sizeof(dict_t *));
for (i = 1; i < argc; i++)
d[i-1] = dict_init (NULL, argv[i], NULL, 0);
while (fgets (line, sizeof(line), stdin) != NULL) {
if ((k = str2words (line, wp, 1024)) < 0)
E_FATAL("Line too long: %s\n", line);
if (k > 2)
E_FATAL("Vocab entry contains too many words\n");
if (k == 0)
continue;
if (k == 1)
wp[1] = wp[0];
/* Look up word in each dictionary until found */
k = 0;
for (i = 0; (i < argc-1) && (k == 0); i++) {
wid = dict_wordid (d[i], wp[1]);
if (NOT_WID(wid))
continue;
for (wid = dict_basewid(d[i], wid);
IS_WID(wid);
wid = dict_nextalt(d[i], wid)) {
k++;
if (k == 1)
printf ("%s\t", wp[0]);
else
printf ("%s(%d)\t", wp[0], k);
for (p = 0; p < dict_pronlen(d[i], wid); p++)
printf (" %s", dict_ciphone_str (d[i], wid, p));
printf ("\n");
}
}
if (k == 0)
E_ERROR("No pronunciation for: '%s'\n", wp[0]);
}
}
static int32 dict_read (FILE *fp, dict_t *d)
{
char line[16384], **wptr;
s3cipid_t p[4096];
int32 lineno, nwd;
s3wid_t w;
int32 i, maxwd;
maxwd = 4092;
wptr = (char **) ckd_calloc (maxwd, sizeof(char *));
lineno = 0;
while (fgets (line, sizeof(line), fp) != NULL) {
lineno++;
if (line[0] == '#') /* Comment line */
continue;
if ((nwd = str2words (line, wptr, maxwd)) < 0)
E_FATAL("str2words(%s) failed; Increase maxwd from %d\n", line, maxwd);
if (nwd == 0) /* Empty line */
continue;
/* wptr[0] is the word-string and wptr[1..nwd-1] the pronunciation sequence */
if (nwd == 1) {
E_ERROR("Line %d: No pronunciation for word %s; ignored\n", lineno, wptr[0]);
continue;
}
/* Convert pronunciation string to CI-phone-ids */
for (i = 1; i < nwd; i++) {
p[i-1] = dict_ciphone_id (d, wptr[i]);
if (NOT_CIPID(p[i-1])) {
E_ERROR("Line %d: Bad ciphone: %s; word %s ignored\n",
lineno, wptr[i], wptr[0]);
break;
}
}
if (i == nwd) { /* All CI-phones successfully converted to IDs */
w = dict_add_word (d, wptr[0], p, nwd-1);
if (NOT_WID(w))
E_ERROR("Line %d: dict_add_word (%s) failed (duplicate?); ignored\n",
lineno, wptr[0]);
}
}
ckd_free (wptr);
return 0;
}
static void process_reffile (char *reffile)
{
FILE *rfp, *afp;
char line[16384], uttid[4096];
int32 i, k, nref;
s3wid_t ref[MAX_UTT_LEN];
s3cipid_t ap[MAX_UTT_LEN];
int8 ap_err[MAX_UTT_LEN];
wseg_t *wseg;
if ((rfp = fopen(reffile, "r")) == NULL)
E_FATAL("fopen(%s,r) failed\n", reffile);
afp = stdin; /* DP file read in from stdin */
while (fgets(line, sizeof(line), rfp) != NULL) {
if ((nref = line2wid (dict, line, ref, MAX_UTT_LEN-1, 0, uttid)) < 0)
E_FATAL("Bad line in file %s: %s\n", reffile, line);
/* Check for unknown words; remove filler words; terminate with BAD_WID */
k = 0;
for (i = 0; i < nref; i++) {
if (NOT_WID(ref[i]))
E_FATAL("Unknown word at position %d in line: %s\n", i, line);
if (! dict_filler_word (dict, ref[i]))
ref[k++] = ref[i];
}
ref[k++] = BAD_WID;
nref = k;
/* Build wseg map for DP line */
if (fgets (line, sizeof(line), afp) == NULL)
E_FATAL("Unexpected EOF(DP-file)\n");
wseg = line2wseg (line, ref, ap, ap_err, MAX_UTT_LEN-1, uttid);
pronerr (uttid, ref, nref, wseg, ap, ap_err);
}
fclose (rfp);
}
dict_t *dict_init (mdef_t *mdef, char *dictfile, char *fillerfile, char comp_sep)
{
FILE *fp, *fp2;
int32 n ;
char line[1024];
dict_t *d;
if (! dictfile)
E_FATAL("No dictionary file\n");
/*
* First obtain #words in dictionary (for hash table allocation).
