/*
* Append a new word to partially build phone-level sentence HMM. (Handle alternative
* pronunciations.) Link new word to end phones of previous words.
* Append optional filler words before w, if indicated.
* Also Link prev_end into the global node list.
* Return value: list of end phone nodes for w. (NOTE: these are not yet linked into
* the global node list.)
*/
static pnode_t *append_transcript_word
(s3wid_t w, /* Transcript word to be appended */
pnode_t *prev_end, /* Previous end points to be attached to w */
s3wid_t nextw, /* Next word to follow w (ignoring optional fillers) */
int32 prefix_filler, /* Whether optional filler words to precede w */
int32 append_filler) /* Whether optional filler words to follow w */
{
int32 i, p;
pnode_t *new_end, *tmp_end, *node;
s3cipid_t pred_ci[256], succ_ci[256];
s3wid_t fw;
if (mdef->n_ciphone >= 256)
E_FATAL("Increase pred_ci, succ_ci array sizes to > #CIphones (%d)\n",
mdef->n_ciphone);
assert (prev_end != NULL);
/* Add optional silence/filler words before w, if indicated */
if (prefix_filler) {
build_pred_ci (prev_end, pred_ci); /* Predecessor CI list for fillers */
build_succ_ci (w, 0, succ_ci); /* Successor CI list for fillers */
new_end = NULL;
for (i = 0; IS_WID(fillwid[i]); i++) {
for (fw = fillwid[i]; IS_WID(fw); fw = dict->word[fw].alt) {
tmp_end = append_word (fw, prev_end, pred_ci, succ_ci);
for (node = tmp_end; node->next; node = node->next);
node->next = new_end;
new_end = tmp_end;
}
}
/* Augment prev_end with new_end for filler words added above */
for (node = prev_end; node->next; node = node->next);
node->next = new_end;
}
/* Add w */
build_pred_ci (prev_end, pred_ci); /* Predecessor CI list for w */
build_succ_ci (nextw, append_filler, succ_ci); /* Successor CI list for w */
new_end = NULL;
for (; IS_WID(w); w = dict->word[w].alt) {
tmp_end = append_word (w, prev_end, pred_ci, succ_ci);
for (node = tmp_end; node->next; node = node->next);
node->next = new_end;
new_end = tmp_end;
}
return (new_end);
}
static int32 refline2wds (char *line, dagnode_t *ref, int32 *noov, char *uttid)
{
int32 i, n, k;
s3wid_t w, wid[MAX_UTT_LEN];
n = 0;
uttid[0] = '\0';
*noov = 0;
if ((n = line2wid (dict, line, wid, MAX_UTT_LEN-1, 1, uttid)) < 0)
E_FATAL("Error in parsing ref line: %s\n", line);
wid[n++] = silwid;
for (i = 0; i < n; i++) {
if (dict_filler_word (dict, wid[i]) && (i < n-1))
E_FATAL("Filler word (%s) in ref: %s\n", dict_wordstr(dict, wid[i]), line);
if (wid[i] >= oovbegin) {
/* Perhaps one of a homophone pair */
w = hom_lookup (wid[i]);
if (IS_WID(w))
wid[i] = w;
if (wid[i] >= oovbegin)
(*noov)++;
}
wid2dagnode (ref+i, i, wid[i]);
}
return n;
}
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");
}
}
}
}
s3wid_t _dict_nextalt (dict_t *d, s3wid_t wid)
{
assert (d);
assert (IS_WID(wid) && (wid < d->n_word));
return (d->word[wid].alt);
}
char *_dict_wordstr (dict_t *d, s3wid_t wid)
{
assert (d);
assert (IS_WID(wid) && (wid < d->n_word));
return (d->word[wid].word);
}
