/* Process utterances in the control file (-ctlfn argument) */
static void process_ctlfile ( void )
{
FILE *ctlfp, *sentfp, *mllrctlfp;
char *ctlfile, *cepdir, *cepext, *sentfile, *outsentfile, *mllrctlfile;
char line[1024], cepfile[1024], ctlspec[1024];
/* CHANGE BY BHIKSHA: ADDED veclen AS A VARIABLE, 6 JAN 98 */
int32 ctloffset, ctlcount, veclen, sf, ef, nfr;
/* END OF CHANGES BY BHIKSHA */
char mllrfile[4096], prevmllr[4096], sent[16384];
char uttid[1024];
int32 i, k;
float32 **mfc;
ctlfile = (char *) cmd_ln_access("-ctlfn");
if ((ctlfp = fopen (ctlfile, "r")) == NULL)
E_FATAL("fopen(%s,r) failed\n", ctlfile);
if ((mllrctlfile = (char *) cmd_ln_access("-mllrctlfn")) != NULL) {
if ((mllrctlfp = fopen (mllrctlfile, "r")) == NULL)
E_FATAL("fopen(%s,r) failed\n", mllrctlfile);
} else
mllrctlfp = NULL;
prevmllr[0] = '\0';
sentfile = (char *) cmd_ln_access("-insentfn");
if ((sentfp = fopen (sentfile, "r")) == NULL)
E_FATAL("fopen(%s,r) failed\n", sentfile);
if ((outsentfile = (char *) cmd_ln_access("-outsentfn")) != NULL) {
if ((outsentfp = fopen (outsentfile, "w")) == NULL)
E_FATAL("fopen(%s,r) failed\n", outsentfile);
}
E_INFO("Processing ctl file %s\n", ctlfile);
cepdir = (char *) cmd_ln_access("-cepdir");
cepext = (char *) cmd_ln_access("-cepext");
assert ((cepdir != NULL) && (cepext != NULL));
/* BHIKSHA: ADDING VECLEN TO ALLOW VECTORS OF DIFFERENT SIZES */
veclen = *((int32 *) cmd_ln_access("-ceplen"));
/* END CHANGES, 6 JAN 1998, BHIKSHA */
ctloffset = *((int32 *) cmd_ln_access("-ctloffset"));
if (! cmd_ln_access("-ctlcount"))
ctlcount = 0x7fffffff; /* All entries processed if no count specified */
else
ctlcount = *((int32 *) cmd_ln_access("-ctlcount"));
if (ctlcount == 0) {
E_INFO("-ctlcount argument = 0!!\n");
fclose (ctlfp);
fclose (sentfp);
if (outsentfp)
fclose (outsentfp);
return;
}
/* Skipping initial offset */
if (ctloffset > 0)
E_INFO("Skipping %d utterances in the beginning of control file\n",
ctloffset);
while ((ctloffset > 0) && (fgets(line, sizeof(line), ctlfp) != NULL)) {
if (sscanf (line, "%s", ctlspec) > 0) {
if (fgets (sent, sizeof(sent), sentfp) == NULL) {
E_ERROR("EOF(%s)\n", sentfile);
ctlcount = 0;
break;
}
if (mllrctlfp) {
if (fscanf (mllrctlfp, "%s", mllrfile) != 1)
E_FATAL ("Unexpected EOF(%s)\n", mllrctlfile);
}
--ctloffset;
}
}
/* Process the specified number of utterance or until end of control file */
while ((ctlcount > 0) && (fgets(line, sizeof(line), ctlfp) != NULL)) {
printf ("\n");
E_INFO("Utterance: %s", line);
sf = 0;
ef = (int32)0x7ffffff0;
if ((k = sscanf (line, "%s %d %d %s", ctlspec, &sf, &ef, uttid)) <= 0)
continue; /* Empty line */
if ((k == 2) || ( (k >= 3) && ((sf >= ef) || (sf < 0))) )
E_FATAL("Bad ctlfile line: %s\n", line);
if (k < 4) {
/* Create utt-id from mfc-filename (and sf/ef if specified) */
for (i = strlen(ctlspec)-1; (i >= 0) && (ctlspec[i] != '/'); --i);
if (k == 3)
sprintf (uttid, "%s_%d_%d", ctlspec+i+1, sf, ef);
else
strcpy (uttid, ctlspec+i+1);
}
if (mllrctlfp) {
if (fscanf (mllrctlfp, "%s", mllrfile) != 1)
E_FATAL ("Unexpected EOF(%s)\n", mllrctlfile);
if (strcmp (prevmllr, mllrfile) != 0) {
float32 ***A, **B;
int32 gid, sid;
uint8 *mgau_xform;
gauden_mean_reload (g, (char *) cmd_ln_access("-meanfn"));
if (mllr_read_regmat (mllrfile, &A, &B, featlen, n_feat) < 0)
E_FATAL("mllr_read_regmat failed\n");
mgau_xform = (uint8 *) ckd_calloc (g->n_mgau, sizeof(uint8));
/* Transform each non-CI mixture Gaussian */
for (sid = 0; sid < sen->n_sen; sid++) {
if (mdef->cd2cisen[sid] != sid) { /* Otherwise it's a CI senone */
gid = sen->mgau[sid];
if (! mgau_xform[gid]) {
mllr_norm_mgau (g->mean[gid], g->n_density, A, B,
featlen, n_feat);
mgau_xform[gid] = 1;
}
}
}