* Reason: The PC NT system doesn't like to grow memory gradually. Better to allocate
* all the required memory in one go.
*/
if ((fp = fopen(dictfile, "r")) == NULL)
E_FATAL_SYSTEM("fopen(%s,r) failed\n", dictfile);
n = 0;
while (fgets (line, sizeof(line), fp) != NULL) {
if (line[0] != '#')
n++;
}
rewind (fp);
if (fillerfile) {
if ((fp2 = fopen(fillerfile, "r")) == NULL)
E_FATAL_SYSTEM("fopen(%s,r) failed\n", fillerfile);
while (fgets (line, sizeof(line), fp2) != NULL) {
if (line[0] != '#')
n++;
}
rewind (fp2);
}
/*
* Allocate dict entries. HACK!! Allow some extra entries for words not in file.
* Also check for type size restrictions.
*/
d = (dict_t *) ckd_calloc (1, sizeof(dict_t));
d->max_words = (n+1024 < MAX_WID) ? n+1024 : MAX_WID;
if (n >= MAX_WID)
E_FATAL("#Words in dictionaries (%d) exceeds limit (%d)\n", n, MAX_WID);
d->word = (dictword_t *) ckd_calloc (d->max_words, sizeof(dictword_t));
d->n_word = 0;
d->mdef = mdef;
if (mdef) {
d->pht = NULL;
d->ciphone_str = NULL;
} else {
d->pht = hash_new (DEFAULT_NUM_PHONE, 1 /* No case */);
d->ciphone_str = (char **) ckd_calloc (DEFAULT_NUM_PHONE, sizeof(char *));
}
d->n_ciphone = 0;
/* Create new hash table for word strings; case-insensitive word strings */
d->ht = hash_new (d->max_words, 1 /* no-case */);
/* Initialize with no compound words */
d->comp_head = NULL;
/* Digest main dictionary file */
E_INFO("Reading main dictionary: %s\n", dictfile);
dict_read (fp, d);
fclose (fp);
E_INFO("%d words read\n", d->n_word);
/* Now the filler dictionary file, if it exists */
d->filler_start = d->n_word;
if (fillerfile) {
E_INFO("Reading filler dictionary: %s\n", fillerfile);
dict_read (fp2, d);
fclose (fp2);
E_INFO("%d words read\n", d->n_word - d->filler_start);
}
d->filler_end = d->n_word-1;
/* Initialize distinguished word-ids */
d->startwid = dict_wordid (d, START_WORD);
d->finishwid = dict_wordid (d, FINISH_WORD);
d->silwid = dict_wordid (d, SILENCE_WORD);
if (NOT_WID(d->startwid))
E_WARN("%s not in dictionary\n", START_WORD);
if (NOT_WID(d->finishwid))
E_WARN("%s not in dictionary\n", FINISH_WORD);
if (NOT_WID(d->silwid))
E_WARN("%s not in dictionary\n", SILENCE_WORD);
/* Identify compound words if indicated */
if (comp_sep) {
E_INFO("Building compound words (separator = '%c')\n", comp_sep);
n = dict_build_comp (d, comp_sep);
E_INFO("%d compound words\n", n);
}
return d;
}
/*
* Build a sentence HMM for the given transcription (wordstr). A two-level DAG is
* built: phone-level and state-level.
* - <s> and </s> always added at the beginning and end of sentence to form an
* augmented transcription.