static void pronerr_output (char *id, s3wid_t *ref, int32 nref,
wseg_t *wseg, s3cipid_t *ap, int8 *ap_err,
int32 ws, int32 we, int32 ps, int32 pe)
{
int32 j;
s3wid_t rcwid, lcwid;
char str[4096];
/* Word sequence for region in error */
sprintf (str, "%s", dict_wordstr (dict, dict_basewid(dict, ref[ws])));
for (j = ws+1; j <= we; j++) {
strcat (str, " ");
strcat (str, dict_wordstr (dict, dict_basewid(dict, ref[j])));
}
printf ("%-22s\t=>\t", str);
/* Print left context phone */
/*lcwid = ((wseg[ws].s < 0) && (ws > 0) && IS_WID(ref[ws-1])) ? ref[ws-1] : BAD_WID;*/
lcwid = (ws > 0) ? ref[ws-1] : BAD_WID;
if (IS_WID(lcwid)) {
j = dict->word[lcwid].pronlen - 1;
sprintf (str, "(%s)", mdef_ciphone_str (mdef, dict->word[lcwid].ciphone[j]));
} else
strcpy (str, "()");
printf ("%-5s", str);
/* Phone sequence for region in error */
for (j = ps; j <= pe; j++) {
strcpy (str, mdef_ciphone_str (mdef, ap[j]));
if (ap_err[j])
ucase (str);
else
lcase (str);
printf (" %s", str);
}
/* Right context if ending in error */
/* rcwid = ((wseg[we].e < 0) && IS_WID(ref[we+1])) ? ref[we+1] : BAD_WID; */
rcwid = ref[we+1];
if (IS_WID(rcwid))
printf ("\t(%s)", mdef_ciphone_str (mdef, dict->word[rcwid].ciphone[0]));
else
printf ("\t()");
printf (" ( %s )\n", id);
}
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[])
{
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 void build_succ_ci (s3wid_t w, int32 append_filler, s3cipid_t *succ_ci)
{
int32 i, p;
pnode_t *node;
for (p = 0; p < mdef->n_ciphone; p++)
succ_ci[p] = 0;
for (; IS_WID(w); w = dict->word[w].alt)
succ_ci[dict->word[w].ciphone[0]] = 1;
if (append_filler) {
for (i = 0; IS_WID(fillwid[i]); i++)
for (w = fillwid[i]; IS_WID(w); w = dict->word[w].alt)
succ_ci[dict->word[w].ciphone[0]] = 1;
}
i = 0;
for (p = 0; p < mdef->n_ciphone; p++) {
if (succ_ci[p])
succ_ci[i++] = p;
}
succ_ci[i] = BAD_CIPID;
}
static void dump_pnode_succ (pnode_t *p)
{
plink_t *l;
printf (" %5d", p->id);
if (IS_WID(p->wid))
printf (" %20s %02d %6d %4s",
dict_wordstr(p->wid), p->pos, p->pid, mdef_ciphone_str (mdef, p->ci));
else
printf (" %20s %02d %6d %4s",
"<phead>", 0, BAD_PID, "");
printf (" %4s %4s",
IS_CIPID(p->lc) ? mdef_ciphone_str (mdef, p->lc) : "-",
IS_CIPID(p->rc) ? mdef_ciphone_str (mdef, p->rc) : "-");
printf ("\t");
for (l = p->succlist; l; l = l->next)
printf (" %5d", l->node->id);
printf ("\n");
}
s3wid_t dict_wids2compwid (dict_t *d, s3wid_t *wid, int32 len)
{
s3wid_t w;
int32 i;
if (! d->comp_head)
return BAD_WID;
assert (len > 1);
for (w = d->comp_head[wid[0]]; IS_WID(w); w = d->comp_head[w]) {
/* w is a compound word beginning with wid[0]; check if rest matches */
assert (d->word[w].n_comp > 1);
assert (d->word[w].comp[0] == wid[0]);
if (d->word[w].n_comp == len) {
for (i = 0; (i < len) && (d->word[w].comp[i] == wid[i]); i++);
if (i == len)
return (dict_basewid(d, w));
}
}
return BAD_WID;
}
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;
}
/*
* 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;
}