/*
* Compute the top-N closest gaussians from the chosen set (mgau,feat)
* for the given input observation vector.
*/
static int32
compute_dist(gauden_dist_t * out_dist, int32 n_top,
mfcc_t * obs, int32 featlen,
mfcc_t ** mean, mfcc_t ** var, mfcc_t * det,
int32 n_density)
{
int32 i, j, d;
gauden_dist_t *worst;
/* Special case optimization when n_density <= n_top */
if (n_top >= n_density)
return (compute_dist_all
(out_dist, obs, featlen, mean, var, det, n_density));
for (i = 0; i < n_top; i++)
out_dist[i].dist = WORST_DIST;
worst = &(out_dist[n_top - 1]);
for (d = 0; d < n_density; d++) {
mfcc_t *m;
mfcc_t *v;
mfcc_t dval;
m = mean[d];
v = var[d];
dval = det[d];
for (i = 0; (i < featlen) && (dval >= worst->dist); i++) {
mfcc_t diff;
#ifdef FIXED_POINT
/* Have to check for underflows here. */
mfcc_t pdval = dval;
diff = obs[i] - m[i];
dval -= MFCCMUL(MFCCMUL(diff, diff), v[i]);
if (dval > pdval) {
dval = WORST_SCORE;
break;
}
#else
diff = obs[i] - m[i];
/* The compiler really likes this to be a single
* expression, for whatever reason. */
dval -= diff * diff * v[i];
#endif
}
if ((i < featlen) || (dval < worst->dist)) /* Codeword d worse than worst */
continue;
/* Codeword d at least as good as worst so far; insert in the ordered list */
for (i = 0; (i < n_top) && (dval < out_dist[i].dist); i++);
assert(i < n_top);
for (j = n_top - 1; j > i; --j)
out_dist[j] = out_dist[j - 1];
out_dist[i].dist = dval;
out_dist[i].id = d;
}
return 0;
}
static void
eval_cb(s2_semi_mgau_t *s, int32 feat, mfcc_t *z)
{
vqFeature_t *worst, *best, *topn;
mfcc_t *mean;
mfcc_t *var, *det, *detP, *detE;
int32 i, ceplen;
best = topn = s->f[feat];
worst = topn + (s->max_topn - 1);
mean = s->means[feat][0];
var = s->vars[feat][0];
det = s->dets[feat];
detE = det + s->n_density;
ceplen = s->veclen[feat];
for (detP = det; detP < detE; ++detP) {
mfcc_t diff, sqdiff, compl; /* diff, diff^2, component likelihood */
mfcc_t d;
mfcc_t *obs;
vqFeature_t *cur;
int32 cw, j;
d = *detP;
obs = z;
cw = detP - det;
for (j = 0; (j < ceplen) && (d >= worst->score); ++j) {
diff = *obs++ - *mean++;
sqdiff = MFCCMUL(diff, diff);
compl = MFCCMUL(sqdiff, *var);
d = GMMSUB(d, compl);
++var;
}
if (j < ceplen) {
/* terminated early, so not in topn */
mean += (ceplen - j);
var += (ceplen - j);
continue;
}
if ((int32)d < worst->score)
continue;
for (i = 0; i < s->max_topn; i++) {
/* already there, so don't need to insert */
if (topn[i].codeword == cw)
break;
}
if (i < s->max_topn)
continue; /* already there. Don't insert */
/* remaining code inserts codeword and dist in correct spot */
for (cur = worst - 1; cur >= best && (int32)d >= cur->score; --cur)
memcpy(cur + 1, cur, sizeof(vqFeature_t));
++cur;
cur->codeword = cw;
cur->score = (int32)d;
}
}
static void