* - Optional <sil> and noise words added between words in the augmented
* transcription.
* wordstr must contain only the transcript; no extraneous stuff such as utterance-id.
* Phone-level HMM structure has replicated nodes to allow for different left and right
* context CI phones; hence, each pnode corresponds to a unique triphone in the sentence
* HMM.
* Return 0 if successful, <0 if any error (eg, OOV word encountered).
*/
int32 align_build_sent_hmm (char *wordstr)
{
s3wid_t w, nextw;
int32 k, oov;
pnode_t *word_end, *node;
char *wd, delim, *wdcopy;
/* HACK HACKA HACK BHIKSHA */
int32 firsttime = 1;
/* END HACK HACKA HACK */
/* Initialize dummy head and tail entries of sent hmm */
phead.wid = BAD_WID;
phead.ci = BAD_CIPID;
phead.lc = BAD_CIPID; /* No predecessor */
phead.rc = BAD_CIPID; /* Any phone can follow head */
phead.pid = BAD_PID;
phead.succlist = NULL;
phead.predlist = NULL;
phead.next = NULL; /* Will ultimately be the head of list of all pnodes */
phead.id = -1; /* Hardwired */
phead.startstate = NULL;
ptail.wid = BAD_WID;
ptail.ci = BAD_CIPID;
ptail.lc = BAD_CIPID; /* Any phone can precede tail */
ptail.rc = BAD_CIPID; /* No successor */
ptail.pid = BAD_PID;
ptail.succlist = NULL;
ptail.predlist = NULL;
ptail.next = NULL;
ptail.id = -2; /* Hardwired */
ptail.startstate = NULL;
n_pnode = 0;
pnode_list = NULL;
oov = 0;
/* State-level DAG initialization should be here in case the build is aborted */
shead.pnode = &phead;
shead.succlist = NULL;
shead.predlist = NULL;
shead.sen = BAD_SENID;
shead.state = mdef->n_emit_state;
shead.hist = NULL;
stail.pnode = &ptail;
stail.succlist = NULL;
stail.predlist = NULL;
stail.sen = BAD_SENID;
stail.state = 0;
stail.hist = NULL;
/* Obtain the first transcript word */
k = nextword (wordstr, " \t\n", &wd, &delim);
if (k < 0)
nextw = finishwid;
else {
wordstr = wd + k;
wdcopy = ckd_salloc (wd);
*wordstr = delim;
nextw = dict_wordid (wdcopy);
if (IS_WID(nextw))
nextw = dict_basewid (nextw);
}
/* Create node(s) for <s> before any transcript word */
/* HACK HACKA HACK BHIKSHA
word_end = append_transcript_word (startwid, &phead, nextw, 0, 1);
END HACK HACKA HACK BHIKSHA */
/* Append each word in transcription to partial sent HMM created so far */
while (k >= 0) {
w = nextw;
if (NOT_WID(w)) {
E_ERROR("%s not in dictionary\n", wdcopy);
oov = 1;
/* Hack!! Temporarily set w to some dummy just to run through sentence */
w = finishwid;
}
ckd_free (wdcopy);
k = nextword (wordstr, " \t\n", &wd, &delim);
if (k < 0)
nextw = finishwid;
else {
wordstr = wd + k;
wdcopy = ckd_salloc (wd);
*wordstr = delim;
nextw = dict_wordid (wdcopy);
if (IS_WID(nextw))
nextw = dict_basewid (nextw);
}
/* HACK HACKA HACK BHIKSHA */
if (firsttime){
word_end = append_transcript_word (w, &phead, nextw, 0, 1);
firsttime = 0;
}
else
if (nextw == finishwid)
word_end = append_transcript_word (w, word_end, BAD_WID, 1, 0);
else
word_end = append_transcript_word (w, word_end, nextw, 1, 1);
/* END HACK HACKA HACK BHIKSHA */
}
if (oov)
return -1;
/* Append phone HMMs for </s> at the end; link to tail node */
/* HACK HACKA HACK BHIKSHA
word_end = append_transcript_word (finishwid, word_end, BAD_WID, 1, 0);
END HACK HACKA HACK BHIKSHA */
for (node = word_end; node; node = node->next)
link_pnodes (node, &ptail);
/* Build state-level DAG from the phone-level one */
build_state_dag ();
