static int
acmod_init_feat(acmod_t *acmod)
{
acmod->fcb =
feat_init(cmd_ln_str_r(acmod->config, "-feat"),
cmn_type_from_str(cmd_ln_str_r(acmod->config,"-cmn")),
cmd_ln_boolean_r(acmod->config, "-varnorm"),
agc_type_from_str(cmd_ln_str_r(acmod->config, "-agc")),
1, cmd_ln_int32_r(acmod->config, "-ceplen"));
if (acmod->fcb == NULL)
return -1;
if (cmd_ln_str_r(acmod->config, "-lda")) {
E_INFO("Reading linear feature transformation from %s\n",
cmd_ln_str_r(acmod->config, "-lda"));
if (feat_read_lda(acmod->fcb,
cmd_ln_str_r(acmod->config, "-lda"),
cmd_ln_int32_r(acmod->config, "-ldadim")) < 0)
return -1;
}
if (cmd_ln_str_r(acmod->config, "-svspec")) {
int32 **subvecs;
E_INFO("Using subvector specification %s\n",
cmd_ln_str_r(acmod->config, "-svspec"));
if ((subvecs = parse_subvecs(cmd_ln_str_r(acmod->config, "-svspec"))) == NULL)
return -1;
if ((feat_set_subvecs(acmod->fcb, subvecs)) < 0)
return -1;
}
if (cmd_ln_exists_r(acmod->config, "-agcthresh")
&& 0 != strcmp(cmd_ln_str_r(acmod->config, "-agc"), "none")) {
agc_set_threshold(acmod->fcb->agc_struct,
cmd_ln_float32_r(acmod->config, "-agcthresh"));
}
if (acmod->fcb->cmn_struct
&& cmd_ln_exists_r(acmod->config, "-cmninit")) {
char *c, *cc, *vallist;
int32 nvals;
vallist = ckd_salloc(cmd_ln_str_r(acmod->config, "-cmninit"));
c = vallist;
nvals = 0;
while (nvals < acmod->fcb->cmn_struct->veclen
&& (cc = strchr(c, ',')) != NULL) {
*cc = '\0';
acmod->fcb->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof_c(c));
c = cc + 1;
++nvals;
}
if (nvals < acmod->fcb->cmn_struct->veclen && *c != '\0') {
acmod->fcb->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof_c(c));
}
ckd_free(vallist);
}
return 0;
}
int
main(int argc, char *argv[])
{
feat_t *fcb;
mfcc_t **in_feats, ***out_feats, ***out_feats2, ***optr;
int32 i, j, ncep, nfr, nfr1, nfr2;
in_feats = (mfcc_t **)ckd_alloc_2d_ptr(6, 13, data, sizeof(mfcc_t));
out_feats = (mfcc_t ***)ckd_calloc_3d(8, 1, 39, sizeof(mfcc_t));
/* Test 1s_c_d_dd features */
fcb = feat_init("1s_c_d_dd", CMN_NONE, 0, AGC_NONE, 1, 13);
ncep = 6;
nfr1 = feat_s2mfc2feat_live(fcb, in_feats, &ncep, 1, 1, out_feats);
printf("Processed %d input %d output frames\n", ncep, nfr1);
for (i = 0; i < nfr1; ++i) {
printf("%d: ", i);
for (j = 0; j < 39; ++j) {
printf("%.3f ", MFCC2FLOAT(out_feats[i][0][j]));
}
printf("\n");
}
feat_free(fcb);
/* Test in "live" mode. */
fcb = feat_init("1s_c_d_dd", CMN_NONE, 0, AGC_NONE, 1, 13);
optr = out_feats2 = (mfcc_t ***)ckd_calloc_3d(8, 1, 39, sizeof(mfcc_t));
nfr2 = 0;
ncep = 2;
nfr = feat_s2mfc2feat_live(fcb, in_feats, &ncep, TRUE, FALSE, optr);
printf("Processed %d input %d output frames\n", ncep, nfr);
nfr2 += nfr;
for (i = 0; i < nfr; ++i) {
printf("%d: ", i);
for (j = 0; j < 39; ++j) {
printf("%.3f ", MFCC2FLOAT(optr[i][0][j]));
}
printf("\n");
}
optr += nfr;
ncep = 2;
nfr = feat_s2mfc2feat_live(fcb, in_feats + 2, &ncep, FALSE, FALSE, optr);
nfr2 += nfr;
printf("Processed %d input %d output frames\n", ncep, nfr);
