int
acmod_process_feat(acmod_t *acmod,
mfcc_t **feat)
{
int i, inptr;
if (acmod->n_feat_frame == acmod->n_feat_alloc) {
if (acmod->grow_feat)
acmod_grow_feat_buf(acmod, acmod->n_feat_alloc * 2);
else
return 0;
}
if (acmod->grow_feat) {
/* Grow to avoid wraparound if grow_feat == TRUE. */
inptr = acmod->feat_outidx + acmod->n_feat_frame;
while (inptr + 1 >= acmod->n_feat_alloc)
acmod_grow_feat_buf(acmod, acmod->n_feat_alloc * 2);
}
else {
inptr = (acmod->feat_outidx + acmod->n_feat_frame) % acmod->n_feat_alloc;
}
for (i = 0; i < feat_dimension1(acmod->fcb); ++i)
memcpy(acmod->feat_buf[inptr][i],
feat[i], feat_dimension2(acmod->fcb, i) * sizeof(**feat));
++acmod->n_feat_frame;
assert(acmod->n_feat_frame <= acmod->n_feat_alloc);
return 1;
}
void
feat_print(feat_t * fcb, mfcc_t *** feat, int32 nfr, FILE * fp)
{
uint32 i, j, k;
for (i = 0; i < nfr; i++) {
fprintf(fp, "%8d:\n", i);
for (j = 0; j < feat_dimension1(fcb); j++) {
fprintf(fp, "\t%2d:", j);
for (k = 0; k < feat_dimension2(fcb, j); k++)
fprintf(fp, " %8.4f", MFCC2FLOAT(feat[i][j][k]));
fprintf(fp, "\n");
}
}
fflush(fp);
}
void
feat_print(feat_t * fcb, mfcc_t *** feat, int32 nfr, FILE * fp)
{
int32 i, j, k;
for (i = 0; i < nfr; i++) {
#ifndef POCKETSPHINX_NET
fprintf(fp, "%8d:\n", i);
#else
net_fprintf(fp, "{0:8}:\n", i);
#endif
for (j = 0; j < feat_dimension1(fcb); j++) {
#ifndef POCKETSPHINX_NET
fprintf(fp, "\t%2d:", j);
#else
net_fprintf(fp, "\t{0:2}:\n", j);
#endif
for (k = 0; k < (int32)feat_dimension2(fcb, j); k++)
{
#ifndef POCKETSPHINX_NET
fprintf(fp, " %8.4f", MFCC2FLOAT(feat[i][j][k]));
#else
//need check
net_fprintf(fp, " {0:8.4}", MFCC2FLOAT(feat[i][j][k]));
#endif
}
#ifndef POCKETSPHINX_NET
fprintf(fp, "\n");
#else
net_fprintf(fp, "\n");
#endif
}
}
fflush(fp);
}
ps_mgau_t *
ms_mgau_init(acmod_t *acmod, logmath_t *lmath, bin_mdef_t *mdef)
{
/* Codebooks */
ms_mgau_model_t *msg;
ps_mgau_t *mg;
gauden_t *g;
senone_t *s;
cmd_ln_t *config;
int i;
static ps_mgaufuncs_t ms_mgau_funcs = {
"ms",
ms_cont_mgau_frame_eval, /* frame_eval */
ms_mgau_mllr_transform, /* transform */
ms_mgau_free /* free */
};
config = acmod->config;
msg = (ms_mgau_model_t *) ckd_calloc(1, sizeof(ms_mgau_model_t));
msg->config = config;
msg->g = 0;
msg->s = 0;
g = msg->g = gauden_init(cmd_ln_str_r(config, "-mean"),
cmd_ln_str_r(config, "-var"),
cmd_ln_float32_r(config, "-varfloor"),
lmath);
/* Verify n_feat and veclen, against acmod. */
if (g->n_feat != feat_dimension1(acmod->fcb)) {
E_ERROR("Number of streams does not match: %d != %d\n",
g->n_feat, feat_dimension1(acmod->fcb));
goto error_out;
}
for (i = 0; i < g->n_feat; ++i) {
if (g->featlen[i] != feat_dimension2(acmod->fcb, i)) {
E_ERROR("Dimension of stream %d does not match: %d != %d\n", i,
g->featlen[i], feat_dimension2(acmod->fcb, i));
goto error_out;
}
}
s = msg->s = senone_init(msg->g,
cmd_ln_str_r(config, "-mixw"),
cmd_ln_str_r(config, "-senmgau"),
cmd_ln_float32_r(config, "-mixwfloor"),
lmath, mdef);
s->aw = cmd_ln_int32_r(config, "-aw");
/* Verify senone parameters against gauden parameters */
if (s->n_feat != g->n_feat)
E_FATAL("#Feature mismatch: gauden= %d, senone= %d\n", g->n_feat,
s->n_feat);
if (s->n_cw != g->n_density)
E_FATAL("#Densities mismatch: gauden= %d, senone= %d\n",
g->n_density, s->n_cw);
if ((int)s->n_gauden > g->n_mgau)
E_FATAL("Senones need more codebooks (%d) than present (%d)\n",