/* Dag must begin and end at shead and stail, respectively */
assert (shead.succlist);
assert (stail.predlist);
assert (! shead.predlist);
assert (! stail.succlist);
#if _DEBUG_ALIGN_
dump_sent_hmm (); /* For debugging */
#endif
k = n_pnode * mdef->n_emit_state;
if (k > active_list_size) { /* Need to grow active list arrays */
if (active_list_size > 0) {
ckd_free (cur_active);
ckd_free (next_active);
}
for (; active_list_size <= k; active_list_size += ACTIVE_LIST_SIZE_INCR);
cur_active = (snode_t **) ckd_calloc (active_list_size, sizeof(snode_t *));
next_active = (snode_t **) ckd_calloc (active_list_size, sizeof(snode_t *));
}
return 0;
}
/*
* Load a DAG from a file: each unique <word-id,start-frame> is a node, i.e. with
* a single start time but it can represent several end times. Links are created
* whenever nodes are adjacent in time.
* Return value: ptr to DAG structure if successful; NULL otherwise.
*/
dag_t *dag_load (char *file)
{
FILE *fp;
dag_t *dag;
int32 seqid, sf, fef, lef, ef;
char line[16384], wd[4096];
int32 i, j, k;
dagnode_t *d, *d2, **darray;
s3wid_t w;
int32 fudge, min_ef_range;
E_INFO("Reading DAG file: %s\n", file);
if ((fp = fopen (file, "r")) == NULL) {
E_ERROR("fopen(%s,r) failed\n", file);
return NULL;
}
dag = ckd_calloc (1, sizeof(dag_t));
dag->node_sf = (dagnode_t **) ckd_calloc (S3_MAX_FRAMES, sizeof(dagnode_t *));
dag->nnode = 0;
dag->nlink = 0;
dag->nfrm = 0;
/* Read Frames parameter */
if ((dag->nfrm = dag_param_read (fp, "Frames")) <= 0)
E_FATAL("%s: Frames parameter missing or invalid\n", file);
/* Read Nodes parameter */
if ((dag->nnode = dag_param_read (fp, "Nodes")) <= 0)
E_FATAL("%s: Nodes parameter missing or invalid\n", file);
/* Read nodes */
darray = (dagnode_t **) ckd_calloc (dag->nnode, sizeof(dagnode_t *));
for (i = 0; i < dag->nnode; i++) {
if (fgets (line, sizeof(line), fp) == NULL)
E_FATAL("%s: Premature EOF\n", file);
if ((k = sscanf (line, "%d %s %d %d %d", &seqid, wd, &sf, &fef, &lef)) != 5)
E_FATAL("%s: Bad line: %s\n", file, line);
if ((sf < 0) || (sf >= dag->nfrm) ||
(fef < 0) || ( fef >= dag->nfrm) ||
(lef < 0) || ( lef >= dag->nfrm))
E_FATAL("%s: Bad frame info: %s\n", file, line);
w = dict_wordid (dict, wd);
if (NOT_WID(w))
E_FATAL("%s: Unknown word: %s\n", file, line);
if (seqid != i)
E_FATAL("%s: Seqno error: %s\n", file, line);
d = (dagnode_t *) listelem_alloc (sizeof(dagnode_t));
darray[i] = d;
d->wid = w;
d->seqid = seqid;
d->reachable = 0;
d->sf = sf;
d->fef = fef;
d->lef = lef;
d->succlist = NULL;
d->predlist = NULL;
d->next = dag->node_sf[sf];
dag->node_sf[sf] = d;
}
/* Read initial node ID */
if (((k = dag_param_read (fp, "Initial")) < 0) || (k >= dag->nnode))
E_FATAL("%s: Initial node parameter missing or invalid\n", file);
dag->entry.src = NULL;
dag->entry.dst = darray[k];
dag->entry.next = NULL;
/* Read final node ID */
if (((k = dag_param_read (fp, "Final")) < 0) || (k >= dag->nnode))
E_FATAL("%s: Final node parameter missing or invalid\n", file);
dag->exit.src = NULL;
dag->exit.dst = darray[k];
dag->exit.next = NULL;
ckd_free (darray); /* That's all I need darray for??? */
/* Read bestsegscore entries; just to make sure all nodes have been read */
if ((k = dag_param_read (fp, "BestSegAscr")) < 0)
E_FATAL("%s: BestSegAscr parameter missing\n", file);
fclose (fp);
/*
* Build edges based on time-adjacency.