for (i = 0; i < nfr; ++i) {
printf("%d: ", i);
for (j = 0; j < 39; ++j) {
printf("%.3f ", MFCC2FLOAT(optr[i][0][j]));
}
printf("\n");
}
optr += nfr;
ncep = 2;
nfr = feat_s2mfc2feat_live(fcb, in_feats + 4, &ncep, FALSE, TRUE, optr);
nfr2 += nfr;
printf("Processed %d input %d output frames\n", ncep, nfr);
for (i = 0; i < nfr; ++i) {
printf("%d: ", i);
for (j = 0; j < 39; ++j) {
printf("%.3f ", MFCC2FLOAT(optr[i][0][j]));
}
printf("\n");
}
optr += nfr;
feat_free(fcb);
TEST_EQUAL(nfr1, nfr2);
for (i = 0; i < nfr1; ++i) {
for (j = 0; j < 39; ++j) {
TEST_EQUAL(out_feats[i][0][j], out_feats2[i][0][j]);
}
}
ckd_free_3d(out_feats2);
ckd_free_3d(out_feats);
ckd_free(in_feats);
return 0;
}
int
main_initialize(int argc,
char *argv[],
lexicon_t **out_lex,
model_def_t **out_omdef,
model_def_t **out_dmdef,
feat_t** out_feat)
{
model_def_t *dmdef = NULL;
model_def_t *omdef = NULL;
lexicon_t *lex = NULL;
feat_t *feat;
const char *fn;
uint32 n_ts;
uint32 n_cb;
const char *ts2cbfn;
parse_cmd_ln(argc, argv);
feat =
feat_init(cmd_ln_str("-feat"),
cmn_type_from_str(cmd_ln_str("-cmn")),
cmd_ln_boolean("-varnorm"),
agc_type_from_str(cmd_ln_str("-agc")),
1, cmd_ln_int32("-ceplen"));
if (cmd_ln_str("-lda")) {
E_INFO("Reading linear feature transformation from %s\n",
cmd_ln_str("-lda"));
if (feat_read_lda(feat,
cmd_ln_str("-lda"),
cmd_ln_int32("-ldadim")) < 0)
return -1;
}
if (cmd_ln_str("-svspec")) {
int32 **subvecs;
E_INFO("Using subvector specification %s\n",
cmd_ln_str("-svspec"));
if ((subvecs = parse_subvecs(cmd_ln_str("-svspec"))) == NULL)
return -1;
if ((feat_set_subvecs(feat, subvecs)) < 0)
return -1;
}
if (cmd_ln_exists("-agcthresh")
&& 0 != strcmp(cmd_ln_str("-agc"), "none")) {
agc_set_threshold(feat->agc_struct,
cmd_ln_float32("-agcthresh"));
}
if (feat->cmn_struct
&& cmd_ln_exists("-cmninit")) {
char *c, *cc, *vallist;
int32 nvals;
vallist = ckd_salloc(cmd_ln_str("-cmninit"));
c = vallist;
nvals = 0;
while (nvals < feat->cmn_struct->veclen
&& (cc = strchr(c, ',')) != NULL) {
*cc = '\0';
feat->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof(c));
c = cc + 1;
++nvals;
}
if (nvals < feat->cmn_struct->veclen && *c != '\0') {
feat->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof(c));
}
ckd_free(vallist);
}
*out_feat = feat;
if (cmd_ln_str("-omoddeffn")) {
E_INFO("Reading output model definitions: %s\n", cmd_ln_str("-omoddeffn"));
/* Read in the model definitions. Defines the set of
CI phones and context dependent phones. Defines the
transition matrix tying and state level tying. */
if (model_def_read(&omdef,
cmd_ln_str("-omoddeffn")) != S3_SUCCESS) {
return S3_ERROR;
}
if (cmd_ln_str("-dmoddeffn")) {
E_INFO("Reading dump model definitions: %s\n", cmd_ln_str("-dmoddeffn"));
if (model_def_read(&dmdef,
cmd_ln_str("-dmoddeffn")) != S3_SUCCESS) {
return S3_ERROR;
}
setup_d2o_map(dmdef, omdef);
}
else {
E_INFO("Assuming dump and output model definitions are identical\n");
}
ts2cbfn = cmd_ln_str("-ts2cbfn");
if (ts2cbfn) {
