cst_utterance *cst_spamf0(cst_utterance *utt)
{
cst_track *spamf0_track = NULL;
cst_track *param_track = NULL;
cst_item *s;
cst_cg_db *cg_db;
const cst_cart *acc_tree, *phrase_tree;
float end, f0val, syldur;
int num_frames, f, i;
cg_db = val_cg_db(utt_feat_val(utt, "cg_db"));
spamf0_track = new_track();
cst_track_resize(spamf0_track,
(utt_feat_int(utt, "param_track_num_frames")), 1);
acc_tree = cg_db->spamf0_accent_tree;
phrase_tree = cg_db->spamf0_phrase_tree;
end = 0.0;
num_frames = 0;
for (s = utt_rel_head(utt, "Segment"); s; s = item_next(s))
{
end = ffeature_float(s, "end");
if (cst_streq("pau", ffeature_string(s, "name")))
{
f0val = 0;
}
else
{
f0val = val_float(cart_interpret(s, phrase_tree));
}
for (;
((num_frames * cg_db->frame_advance) <= end)
&& (num_frames < utt_feat_int(utt, "param_track_num_frames"));
num_frames++)
{
spamf0_track->frames[num_frames][0] = f0val;
}
}
for (s = utt_rel_head(utt, "Syllable"); s; s = item_next(s))
{
f = val_int(cart_interpret(s, acc_tree));
syldur = ffeature_float(s, "R:SylStructure.daughtern.R:Segment.end")
- ffeature_float(s, "R:SylStructure.daughter1.R:Segment.p.end");
cst_synthtilt(cg_db,
ffeature_float(s,
"R:SylStructure.daughter1.R:Segment.p.end"),
cg_db->spamf0_accent_vectors[f][0],
cg_db->spamf0_accent_vectors[f][2], syldur,
cg_db->spamf0_accent_vectors[f][6], spamf0_track);
}
param_track = val_track(utt_feat_val(utt, "param_track"));
for (i = 0; i < utt_feat_int(utt, "param_track_num_frames"); i++)
{
param_track->frames[i][0] = spamf0_track->frames[i][0];
}
delete_track(spamf0_track);
return utt;
}
static cst_utterance *cg_make_hmmstates(cst_utterance *utt)
{
/* Build HMM state structure below the segment structure */
cst_cg_db *cg_db;
cst_relation *hmmstate, *segstate;
cst_item *seg, *s, *ss;
const char *segname;
int sp,p;
cg_db = val_cg_db(utt_feat_val(utt,"cg_db"));
hmmstate = utt_relation_create(utt,"HMMstate");
segstate = utt_relation_create(utt,"segstate");
for (seg = utt_rel_head(utt,"Segment"); seg; seg=item_next(seg))
{
ss = relation_append(segstate,seg);
segname = item_feat_string(seg,"name");
for (p=0; cg_db->phone_states[p]; p++)
if (cst_streq(segname,cg_db->phone_states[p][0]))
break;
if (cg_db->phone_states[p] == NULL)
p = 0; /* unknown phoneme */
for (sp=1; cg_db->phone_states[p][sp]; sp++)
{
s = relation_append(hmmstate,NULL);
item_add_daughter(ss,s);
item_set_string(s,"name",cg_db->phone_states[p][sp]);
item_set_int(s,"statepos",sp);
}
}
return utt;
}
static cst_utterance *cg_make_params(cst_utterance *utt)
{
/* puts in the frame items */
/* historically called "mcep" but can actually be any random vectors */
cst_cg_db *cg_db;
cst_relation *mcep, *mcep_link;
cst_item *s, *mcep_parent, *mcep_frame;
int num_frames;
float start, end;
float dur_stretch, tok_stretch;
cg_db = val_cg_db(utt_feat_val(utt,"cg_db"));
mcep = utt_relation_create(utt,"mcep");
mcep_link = utt_relation_create(utt,"mcep_link");
end = 0.0;
num_frames = 0;
dur_stretch = get_param_float(utt->features,"duration_stretch", 1.0);
for (s = utt_rel_head(utt,"HMMstate"); s; s=item_next(s))
{
start = end;
tok_stretch = ffeature_float(s,"R:segstate.parent.R:SylStructure.parent.parent.R:Token.parent.local_duration_stretch");
if (tok_stretch == 0)
tok_stretch = 1.0;
end = start + (tok_stretch*dur_stretch*cg_state_duration(s,cg_db));
item_set_float(s,"end",end);
mcep_parent = relation_append(mcep_link, s);
for ( ; (num_frames * cg_db->frame_advance) <= end; num_frames++ )
{
mcep_frame = relation_append(mcep,NULL);
item_add_daughter(mcep_parent,mcep_frame);
item_set_int(mcep_frame,"frame_number",num_frames);
item_set(mcep_frame,"name",item_feat(mcep_parent,"name"));
}
}
/* Copy duration up onto Segment relation */
for (s = utt_rel_head(utt,"Segment"); s; s=item_next(s))
item_set(s,"end",ffeature(s,"R:segstate.daughtern.end"));
utt_set_feat_int(utt,"param_track_num_frames",num_frames);
return utt;
}
static void cg_smooth_F0(cst_utterance *utt,cst_cg_db *cg_db,
cst_track *param_track)
{
/* Smooth F0 and mark unnoived frames as 0.0 */
cst_item *mcep;
int i, c;
float l, s;
float mean, stddev;
l = 0.0;
for (i=0; i<param_track->num_frames-1; i++)
{
c = 0; s = 0;
if (l > 0.0)
{
c++; s+=l;
}
if (param_track->frames[i+1][0] > 0.0)
{
c++; s+=param_track->frames[i+1][0];
}
l = param_track->frames[i][0];
if (param_track->frames[i][0] > 0.0)
{
c++; s+=param_track->frames[i][0];
param_track->frames[i][0] = s/c;
}
}
mean = get_param_float(utt->features,"int_f0_target_mean", cg_db->f0_mean);
mean *= get_param_float(utt->features,"f0_shift", 1.0);
stddev =
get_param_float(utt->features,"int_f0_target_stddev", cg_db->f0_stddev);
for (i=0,mcep=utt_rel_head(utt,"mcep"); mcep; i++,mcep=item_next(mcep))
{
if (voiced_frame(mcep))
{
/* scale the F0 -- which normally wont change it at all */
param_track->frames[i][0] =
(((param_track->frames[i][0]-cg_db->f0_mean)/cg_db->f0_stddev)
*stddev)+mean;
/* Some safety checks */
if (param_track->frames[i][0] < 50)
param_track->frames[i][0] = 50;
if (param_track->frames[i][0] > 700)
param_track->frames[i][0] = 700;
}
else /* Unvoice it */
param_track->frames[i][0] = 0.0;
}
return;
}
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;
}