* min_ef_range = min. endframes that a node must persist for it to be not ignored.
* fudge = #frames to be fudged around word begin times
*/
min_ef_range = *((int32 *) cmd_ln_access ("-min_endfr"));
fudge = *((int32 *) cmd_ln_access ("-dagfudge"));
if (min_ef_range <= 0)
E_FATAL("Bad min_endfr argument: %d\n", min_ef_range);
if ((fudge < 0) || (fudge > 2))
E_FATAL("Bad dagfudge argument: %d\n", fudge);
dag->nlink = 0;
for (sf = 0; sf < dag->nfrm; sf++) {
for (d = dag->node_sf[sf]; d; d = d->next) {
if ((d->lef - d->fef < min_ef_range - 1) && (d != dag->entry.dst))
continue;
if (d->wid == finishwid)
continue;
for (ef = d->fef - fudge + 1; ef <= d->lef + 1; ef++) {
for (d2 = dag->node_sf[ef]; d2; d2 = d2->next) {
if ((d2->lef - d2->fef < min_ef_range - 1) && (d2 != dag->exit.dst))
continue;
dag_link (d, d2);
dag->nlink++;
}
}
}
}
return dag;
}
/*
* Load and cross-check all models (acoustic/lexical/linguistic).
*/
static void models_init ( void )
{
float32 varfloor, mixwfloor, tpfloor;
int32 i, s;
s3cipid_t ci;
s3wid_t w;
char *arg;
dict_t *dict;
/* HMM model definition */
mdef = mdef_init ((char *) cmd_ln_access("-mdeffn"));
/* Dictionary */
dict = dict_init ((char *) cmd_ln_access("-dictfn"),
(char *) cmd_ln_access("-fdictfn"));
/* HACK!! Make sure SILENCE_WORD, START_WORD and FINISH_WORD are in dictionary */
silwid = dict_wordid (SILENCE_WORD);
startwid = dict_wordid (START_WORD);
finishwid = dict_wordid (FINISH_WORD);
if (NOT_WID(silwid) || NOT_WID(startwid) || NOT_WID(finishwid)) {
E_FATAL("%s, %s, or %s missing from dictionary\n",
SILENCE_WORD, START_WORD, FINISH_WORD);
}
if ((dict->filler_start > dict->filler_end) || (! dict_filler_word (silwid)))
E_FATAL("%s must occur (only) in filler dictionary\n", SILENCE_WORD);
/* No check that alternative pronunciations for filler words are in filler range!! */
/* Codebooks */
varfloor = *((float32 *) cmd_ln_access("-varfloor"));
g = gauden_init ((char *) cmd_ln_access("-meanfn"),
(char *) cmd_ln_access("-varfn"),
varfloor);
/* Verify codebook feature dimensions against libfeat */
n_feat = feat_featsize (&featlen);
if (n_feat != g->n_feat)
E_FATAL("#feature mismatch: s2= %d, mean/var= %d\n", n_feat, g->n_feat);
for (i = 0; i < n_feat; i++)
if (featlen[i] != g->featlen[i])
E_FATAL("featlen[%d] mismatch: s2= %d, mean/var= %d\n", i,
featlen[i], g->featlen[i]);
/* Senone mixture weights */
mixwfloor = *((float32 *) cmd_ln_access("-mwfloor"));
sen = senone_init ((char *) cmd_ln_access("-mixwfn"),
(char *) cmd_ln_access("-senmgaufn"),
mixwfloor);
/* Verify senone parameters against gauden parameters */
if (sen->n_feat != g->n_feat)
E_FATAL("#Feature mismatch: gauden= %d, senone= %d\n", g->n_feat, sen->n_feat);