if (strcmp(SEMI_LABEL, ts2cbfn) == 0) {
omdef->cb = semi_ts2cb(omdef->n_tied_state);
n_ts = omdef->n_tied_state;
n_cb = 1;
}
else if (strcmp(CONT_LABEL, ts2cbfn) == 0) {
omdef->cb = cont_ts2cb(omdef->n_tied_state);
n_ts = omdef->n_tied_state;
n_cb = omdef->n_tied_state;
}
else if (strcmp(PTM_LABEL, ts2cbfn) == 0) {
omdef->cb = ptm_ts2cb(omdef);
n_ts = omdef->n_tied_state;
n_cb = omdef->acmod_set->n_ci;
}
else if (s3ts2cb_read(cmd_ln_str("-ts2cbfn"),
&omdef->cb,
&n_ts,
&n_cb) != S3_SUCCESS) {
return S3_ERROR;
}
if (omdef->n_tied_state != n_ts) {
E_FATAL("Model definition file n_tied_state = %u, but %u mappings in ts2cb\n",
omdef->n_tied_state, n_ts);
}
}
}
else {
E_INFO("No mdef files. Assuming 1-class init\n");
}
*out_omdef = omdef;
*out_dmdef = dmdef;
fn = cmd_ln_str("-dictfn");
if (fn) {
E_INFO("Reading main lexicon: %s\n", fn);
lex = lexicon_read(NULL,
fn,
omdef->acmod_set);
if (lex == NULL)
return S3_ERROR;
}
fn = cmd_ln_str("-fdictfn");
if (fn) {
E_INFO("Reading filler lexicon: %s\n", fn);
(void)lexicon_read(lex,
fn,
omdef->acmod_set);
}
*out_lex = lex;
stride = cmd_ln_int32("-stride");
return S3_SUCCESS;
}
main (int32 argc, char *argv[])
{
char *str;
#if 0
ckd_debug(100000);
#endif
E_INFO("%s COMPILED ON: %s, AT: %s\n\n", argv[0], __DATE__, __TIME__);
/* Digest command line argument definitions */
cmd_ln_define (defn);
if ((argc == 2) && (strcmp (argv[1], "help") == 0)) {
cmd_ln_print_definitions();
exit(1);
}
/* Look for default or specified arguments file */
str = NULL;
if ((argc == 2) && (argv[1][0] != '-'))
str = argv[1];
else if (argc == 1) {
str = "s3align.arg";
E_INFO("Looking for default argument file: %s\n", str);
}
if (str) {
/* Build command line argument list from file */
if ((argc = load_argfile (str, argv[0], &argv)) < 0) {
fprintf (stderr, "Usage:\n");
fprintf (stderr, "\t%s argument-list, or\n", argv[0]);
fprintf (stderr, "\t%s [argument-file] (default file: s3align.arg)\n\n",
argv[0]);
cmd_ln_print_definitions();
exit(1);
}
}
cmdline_parse (argc, argv);
if ((cmd_ln_access("-mdeffn") == NULL) ||
(cmd_ln_access("-meanfn") == NULL) ||
(cmd_ln_access("-varfn") == NULL) ||
(cmd_ln_access("-mixwfn") == NULL) ||
(cmd_ln_access("-tmatfn") == NULL) ||
(cmd_ln_access("-dictfn") == NULL))
E_FATAL("Missing -mdeffn, -meanfn, -varfn, -mixwfn, -tmatfn, or -dictfn argument\n");
if ((cmd_ln_access("-ctlfn") == NULL) || (cmd_ln_access("-insentfn") == NULL))
E_FATAL("Missing -ctlfn or -insentfn argument\n");
if ((cmd_ln_access ("-s2stsegdir") == NULL) &&
(cmd_ln_access ("-stsegdir") == NULL) &&
(cmd_ln_access ("-phsegdir") == NULL) &&
(cmd_ln_access ("-wdsegdir") == NULL) &&
(cmd_ln_access ("-outsentfn") == NULL))
E_FATAL("Missing output file/directory argument(s)\n");
tm_utt = timing_new ();
/*
* Initialize log(S3-base). All scores (probs...) computed in log domain to avoid
* underflow. At the same time, log base = 1.0001 (1+epsilon) to allow log values
* to be maintained in int32 variables without significant loss of precision.