eval_cb_kdtree(s2_semi_mgau_t *s, int32 feat, mfcc_t *z,
kd_tree_node_t *node, uint32 maxbbi)
{
vqFeature_t *worst, *best, *topn;
int32 i, ceplen;
best = topn = s->f[feat];
worst = topn + (s->max_topn - 1);
ceplen = s->veclen[feat];
for (i = 0; i < maxbbi; ++i) {
mfcc_t *mean, diff, sqdiff, compl; /* diff, diff^2, component likelihood */
mfcc_t *var, d;
mfcc_t *obs;
vqFeature_t *cur;
int32 cw, j, k;
cw = node->bbi[i];
mean = s->means[feat] + cw * ceplen;
var = s->vars[feat] + cw * ceplen;
d = s->dets[feat][cw];
obs = z;
for (j = 0; (j < ceplen) && (d >= worst->score); j++) {
diff = *obs++ - *mean++;
sqdiff = MFCCMUL(diff, diff);
compl = MFCCMUL(sqdiff, *var);
d = GMMSUB(d, compl);
++var;
}
if (j < ceplen)
continue;
if ((int32)d < worst->score)
continue;
for (k = 0; k < s->max_topn; k++) {
/* already there, so don't need to insert */
if (topn[k].codeword == cw)
break;
}
if (k < s->max_topn)
continue; /* already there. Don't insert */
/* remaining code inserts codeword and dist in correct spot */
for (cur = worst - 1; cur >= best && (int32)d >= cur->score; --cur)
memcpy(cur + 1, cur, sizeof(vqFeature_t));
++cur;
cur->codeword = cw;
cur->score = (int32)d;
}
}
static void
cmn_prior_shiftwin(cmn_t *cmn)
{
mfcc_t sf;
int32 i;
E_INFO("cmn_prior_update: from < ");
for (i = 0; i < cmn->veclen; i++)
E_INFOCONT("%5.2f ", MFCC2FLOAT(cmn->cmn_mean[i]));
E_INFOCONT(">\n");
sf = FLOAT2MFCC(1.0) / cmn->nframe;
for (i = 0; i < cmn->veclen; i++)
cmn->cmn_mean[i] = cmn->sum[i] / cmn->nframe; /* sum[i] * sf */
/* Make the accumulation decay exponentially */
if (cmn->nframe >= CMN_WIN_HWM) {
sf = CMN_WIN * sf;
for (i = 0; i < cmn->veclen; i++)
cmn->sum[i] = MFCCMUL(cmn->sum[i], sf);
cmn->nframe = CMN_WIN;
}
E_INFO("cmn_prior_update: to < ");
for (i = 0; i < cmn->veclen; i++)
E_INFOCONT("%5.2f ", MFCC2FLOAT(cmn->cmn_mean[i]));
E_INFOCONT(">\n");
}
/* See compute_dist below */
static int32
compute_dist_all(gauden_dist_t * out_dist, mfcc_t* obs, int32 featlen,
mfcc_t ** mean, mfcc_t ** var, mfcc_t * det,
int32 n_density)
{
int32 i, d;
for (d = 0; d < n_density; ++d) {
mfcc_t *m;
mfcc_t *v;
mfcc_t dval;
m = mean[d];
v = var[d];
dval = det[d];
for (i = 0; i < featlen; i++) {
mfcc_t diff;
#ifdef FIXED_POINT
/* Have to check for underflows here. */
mfcc_t pdval = dval;
diff = obs[i] - m[i];
dval -= MFCCMUL(MFCCMUL(diff, diff), v[i]);
if (dval > pdval) {
dval = WORST_SCORE;
break;
}
#else
diff = obs[i] - m[i];
/* The compiler really likes this to be a single
* expression, for whatever reason. */
dval -= diff * diff * v[i];
#endif
}
out_dist[d].dist = dval;
out_dist[d].id = d;
}
return 0;
}
static void
eval_topn(s2_semi_mgau_t *s, int32 feat, mfcc_t *z)
{
int i, ceplen;
vqFeature_t *topn;
topn = s->f[feat];
ceplen = s->veclen[feat];