if (sen->n_cw != g->n_density)
E_FATAL("#Densities mismatch: gauden= %d, senone= %d\n", g->n_density, sen->n_cw);
if (sen->n_gauden > g->n_mgau)
E_FATAL("Senones need more codebooks (%d) than present (%d)\n",
sen->n_gauden, g->n_mgau);
if (sen->n_gauden < g->n_mgau)
E_ERROR("Senones use fewer codebooks (%d) than present (%d)\n",
sen->n_gauden, g->n_mgau);
/* Verify senone parameters against model definition parameters */
if (mdef->n_sen != sen->n_sen)
E_FATAL("Model definition has %d senones; but #senone= %d\n",
mdef->n_sen, sen->n_sen);
/* CD/CI senone interpolation weights file, if present */
if ((arg = (char *) cmd_ln_access ("-lambdafn")) != NULL) {
interp = interp_init (arg);
/* Verify interpolation weights size with senones */
if (interp->n_sen != sen->n_sen)
E_FATAL("Interpolation file has %d weights; but #senone= %d\n",
interp->n_sen, sen->n_sen);
} else
interp = NULL;
/* Transition matrices */
tpfloor = *((float32 *) cmd_ln_access("-tpfloor"));
tmat = tmat_init ((char *) cmd_ln_access("-tmatfn"), tpfloor);
/* Verify transition matrices parameters against model definition parameters */
if (mdef->n_tmat != tmat->n_tmat)
E_FATAL("Model definition has %d tmat; but #tmat= %d\n",
mdef->n_tmat, tmat->n_tmat);
if (mdef->n_emit_state != tmat->n_state-1)
E_FATAL("#Emitting states in model definition = %d, #states in tmat = %d\n",
mdef->n_emit_state, tmat->n_state);
arg = (char *) cmd_ln_access ("-agc");
if ((strcmp (arg, "max") != 0) && (strcmp (arg, "none") != 0))
E_FATAL("Unknown -agc argument: %s\n", arg);
arg = (char *) cmd_ln_access ("-cmn");
if ((strcmp (arg, "current") != 0) && (strcmp (arg, "none") != 0))
E_FATAL("Unknown -cmn argument: %s\n", arg);
}
void lmcontext_load (corpus_t *corp, char *uttid, s3wid_t *pred, s3wid_t *succ)
{
char *str, wd[4096], *strp;
s3wid_t w[3];
int32 i, n;
dict_t *dict;
s3lmwid_t lwid;
if ((str = corpus_lookup (corp, uttid)) == NULL)
E_FATAL("Couldn't find LM context for %s\n", uttid);
dict = dict_getdict ();
strp = str;
for (i = 0; i < 4; i++) {
if (sscanf (strp, "%s%n", wd, &n) != 1) {
if (i < 3)
E_FATAL("Bad LM context spec for %s: %s\n", uttid, str);
else
break;
}
strp += n;
if (strcmp (wd, "-") == 0)
w[i] = BAD_WID;
else {
w[i] = dict_wordid (wd);
if (NOT_WID(w[i]))
E_FATAL("LM context word (%s) for %s not in dictionary\n", wd, uttid);
w[i] = dict_basewid(w[i]);
switch (i) {
case 0:
if ((n = dict->word[w[0]].n_comp) > 0)
w[0] = dict->word[w[0]].comp[n-1].wid;
break;
case 1:
if ((n = dict->word[w[1]].n_comp) > 0) {
w[0] = dict->word[w[1]].comp[n-2].wid;
w[1] = dict->word[w[1]].comp[n-1].wid;
}
break;
case 2:
if (w[2] != dict_wordid(FINISH_WORD))
E_FATAL("Illegal successor LM context for %s: %s\n", uttid, str);