*/
if (cmd_ln_access("-logbase") == NULL)
logs3_init (1.0001);
else {
float32 logbase;
logbase = *((float32 *) cmd_ln_access("-logbase"));
if (logbase <= 1.0)
E_FATAL("Illegal log-base: %e; must be > 1.0\n", logbase);
if (logbase > 1.1)
E_WARN("Logbase %e perhaps too large??\n", logbase);
logs3_init ((float64) logbase);
}
/* Initialize feature stream type */
feat_init ((char *) cmd_ln_access ("-feat"));
/* BHIKSHA: PASS CEPSIZE TO FEAT_CEPSIZE, 6 Jan 98 */
cepsize = *((int32 *) cmd_ln_access("-ceplen"));
cepsize = feat_cepsize (cepsize);
/* END CHANGES BY BHIKSHA */
/* Read in input databases */
models_init ();
senscale = (int32 *) ckd_calloc (S3_MAX_FRAMES, sizeof(int32));
tmr_utt = cyctimer_new ("U");
tmr_gauden = cyctimer_new ("G");
tmr_senone = cyctimer_new ("S");
tmr_align = cyctimer_new ("A");
/* Initialize align module */
align_init ();
printf ("\n");
tot_nfr = 0;
process_ctlfile ();
if (tot_nfr > 0) {
printf ("\n");
printf("TOTAL FRAMES: %8d\n", tot_nfr);
printf("TOTAL CPU TIME: %11.2f sec, %7.2f xRT\n",
tm_utt->t_tot_cpu, tm_utt->t_tot_cpu/(tot_nfr*0.01));
printf("TOTAL ELAPSED TIME: %11.2f sec, %7.2f xRT\n",
tm_utt->t_tot_elapsed, tm_utt->t_tot_elapsed/(tot_nfr*0.01));
}
#if (! WIN32)
system ("ps aguxwww | grep s3align");
#endif
/* Hack!! To avoid hanging problem under Linux */
if (logfp) {
fclose (logfp);
*stdout = orig_stdout;
*stderr = orig_stderr;
}
exit(0);
}
static void
models_init(cmd_ln_t *config)
{
int32 cisencnt;
kbc = New_kbcore(config);
kbc->logmath = logs3_init(cmd_ln_float64_r(config, "-logbase"), 1,
cmd_ln_int32_r(config, "-log3table"));
/* Initialize feaure stream type */
kbc->fcb = feat_init(cmd_ln_str_r(config, "-feat"),
cmn_type_from_str(cmd_ln_str_r(config, "-cmn")),
cmd_ln_boolean_r(config, "-varnorm"),
agc_type_from_str(cmd_ln_str_r(config, "-agc")), 1,
cmd_ln_int32_r(config, "-ceplen"));
s3_am_init(kbc);
/* Initialize the front end if -adcin is specified */
if (cmd_ln_exists_r(config, "-adcin") && cmd_ln_boolean_r(config, "-adcin")) {
if ((fe = fe_init_auto_r(config)) == NULL) {
E_FATAL("fe_init_auto_r() failed\n");
}
}
assert(kbc);
assert(kbc->mdef);
assert(kbc->tmat);
/* Dictionary */
dict = dict_init(kbc->mdef, cmd_ln_str_r(config, "-dict"),
cmd_ln_str_r(config, "-fdict"),
cmd_ln_int32_r(config, "-lts_mismatch"),
cmd_ln_boolean_r(config, "-mdef_fillers"),
/* Never do mdef filler phones. */
FALSE,
TRUE);
for (cisencnt = 0; cisencnt == kbc->mdef->cd2cisen[cisencnt];
cisencnt++);
ascr = ascr_init(kbc->mdef->n_sen, 0, /* No composite senone */
mdef_n_sseq(kbc->mdef), 0, /* No composite senone sequence */
1, /* Phoneme lookahead window =1. Not enabled phoneme lookahead at this moment */
cisencnt);
fastgmm = fast_gmm_init(cmd_ln_int32_r(config, "-ds"),
cmd_ln_int32_r(config, "-cond_ds"),
cmd_ln_int32_r(config, "-dist_ds"),
cmd_ln_int32_r(config, "-gs4gs"),
cmd_ln_int32_r(config, "-svq4svq"),
cmd_ln_float64_r(config, "-subvqbeam"),
cmd_ln_float64_r(config, "-ci_pbeam"),
