/* 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);
}
static void decode_utt (void *data, char *uttfile, int32 sf, int32 ef, char *uttid)
{
kb_t *kb;
acoustic_t *am;
int32 featwin, nfr, min_utt_frames, n_vithist;
char cepfile[4096], latfile[4096];
vithist_t *finalhist;
int32 i, f;
glist_t hyplist;
FILE *latfp;
printf ("\n");
fflush (stdout);
E_INFO("Utterance %s\n", uttid);
kb = (kb_t *)data;
am = kb->am;
featwin = feat_window_size(am->fcb);
/* Build complete cepfile name and read cepstrum data; check for min length */
ctl_infile (cepfile, cmd_ln_str("-cepdir"), cmd_ln_str("-cepext"), uttfile);
if ((nfr = s2mfc_read (cepfile, sf, ef, featwin, am->mfc, S3_MAX_FRAMES)) < 0) {
E_ERROR("%s: MFC read failed\n", uttid);
return;
}
E_INFO("%s: %d frames\n", uttid, nfr-(featwin<<1));
ptmr_reset (kb->tm);
ptmr_reset (kb->tm_search);
ptmr_start (kb->tm);
min_utt_frames = (featwin<<1) + 1;
if (nfr < min_utt_frames) {
E_ERROR("%s: Utterance shorter than %d frames; ignored\n",
uttid, min_utt_frames, nfr);
return;
}
/* CMN/AGC */
if (strcmp (cmd_ln_str("-cmn"), "current") == 0)
cmn (am->mfc, nfr, feat_cepsize(am->fcb));
if (strcmp (cmd_ln_str("-agc"), "max") == 0)
agc_max (am->mfc, nfr);
/* Process utterance */
lextree_vit_start (kb, uttid);
for (i = featwin, f = 0; i < nfr-featwin; i++, f++) {
am->senscale[f] = acoustic_eval (am, i);
ptmr_start (kb->tm_search);
lextree_vit_frame (kb, f, uttid);
printf (" %d,%d,%d", f, glist_count (kb->vithist[f]), glist_count (kb->lextree_active));
fflush (stdout);
ptmr_stop (kb->tm_search);
}
printf ("\n");
finalhist = lextree_vit_end (kb, f, uttid);
hyplist = vithist_backtrace (finalhist, kb->am->senscale);
hyp_log (stdout, hyplist, _dict_wordstr, (void *)kb->dict);
hyp_myfree (hyplist);
printf ("\n");
/* Log the entire Viterbi word lattice */
sprintf (latfile, "%s.lat", uttid);
if ((latfp = fopen(latfile, "w")) == NULL) {
E_ERROR("fopen(%s,w) failed; using stdout\n", latfile);
latfp = stdout;
}
n_vithist = vithist_log (latfp, kb->vithist, f, _dict_wordstr, (void *)kb->dict);
if (latfp != stdout)
fclose (latfp);
else {
printf ("\n");
fflush (stdout);
}
ptmr_stop (kb->tm);
if (f > 0) {
printf("TMR(%s): %5d frames; %.1fs CPU, %.2f xRT; %.1fs CPU(search), %.2f xRT; %.1fs Elapsed, %.2f xRT\n",
uttid, f,
kb->tm->t_cpu, kb->tm->t_cpu * 100.0 / f,
kb->tm_search->t_cpu, kb->tm_search->t_cpu * 100.0 / f,
kb->tm->t_elapsed, kb->tm->t_elapsed * 100.0 / f);
printf("CTR(%s): %5d frames; %d Sen (%.1f/fr); %d HMM (%.1f/fr); %d Words (%.1f/fr)\n",
uttid, f,
kb->n_sen_eval, ((float64)kb->n_sen_eval) / f,
kb->n_hmm_eval, ((float64)kb->n_hmm_eval) / f,
n_vithist, ((float64) n_vithist) / f);
}
/* Cleanup */
glist_free (kb->lextree_active);
kb->lextree_active = NULL;
for (; f >= -1; --f) { /* I.e., including dummy START_WORD node at frame -1 */
glist_myfree (kb->vithist[f], sizeof(vithist_t));
kb->vithist[f] = NULL;
}
lm_cache_reset (kb->lm);
}