s->n_gauden, g->n_mgau);
if ((int)s->n_gauden < g->n_mgau)
E_ERROR("Senones use fewer codebooks (%d) than present (%d)\n",
s->n_gauden, g->n_mgau);
msg->topn = cmd_ln_int32_r(config, "-topn");
E_INFO("The value of topn: %d\n", msg->topn);
if (msg->topn == 0 || msg->topn > msg->g->n_density) {
E_WARN
("-topn argument (%d) invalid or > #density codewords (%d); set to latter\n",
msg->topn, msg->g->n_density);
msg->topn = msg->g->n_density;
}
msg->dist = (gauden_dist_t ***)
ckd_calloc_3d(g->n_mgau, g->n_feat, msg->topn,
sizeof(gauden_dist_t));
msg->mgau_active = (uint8*)ckd_calloc(g->n_mgau, sizeof(int8));
mg = (ps_mgau_t *)msg;
mg->vt = &ms_mgau_funcs;
return mg;
error_out:
ms_mgau_free(ps_mgau_base(msg));
return 0;
}
s2_semi_mgau_t *
s2_semi_mgau_init(cmd_ln_t *config, logmath_t *lmath, feat_t *fcb, mdef_t *mdef)
{
s2_semi_mgau_t *s;
char const *sendump_path;
float32 **fgau;
int i;
s = ckd_calloc(1, sizeof(*s));
s->config = config;
s->lmath = logmath_retain(lmath);
/* Log-add table. */
s->lmath_8b = logmath_init(logmath_get_base(lmath), SENSCR_SHIFT, TRUE);
if (s->lmath_8b == NULL) {
s2_semi_mgau_free(s);
return NULL;
}
/* Ensure that it is only 8 bits wide so that fast_logmath_add() works. */
if (logmath_get_width(s->lmath_8b) != 1) {
E_ERROR("Log base %f is too small to represent add table in 8 bits\n",
logmath_get_base(s->lmath_8b));
s2_semi_mgau_free(s);
return NULL;
}
/* Inherit stream dimensions from acmod, will be checked below. */
s->n_feat = feat_dimension1(fcb);
s->veclen = ckd_calloc(s->n_feat, sizeof(int32));
for (i = 0; i < s->n_feat; ++i)
s->veclen[i] = feat_dimension2(fcb, i);
/* Read means and variances. */
if (s3_read_mgau(s, cmd_ln_str_r(s->config, "-mean"), &fgau) < 0) {
s2_semi_mgau_free(s);
return NULL;
}
s->means = (mfcc_t **)fgau;
if (s3_read_mgau(s, cmd_ln_str_r(s->config, "-var"), &fgau) < 0) {
s2_semi_mgau_free(s);
return NULL;
}
s->vars = (mfcc_t **)fgau;
/* Precompute (and fixed-point-ize) means, variances, and determinants. */
s->dets = (mfcc_t **)ckd_calloc_2d(s->n_feat, s->n_density, sizeof(**s->dets));
s3_precomp(s, s->lmath, cmd_ln_float32_r(s->config, "-varfloor"));
/* Read mixture weights */
if ((sendump_path = cmd_ln_str_r(s->config, "-sendump")))
read_sendump(s, mdef, sendump_path);
else
read_mixw(s, cmd_ln_str_r(s->config, "-mixw"),
cmd_ln_float32_r(s->config, "-mixwfloor"));
s->ds_ratio = cmd_ln_int32_r(s->config, "-ds");
/* Determine top-N for each feature */
s->topn_beam = ckd_calloc(s->n_feat, sizeof(*s->topn_beam));
s->max_topn = cmd_ln_int32_r(s->config, "-topn");
split_topn(cmd_ln_str_r(s->config, "-topn_beam"), s->topn_beam, s->n_feat);
E_INFO("Maximum top-N: %d ", s->max_topn);
E_INFOCONT("Top-N beams:");
for (i = 0; i < s->n_feat; ++i) {
E_INFOCONT(" %d", s->topn_beam[i]);
}
E_INFOCONT("\n");
/* Top-N scores from recent frames */
s->n_topn_hist = cmd_ln_int32_r(s->config, "-pl_window") + 2;
s->topn_hist = (vqFeature_t ***)
ckd_calloc_3d(s->n_topn_hist, s->n_feat, s->max_topn,
sizeof(***s->topn_hist));
s->topn_hist_n = ckd_calloc_2d(s->n_topn_hist, s->n_feat,
sizeof(**s->topn_hist_n));
for (i = 0; i < s->n_topn_hist; ++i) {
int j;
for (j = 0; j < s->n_feat; ++j) {
int k;
for (k = 0; k < s->max_topn; ++k) {
s->topn_hist[i][j][k].score = WORST_DIST;
s->topn_hist[i][j][k].codeword = k;
}
}
}
return s;
}