cmd_ln_float64_r(config, "-tighten_factor"),
cmd_ln_int32_r(config, "-maxcdsenpf"),
kbc->mdef->n_ci_sen,
kbc->logmath);
adapt_am = adapt_am_init();
}
static int
initialize(int argc,
char *argv[])
{
const char *fdictfn;
const char *dictfn;
const char *ts2cbfn;
uint32 n_ts;
uint32 n_cb;
/* define, parse and (partially) validate the command line */
parse_cmd_ln(argc, argv);
feat =
feat_init(cmd_ln_str("-feat"),
cmn_type_from_str(cmd_ln_str("-cmn")),
cmd_ln_boolean("-varnorm"),
agc_type_from_str(cmd_ln_str("-agc")),
1, cmd_ln_int32("-ceplen"));
if (cmd_ln_str("-lda")) {
E_INFO("Reading linear feature transformation from %s\n",
cmd_ln_str("-lda"));
if (feat_read_lda(feat,
cmd_ln_str("-lda"),
cmd_ln_int32("-ldadim")) < 0)
return -1;
}
if (cmd_ln_str("-svspec")) {
int32 **subvecs;
E_INFO("Using subvector specification %s\n",
cmd_ln_str("-svspec"));
if ((subvecs = parse_subvecs(cmd_ln_str("-svspec"))) == NULL)
return -1;
if ((feat_set_subvecs(feat, subvecs)) < 0)
return -1;
}
if (cmd_ln_exists("-agcthresh")
&& 0 != strcmp(cmd_ln_str("-agc"), "none")) {
agc_set_threshold(feat->agc_struct,
cmd_ln_float32("-agcthresh"));
}
if (feat->cmn_struct
&& cmd_ln_exists("-cmninit")) {
char *c, *cc, *vallist;
int32 nvals;
vallist = ckd_salloc(cmd_ln_str("-cmninit"));
c = vallist;
nvals = 0;
while (nvals < feat->cmn_struct->veclen
&& (cc = strchr(c, ',')) != NULL) {
*cc = '\0';
feat->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof(c));
c = cc + 1;
++nvals;
}
if (nvals < feat->cmn_struct->veclen && *c != '\0') {
feat->cmn_struct->cmn_mean[nvals] = FLOAT2MFCC(atof(c));
}
ckd_free(vallist);
}
if (cmd_ln_str("-segdir"))
corpus_set_seg_dir(cmd_ln_str("-segdir"));
if (cmd_ln_str("-segext"))
corpus_set_seg_ext(cmd_ln_str("-segext"));
corpus_set_mfcc_dir(cmd_ln_str("-cepdir"));
corpus_set_mfcc_ext(cmd_ln_str("-cepext"));
if (cmd_ln_str("-lsnfn"))
corpus_set_lsn_filename(cmd_ln_str("-lsnfn"));
corpus_set_ctl_filename(cmd_ln_str("-ctlfn"));
if (cmd_ln_int32("-nskip") && cmd_ln_int32("-runlen")) {
corpus_set_interval(cmd_ln_int32("-nskip"),
cmd_ln_int32("-runlen"));
} else if (cmd_ln_int32("-part") && cmd_ln_int32("-npart")) {
corpus_set_partition(cmd_ln_int32("-part"),
cmd_ln_int32("-npart"));
}
if (corpus_init() != S3_SUCCESS) {
return S3_ERROR;
}
if (cmd_ln_str("-moddeffn")) {
E_INFO("Reading %s\n", cmd_ln_str("-moddeffn"));
/* Read in the model definitions. Defines the set of
CI phones and context dependent phones. Defines the
transition matrix tying and state level tying. */
if (model_def_read(&mdef,
cmd_ln_str("-moddeffn")) != S3_SUCCESS) {
return S3_ERROR;
}
ts2cbfn = cmd_ln_str("-ts2cbfn");
if (strcmp(SEMI_LABEL, ts2cbfn) == 0) {
mdef->cb = semi_ts2cb(mdef->n_tied_state);
n_ts = mdef->n_tied_state;
n_cb = 1;
}
else if (strcmp(CONT_LABEL, ts2cbfn) == 0) {
mdef->cb = cont_ts2cb(mdef->n_tied_state);
n_ts = mdef->n_tied_state;
n_cb = mdef->n_tied_state;
}
else if (strcmp(PTM_LABEL, ts2cbfn) == 0) {
mdef->cb = ptm_ts2cb(mdef);
n_ts = mdef->n_tied_state;