break;
default:
assert (0); /* Should never get here */
break;
}
}
}
if (IS_WID(w[0]) && NOT_WID(w[1]))
E_FATAL("Bad LM context spec for %s: %s\n", uttid, str);
for (i = 0; i < 3; i++) {
if (IS_WID(w[i])) {
lwid = lm_lmwid (w[i]);
if (NOT_LMWID(lwid))
E_FATAL("LM context word (%s) for %s not in LM\n", wd, uttid);
}
}
pred[0] = w[0];
pred[1] = w[1];
*succ = w[2];
}
static wseg_t *line2wseg (char *line, s3wid_t *ref,
s3cipid_t *ap, int8 *ap_err, int32 aplen, char *id)
{
char word[1024], uttid[1024], *lp;
int32 i, k, n_hypci, n_refwd, n_refci, pronlen;
s3cipid_t ci;
typedef enum {CORR=0, REFERR=1, HYPERR=2} state_t;
state_t state;
static wseg_t *wseg = NULL;
if (! wseg)
wseg = (wseg_t *) ckd_calloc (MAX_UTT_LEN, sizeof(wseg_t));
lp = line;
n_hypci = n_refci = pronlen = 0;
n_refwd = -1;
uttid[0] = '\0';
state = CORR;
while (sscanf (lp, "%s%n", word, &k) == 1) {
lp += k;
if (is_uttid (word, uttid))
break;
if (strcmp (word, "[[") == 0) {
if (state != CORR)
E_FATAL("%s: Illegal [[\n", id);
state = REFERR;
if (n_refci < pronlen)
wseg[n_refwd].err = 1;
} else if (strcmp (word, "]]") == 0) {
if (state != HYPERR)
E_FATAL("%s: Illegal ]]\n", id);
state = CORR;
} else if (strcmp (word, "=>") == 0) {
if (state != REFERR)
E_FATAL("%s: Illegal =>\n", id);
state = HYPERR;
} else {
ci = mdef_ciphone_id (mdef, word);
if (NOT_CIPID(ci))
E_FATAL("%s: Unknown CIphone %s\n", id, word);
if (state != HYPERR) { /* Check if matches next pron in ref word */
if (n_refci >= pronlen) {
assert (n_refci == pronlen);
n_refwd++;
pronlen = dict->word[ref[n_refwd]].pronlen;
assert (pronlen > 0);
wseg[n_refwd].s = (state == CORR) ? n_hypci : -1;
wseg[n_refwd].e = -1;
wseg[n_refwd].err = 0;
n_refci = 0;
}
if (NOT_WID(ref[n_refwd]))
E_FATAL("%s: Premature end of ref wid\n", id);
if (dict->word[ref[n_refwd]].ciphone[n_refci] != ci)
E_FATAL("%s: CIphone mismatch at word %d, ciphone %d\n",
id, n_refwd, n_refci);
n_refci++;
if ((n_refci == pronlen) && (state == CORR))
wseg[n_refwd].e = n_hypci;
if (state != CORR)
wseg[n_refwd].err = 1;
}
if (state != REFERR) {
if (n_hypci >= aplen)
E_FATAL("%s: Too many CIphones: >%d\n", id, aplen);
ap[n_hypci] = ci;
ap_err[n_hypci] = (state == CORR) ? 0 : 1;
n_hypci++;
}
}
}
assert (n_refci == pronlen);
n_refwd++;
assert (NOT_WID(ref[n_refwd]));
wseg[n_refwd].s = wseg[n_refwd].e = n_hypci;
wseg[n_refwd].err = 0;
ap[n_hypci] = BAD_CIPID;
ap_err[n_hypci] = 1;
if (strcmp (uttid, id) != 0)
E_FATAL("Uttid mismatch: %s expected, %s found\n", id, uttid);
#if 0
for (i = 0; IS_WID(ref[i]); i++) {
printf ("%s: %4d %4d %d %s\n", id, wseg[i].s, wseg[i].e, wseg[i].err,
dict_wordstr (dict, ref[i]));
}
#endif
return wseg;
}