ckd_free (mgau_xform);
mllr_free_regmat (A, B, featlen, n_feat);
strcpy (prevmllr, mllrfile);
}
}
if (ctlspec[0] != '/')
sprintf (cepfile, "%s/%s.%s", cepdir, ctlspec, cepext);
else
sprintf (cepfile, "%s.%s", ctlspec, cepext);
/* Read utterance transcript */
if (fgets (sent, sizeof(sent), sentfp) == NULL) {
E_ERROR("EOF(%s)\n", sentfile);
break;
}
/* Strip utterance id from the end of the transcript */
for (k = strlen(sent) - 1;
(k > 0) && ((sent[k] == '\n') || (sent[k] == '\t') || (sent[k] == ' '));
--k);
if ((k > 0) && (sent[k] == ')')) {
for (--k; (k >= 0) && (sent[k] != '('); --k);
if ((k >= 0) && (sent[k] == '(')) {
sent[k] = '\0';
/* Check that uttid in transcript and control file match */
for (i = ++k;
sent[i] && (sent[i] != ')') &&
(sent[i] != '\n') && (sent[i] != '\t') && (sent[i] != ' ');
i++);
sent[i] = '\0';
if (id_cmp (sent+k, uttid) != 0)
E_ERROR("Uttid mismatch: ctlfile = \"%s\"; transcript = \"%s\"\n",
uttid, sent+k);
}
}
/* Read and process mfc file */
/* CHANGE BY BHIKSHA; PASSING VECLEN TO s2mfc_read(), 6 JAN 98 */
/* Read mfc file */
/* HACK HACKA HACK BHIKSHA */
{
int32 asf, aef;
asf = sf; sf = asf - 4;
aef = ef; ef = aef + 4;
if (sf < 0 ) {
E_ERROR("Utterance %s begin %d < 4; ignored\n", uttid, asf);
return;
}
if ((nfr = s2mfc_read (cepfile, sf, ef, &mfc, veclen)) <= 0)
E_ERROR("Utt %s: MFC file read (%s) failed\n", uttid, cepfile);
/* END CHANGES BY BHIKSHA */
else {
E_INFO ("%d mfc frames\n", nfr-8); /* -8 HACK HACKA HACK */
/* Align utterance */
align_utt (sent, mfc+4, nfr-8, ctlspec, uttid); /* +4 HACKA HACK */
}
}
/* END HACK HACKA HACK */
--ctlcount;
}
printf ("\n");
while (fgets(line, sizeof(line), ctlfp) != NULL) {
if (sscanf (line, "%s", ctlspec) > 0) {
E_INFO("Skipping rest of control file beginning with:\n\t%s", line);
break;
}
}
fclose (ctlfp);
fclose (sentfp);
if (outsentfp)
fclose (outsentfp);
if (mllrctlfp)
fclose (mllrctlfp);
}
void
adapt_set_mllr(adapt_am_t * ad, mgau_model_t * g, const char *mllrfile,
const char *cb2mllrname, mdef_t * mdef, cmd_ln_t *config)
{
int32 *cb2mllr;
float32 varfloor;
varfloor = cmd_ln_float32_r(config, "-varfloor");
/* Reread the gaussian mean from the file again */
E_INFO("Reloading mean\n");
mgau_mean_reload(g, cmd_ln_str_r(config, "-mean"));
/* Reread the gaussian variance from the file again */
E_INFO("Reloading variance\n");
mgau_var_reload(g, cmd_ln_str_r(config, "-var"));
mgau_uninit_compact(g); /* Delete uninitialized components */
if (g->mgau[0].var && varfloor > 0.0)
mgau_var_floor(g, varfloor); /* Variance floor after above compaction */
#if MLLR_DEBUG
/*This generates huge amount of information */
/* mgau_dump(g,1); */
#endif
/* Read in the mllr matrix */
mllr_read_regmat(mllrfile,
&(ad->regA),
&(ad->regB),
&(ad->regH),
&(ad->mllr_nclass), mgau_veclen(g));
if (cb2mllrname && strcmp(cb2mllrname, ".1cls.") != 0) {
uint32 ncb, nmllr;
cb2mllr_read(cb2mllrname, &cb2mllr, &ncb, &nmllr);
if (nmllr != ad->mllr_nclass)
E_FATAL
("Number of classes in cb2mllr does not match mllr (%d != %d)\n",
ncb, ad->mllr_nclass);
if (ncb != mdef->n_sen)
E_FATAL
("Number of senones in cb2mllr does not match mdef (%d != %d)\n",
ncb, mdef->n_sen);
}
else
cb2mllr = NULL;
/* Transform mean and variance vectors */
mllr_norm_mgau(g, ad->regA, ad->regB, ad->regH, ad->mllr_nclass, cb2mllr);
ckd_free(cb2mllr);
if (g->mgau[0].var && varfloor > 0.0)
mgau_var_floor(g, varfloor); /* Variance floor after above transform */
/* Re-precompute variance things */
mgau_precomp(g);
#if MLLR_DEBUG
/*#if 1 */
mllr_dump(ad->regA, ad->regB, ad->regH, mgau_veclen(g), g->mllr_class, cb2mllr);
/*This generates huge amount of information */
/*mgau_dump(kbcore_mgau(kb->kbcore),1); */
#endif
}