n_cb = mdef->acmod_set->n_ci;
}
else if (s3ts2cb_read(ts2cbfn,
&mdef->cb,
&n_ts,
&n_cb) != S3_SUCCESS) {
return S3_ERROR;
}
dictfn = cmd_ln_str("-dictfn");
if (dictfn == NULL) {
E_FATAL("You must specify a content dictionary using -dictfn\n");
}
E_INFO("Reading %s\n", dictfn);
lex = lexicon_read(NULL, /* no lexicon to start */
dictfn,
mdef->acmod_set);
if (lex == NULL)
return S3_ERROR;
fdictfn = cmd_ln_str("-fdictfn");
if (fdictfn) {
E_INFO("Reading %s\n", fdictfn);
(void)lexicon_read(lex, /* add filler words content lexicon */
fdictfn,
mdef->acmod_set);
}
}
return S3_SUCCESS;
}
kbcore_t *kbcore_init (float64 logbase,
char *feattype,
char *cmn,
char *varnorm,
char *agc,
char *mdeffile,
char *dictfile,
char *fdictfile,
char *compsep,
char *lmfile,
char *fillpenfile,
float64 silprob,
float64 fillprob,
float64 langwt,
float64 inspen,
float64 uw,
char *meanfile,
char *varfile,
float64 varfloor,
char *mixwfile,
float64 mixwfloor,
char *subvqfile,
char *tmatfile,
float64 tmatfloor)
{
kbcore_t *kb;
E_INFO("Initializing core models:\n");
kb = (kbcore_t *) ckd_calloc (1, sizeof(kbcore_t));
kb->fcb = NULL;
kb->mdef = NULL;
kb->dict = NULL;
kb->dict2pid = NULL;
kb->lm = NULL;
kb->fillpen = NULL;
kb->dict2lmwid = NULL;
kb->mgau = NULL;
kb->svq = NULL;
kb->tmat = NULL;
logs3_init (logbase);
if (feattype) {
if ((kb->fcb = feat_init (feattype, cmn, varnorm, agc)) == NULL)
E_FATAL("feat_init(%s) failed\n", feattype);
if (feat_n_stream(kb->fcb) != 1)
E_FATAL("#Feature streams(%d) != 1\n", feat_n_stream(kb->fcb));
}
if (mdeffile) {
if ((kb->mdef = mdef_init (mdeffile)) == NULL)
E_FATAL("mdef_init(%s) failed\n", mdeffile);
}
if (dictfile) {
if (! compsep)
compsep = "";
else if ((compsep[0] != '\0') && (compsep[1] != '\0')) {
E_FATAL("Compound word separator(%s) must be empty or single character string\n",
compsep);
}
if ((kb->dict = dict_init (kb->mdef, dictfile, fdictfile, compsep[0])) == NULL)
E_FATAL("dict_init(%s,%s,%s) failed\n", dictfile,
fdictfile ? fdictfile : "", compsep);
}
if (lmfile) {
if ((kb->lm = lm_read (lmfile, langwt, inspen, uw)) == NULL)
E_FATAL("lm_read(%s, %e, %e, %e) failed\n", lmfile, langwt, inspen, uw);
}
if (fillpenfile || (lmfile && kb->dict)) {
if (! kb->dict) /* Sic */
E_FATAL("No dictionary for associating filler penalty file(%s)\n", fillpenfile);
if ((kb->fillpen = fillpen_init (kb->dict, fillpenfile, silprob, fillprob,
langwt, inspen)) == NULL)
E_FATAL("fillpen_init(%s) failed\n", fillpenfile);
}
if (meanfile) {
if ((! varfile) || (! mixwfile))
E_FATAL("Varfile or mixwfile not specified along with meanfile(%s)\n", meanfile);
kb->mgau = mgau_init (meanfile, varfile, varfloor, mixwfile, mixwfloor, TRUE);
if (kb->mgau == NULL)
E_FATAL("gauden_init(%s, %s, %e) failed\n", meanfile, varfile, varfloor);
if (subvqfile) {
if ((kb->svq = subvq_init (subvqfile, varfloor, -1, kb->mgau)) == NULL)
E_FATAL("subvq_init (%s, %e, -1) failed\n", subvqfile, varfloor);
}
}
if (tmatfile) {
if ((kb->tmat = tmat_init (tmatfile, tmatfloor)) == NULL)