for (i = 0; i < s->max_topn; i++) {
mfcc_t *mean, diff, sqdiff, compl; /* diff, diff^2, component likelihood */
vqFeature_t vtmp;
mfcc_t *var, d;
mfcc_t *obs;
int32 cw, j;
cw = topn[i].codeword;
mean = s->means[feat][0] + cw * ceplen;
var = s->vars[feat][0] + cw * ceplen;
d = s->dets[feat][cw];
obs = z;
for (j = 0; j < ceplen; j++) {
diff = *obs++ - *mean++;
sqdiff = MFCCMUL(diff, diff);
compl = MFCCMUL(sqdiff, *var);
d = GMMSUB(d, compl);
++var;
}
topn[i].score = (int32)d;
if (i == 0)
continue;
vtmp = topn[i];
for (j = i - 1; j >= 0 && (int32)d > topn[j].score; j--) {
topn[j + 1] = topn[j];
}
topn[j + 1] = vtmp;
}
}
void
feat_lda_transform(feat_t *fcb, mfcc_t ***inout_feat, uint32 nfr)
{
mfcc_t *tmp;
uint32 i, j, k;
tmp = ckd_calloc(fcb->stream_len[0], sizeof(mfcc_t));
for (i = 0; i < nfr; ++i) {
/* Do the matrix multiplication inline here since fcb->lda
* is transposed (eigenvectors in rows not columns). */
/* FIXME: In the future we ought to use the BLAS. */
memset(tmp, 0, sizeof(mfcc_t) * fcb->stream_len[0]);
for (j = 0; j < feat_dimension(fcb); ++j) {
for (k = 0; k < fcb->stream_len[0]; ++k) {
tmp[j] += MFCCMUL(inout_feat[i][0][k], fcb->lda[0][j][k]);
}
}
memcpy(inout_feat[i][0], tmp, fcb->stream_len[0] * sizeof(mfcc_t));
}
ckd_free(tmp);
}
void
cmn(cmn_t *cmn, mfcc_t ** mfc, int32 varnorm, int32 n_frame)
{
mfcc_t *mfcp;
mfcc_t t;
int32 i, f;
oe_assert(mfc != NULL);
if (n_frame <= 0)
return;
/* If cmn->cmn_mean wasn't NULL, we need to zero the contents */
memset(cmn->cmn_mean, 0, cmn->veclen * sizeof(mfcc_t));
/* Find mean cep vector for this utterance */
for (f = 0; f < n_frame; f++) {
mfcp = mfc[f];
for (i = 0; i < cmn->veclen; i++) {
cmn->cmn_mean[i] += mfcp[i];
}
}
for (i = 0; i < cmn->veclen; i++)
cmn->cmn_mean[i] /= n_frame;
E_INFO("CMN: ");
for (i = 0; i < cmn->veclen; i++)
E_INFOCONT("%5.2f ", MFCC2FLOAT(cmn->cmn_mean[i]));
E_INFOCONT("\n");
if (!varnorm) {
/* Subtract mean from each cep vector */
for (f = 0; f < n_frame; f++) {
mfcp = mfc[f];
for (i = 0; i < cmn->veclen; i++)
mfcp[i] -= cmn->cmn_mean[i];
}
}
else {
/* Scale cep vectors to have unit variance along each dimension, and subtract means */
/* If cmn->cmn_var wasn't NULL, we need to zero the contents */
memset(cmn->cmn_var, 0, cmn->veclen * sizeof(mfcc_t));
for (f = 0; f < n_frame; f++) {
mfcp = mfc[f];
for (i = 0; i < cmn->veclen; i++) {
t = mfcp[i] - cmn->cmn_mean[i];
cmn->cmn_var[i] += MFCCMUL(t, t);
}
}
for (i = 0; i < cmn->veclen; i++)
/* Inverse Std. Dev, RAH added type case from sqrt */
cmn->cmn_var[i] = FLOAT2MFCC(sqrt((float64)n_frame / MFCC2FLOAT(cmn->cmn_var[i])));
for (f = 0; f < n_frame; f++) {
mfcp = mfc[f];
for (i = 0; i < cmn->veclen; i++)
mfcp[i] = MFCCMUL((mfcp[i] - cmn->cmn_mean[i]), cmn->cmn_var[i]);
}
}
}
