cst_utterance *f0_targets_to_pm(cst_utterance *utt)
{
cst_item *t;
float pos,lpos,f0,lf0,m;
double time;
int pm;
cst_sts_list *sts_list;
cst_lpcres *target_lpcres;
sts_list = val_sts_list(utt_feat_val(utt,"sts_list"));
lpos = 0;
lf0 = 120; /* hmm */
pm = 0;
time = 0;
/* First pass to count how many pms will be required */
for (t=relation_head(utt_relation(utt,"Target"));
t;
t=item_next(t), lf0 = f0, lpos = pos) /* changed by dhopkins */
{
pos = item_feat_float(t,"pos");
f0 = item_feat_float(t,"f0");
if (time == pos) continue;
m = (f0-lf0)/(pos-lpos);
for ( ; time < pos; pm++)
{
time += 1/(lf0 + ((time-lpos)*m));
}
}
target_lpcres = new_lpcres();
lpcres_resize_frames(target_lpcres,pm);
lpos = 0;
lf0 = 120;
pm = 0;
time = 0;
/* Second pass puts the values in */
for (t=relation_head(utt_relation(utt,"Target"));
t;
t=item_next(t), lf0 = f0, lpos = pos) /* changed by dhopkins */
{
pos = item_feat_float(t,"pos");
f0 = item_feat_float(t,"f0");
if (time == pos) continue;
m = (f0-lf0)/(pos-lpos);
for ( ; time < pos; pm++)
{
time += 1/(lf0 + ((time-lpos)*m));
target_lpcres->times[pm] = sts_list->sample_rate * time;
}
}
utt_set_feat(utt,"target_lpcres",lpcres_val(target_lpcres));
return utt;
}
cst_utterance *asis_to_pm(cst_utterance *utt)
{
/* Copy the PM structure from the units unchanged */
cst_item *u;
cst_lpcres *target_lpcres;
int unit_start, unit_end;
int utt_pms, utt_size, i;
cst_sts_list *sts_list;
sts_list = val_sts_list(utt_feat_val(utt,"sts_list"));
target_lpcres = new_lpcres();
/* Pass one to find the size */
utt_pms = utt_size = 0;
for (u=relation_head(utt_relation(utt,"Unit"));
u;
u=item_next(u))
{
unit_start = item_feat_int(u,"unit_start");
unit_end = item_feat_int(u,"unit_end");
utt_size += get_unit_size(sts_list,unit_start,unit_end);
utt_pms += unit_end - unit_start;
item_set_int(u,"target_end",utt_size);
}
lpcres_resize_frames(target_lpcres,utt_pms);
/* Pass two to fill in the values */
utt_pms = utt_size = 0;
for (u=relation_head(utt_relation(utt,"Unit"));
u;
u=item_next(u))
{
unit_start = item_feat_int(u,"unit_start");
unit_end = item_feat_int(u,"unit_end");
for (i=unit_start; i<unit_end; i++,utt_pms++)
{
utt_size += get_frame_size(sts_list, i);
target_lpcres->times[utt_pms] = utt_size;
}
}
utt_set_feat(utt,"target_lpcres",lpcres_val(target_lpcres));
return utt;
}
static cst_utterance *cg_predict_params(cst_utterance *utt)
{
cst_cg_db *cg_db;
cst_track *param_track;
cst_track *str_track = NULL;
cst_item *mcep;
const cst_cart *mcep_tree, *f0_tree;
int i,j,f,p,fd,o;
const char *mname;
float f0_val;
int fff;
int extra_feats = 0;
cg_db = val_cg_db(utt_feat_val(utt,"cg_db"));
param_track = new_track();
if (cg_db->do_mlpg) /* which should be the default */
fff = 1; /* copy details with stddevs */
else
fff = 2; /* copy details without stddevs */
extra_feats = 1; /* voicing */
if (cg_db->mixed_excitation)
{
extra_feats += 5;
str_track = new_track();
cst_track_resize(str_track,
utt_feat_int(utt,"param_track_num_frames"),
5);
}
cst_track_resize(param_track,
utt_feat_int(utt,"param_track_num_frames"),
(cg_db->num_channels0/fff)-
(2 * extra_feats));/* no voicing or str */
for (i=0,mcep=utt_rel_head(utt,"mcep"); mcep; i++,mcep=item_next(mcep))
{
mname = item_feat_string(mcep,"name");
for (p=0; cg_db->types[p]; p++)
if (cst_streq(mname,cg_db->types[p]))
break;
if (cg_db->types[0] == NULL)
p=0; /* if there isn't a matching tree, use the first one */
/* Predict F0 */
f0_tree = cg_db->f0_trees[p];
f0_val = val_float(cart_interpret(mcep,f0_tree));
param_track->frames[i][0] = f0_val;
/* what about stddev ? */
if (cg_db->multimodel)
{ /* MULTI model */
f = val_int(cart_interpret(mcep,cg_db->param_trees0[p]));
fd = val_int(cart_interpret(mcep,cg_db->param_trees1[p]));
item_set_int(mcep,"clustergen_param_frame",f);
param_track->frames[i][0] =
(param_track->frames[i][0]+
CG_MODEL_VECTOR(cg_db,model_vectors0,f,0)+
CG_MODEL_VECTOR(cg_db,model_vectors1,fd,0))/3.0;
for (j=2; j<param_track->num_channels; j++)
param_track->frames[i][j] =
(CG_MODEL_VECTOR(cg_db,model_vectors0,f,(j)*fff)+
CG_MODEL_VECTOR(cg_db,model_vectors1,fd,(j)*fff))/2.0;
if (cg_db->mixed_excitation)
{
o = j;
for (j=0; j<5; j++)
{
str_track->frames[i][j] =
(CG_MODEL_VECTOR(cg_db,model_vectors0,f,(o+(2*j))*fff)+
CG_MODEL_VECTOR(cg_db,model_vectors1,fd,(o+(2*j))*fff))/2.0;
}
}
}
else
{ /* SINGLE model */
/* Predict Spectral */
mcep_tree = cg_db->param_trees0[p];
f = val_int(cart_interpret(mcep,mcep_tree));
item_set_int(mcep,"clustergen_param_frame",f);
param_track->frames[i][0] =
(param_track->frames[i][0]+
CG_MODEL_VECTOR(cg_db,model_vectors0,f,0))/2.0;
for (j=2; j<param_track->num_channels; j++)
param_track->frames[i][j] =
CG_MODEL_VECTOR(cg_db,model_vectors0,f,(j)*fff);
if (cg_db->mixed_excitation)
{
o = j;
for (j=0; j<5; j++)
{
str_track->frames[i][j] =
CG_MODEL_VECTOR(cg_db,model_vectors0,f,(o+(2*j))*fff);
}
}
}
/* last coefficient is average voicing for cluster */
item_set_float(mcep,"voicing",
CG_MODEL_VECTOR(cg_db,model_vectors0,f,
cg_db->num_channels0-2));
param_track->times[i] = i * cg_db->frame_advance;
}
cg_smooth_F0(utt,cg_db,param_track);
utt_set_feat(utt,"param_track",track_val(param_track));
if (cg_db->mixed_excitation)
utt_set_feat(utt,"str_track",track_val(str_track));
return utt;
}