E_FATAL("tmat_init (%s, %e) failed\n", tmatfile, tmatfloor);
}
if (kb->dict && kb->lm) { /* Initialize dict2lmwid */
if ((kb->dict2lmwid = wid_dict_lm_map (kb->dict, kb->lm)) == NULL)
E_FATAL("Dict/LM word-id mapping failed\n");
}
if (kb->mdef && kb->dict) { /* Initialize dict2pid */
kb->dict2pid = dict2pid_build (kb->mdef, kb->dict);
}
/* ***************** Verifications ***************** */
E_INFO("Verifying models consistency:\n");
if (kb->fcb && kb->mgau) {
/* Verify feature streams against gauden codebooks */
if (feat_stream_len(kb->fcb, 0) != mgau_veclen(kb->mgau))
E_FATAL("Feature streamlen(%d) != mgau streamlen(%d)\n",
feat_stream_len(kb->fcb, 0), mgau_veclen(kb->mgau));
}
if (kb->mdef && kb->mgau) {
/* Verify senone parameters against model definition parameters */
if (kb->mdef->n_sen != mgau_n_mgau(kb->mgau))
E_FATAL("Mdef #senones(%d) != mgau #senones(%d)\n",
kb->mdef->n_sen, mgau_n_mgau(kb->mgau));
}
if (kb->mdef && kb->tmat) {
/* Verify transition matrices parameters against model definition parameters */
if (kb->mdef->n_tmat != kb->tmat->n_tmat)
E_FATAL("Mdef #tmat(%d) != tmatfile(%d)\n", kb->mdef->n_tmat, kb->tmat->n_tmat);
if (kb->mdef->n_emit_state != kb->tmat->n_state)
E_FATAL("Mdef #states(%d) != tmat #states(%d)\n",
kb->mdef->n_emit_state, kb->tmat->n_state);
}
return kb;
}
int
main(int argc, char *argv[])
{
feat_t *fcb;
mfcc_t **in_feats, ***out_feats;
int32 i, j, ncep;
/* Test "raw" features without concatenation */
fcb = feat_init("13", CMN_NONE, 0, AGC_NONE, 1, 13);
in_feats = (mfcc_t **)ckd_alloc_2d_ptr(6, 13, data, sizeof(mfcc_t));
out_feats = (mfcc_t ***)ckd_calloc_3d(6, 1, 13, sizeof(mfcc_t));
ncep = 6;
feat_s2mfc2feat_live(fcb, in_feats, &ncep, 1, 1, out_feats);
for (i = 0; i < 6; ++i) {
for (j = 0; j < 13; ++j) {
printf("%.3f ", MFCC2FLOAT(out_feats[i][0][j]));
}
printf("\n");
}
feat_free(fcb);
ckd_free(in_feats);
ckd_free_3d(out_feats);
/* Test "raw" features with concatenation */
fcb = feat_init("13:1", CMN_NONE, 0, AGC_NONE, 1, 13);
in_feats = (mfcc_t **)ckd_alloc_2d_ptr(6, 13, data, sizeof(mfcc_t));
out_feats = (mfcc_t ***)ckd_calloc_3d(8, 1, 39, sizeof(mfcc_t));
ncep = 6;
feat_s2mfc2feat_live(fcb, in_feats, &ncep, 1, 1, out_feats);
for (i = 0; i < 6; ++i) {
for (j = 0; j < 39; ++j) {
printf("%.3f ", MFCC2FLOAT(out_feats[i][0][j]));
}
printf("\n");
}
feat_free(fcb);
/* Test 1s_c_d_dd features */
fcb = feat_init("1s_c_d_dd", CMN_NONE, 0, AGC_NONE, 1, 13);
ncep = 6;
feat_s2mfc2feat_live(fcb, in_feats, &ncep, 1, 1, out_feats);
for (i = 0; i < 6; ++i) {
for (j = 0; j < 39; ++j) {
printf("%.3f ", MFCC2FLOAT(out_feats[i][0][j]));
}
printf("\n");
}
/* Verify that the deltas are correct. */
for (i = 2; i < 4; ++i) {
for (j = 0; j < 13; ++j) {
if (fabs(MFCC2FLOAT(out_feats[i][0][13+j] -
(out_feats[i+2][0][j]
- out_feats[i-2][0][j]))) > 0.01) {
printf("Delta mismatch in [%d][%d]\n", i, j);
return 1;
}
}
}
feat_free(fcb);
ckd_free(in_feats);
ckd_free_3d(out_feats);
return 0;
}