cst_utterance *concat_units(cst_utterance *utt)
{
cst_lpcres *target_lpcres;
cst_item *u;
int pm_i;
int unit_size, unit_start, unit_end;
int rpos, nearest_u_pm;
int target_end, target_start;
float m, u_index;
cst_sts_list *sts_list;
const char *residual_type;
sts_list = val_sts_list(utt_feat_val(utt, "sts_list"));
if (sts_list->codec == NULL)
residual_type = "ulaw";
else
residual_type = sts_list->codec;
target_lpcres = val_lpcres(utt_feat_val(utt, "target_lpcres"));
target_lpcres->lpc_min = sts_list->coeff_min;
target_lpcres->lpc_range = sts_list->coeff_range;
target_lpcres->num_channels = sts_list->num_channels;
target_lpcres->sample_rate = sts_list->sample_rate;
lpcres_resize_samples(target_lpcres,
target_lpcres->times[target_lpcres->num_frames -
1]);
if (utt_feat_val(utt, "delayed_decoding"))
{
target_lpcres->delayed_decoding = 1;
target_lpcres->packed_residuals =
cst_alloc(const unsigned char *, target_lpcres->num_frames);
}
static cst_utterance *cg_resynth(cst_utterance *utt)
{
cst_cg_db *cg_db;
cst_wave *w;
cst_track *param_track;
cst_track *str_track = NULL;
cst_track *smoothed_track;
const cst_val *streaming_info_val;
cst_audio_streaming_info *asi = NULL;
streaming_info_val=get_param_val(utt->features,"streaming_info",NULL);
if (streaming_info_val)
asi = val_audio_streaming_info(streaming_info_val);
cg_db = val_cg_db(utt_feat_val(utt,"cg_db"));
param_track = val_track(utt_feat_val(utt,"param_track"));
if (cg_db->mixed_excitation)
str_track = val_track(utt_feat_val(utt,"str_track"));
if (cg_db->do_mlpg)
{
smoothed_track = mlpg(param_track, cg_db);
w = mlsa_resynthesis(smoothed_track,str_track,cg_db,asi);
delete_track(smoothed_track);
}
else
w=mlsa_resynthesis(param_track,str_track,cg_db,asi);
utt_set_wave(utt,w);
return utt;
}
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;
}
cst_utterance *join_units_simple(cst_utterance *utt)
{
cst_wave *w = 0;
cst_lpcres *lpcres;
const char *resynth_type;
const cst_val *streaming_info_val;
resynth_type = get_param_string(utt->features, "resynth_type", "fixed");
asis_to_pm(utt);
concat_units(utt);
lpcres = val_lpcres(utt_feat_val(utt, "target_lpcres"));
streaming_info_val = get_param_val(utt->features, "streaming_info", NULL);
if (streaming_info_val)
{
lpcres->asi = val_audio_streaming_info(streaming_info_val);
lpcres->asi->utt = utt;
}
if (cst_streq(resynth_type, "fixed"))
w = lpc_resynth_fixedpoint(lpcres);
else
{
cst_errmsg("unknown resynthesis type %s\n", resynth_type);
cst_error(); /* Should not happen */
}
utt_set_wave(utt, w);
return utt;
}
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 *cg_synth(cst_utterance *utt)
{
cst_cg_db *cg_db;
cg_db = val_cg_db(utt_feat_val(utt,"cg_db"));
cg_make_hmmstates(utt);
cg_make_params(utt);
cg_predict_params(utt);
if (cg_db->spamf0)
{
cst_spamf0(utt);
}
cg_resynth(utt);
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;
}
cst_utterance *join_units_modified_lpc(cst_utterance *utt)
{
cst_wave *w = 0;
cst_lpcres *lpcres;
const char *resynth_type;
const cst_val *streaming_info_val;
resynth_type = get_param_string(utt->features, "resynth_type", "float");
f0_targets_to_pm(utt);
concat_units(utt);
lpcres = val_lpcres(utt_feat_val(utt, "target_lpcres"));
streaming_info_val = get_param_val(utt->features, "streaming_info", NULL);
if (streaming_info_val)
{
lpcres->asi = val_audio_streaming_info(streaming_info_val);
lpcres->asi->utt = utt;
}
if (cst_streq(resynth_type, "float"))
w = lpc_resynth(lpcres);
else if (cst_streq(resynth_type, "fixed"))
{
w = lpc_resynth_fixedpoint(lpcres);
}
else
{
cst_errmsg("unknown resynthesis type %s\n", resynth_type);
cst_error(); /* Should not happen */
}
if (w == NULL)
{
/* Synthesis Failed, probably because it was interrupted */
utt_set_feat_int(utt, "Interrupted", 1);
w = new_wave();
}
utt_set_wave(utt, w);
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 *tokentosegs(cst_utterance *u)
{
cst_item *t;
cst_relation *seg, *syl, *sylstructure, *word;
cst_item *sylitem, *sylstructureitem, *worditem, *sssyl;
cst_phoneset *ps;
ps = val_phoneset(utt_feat_val(u, "phoneset"));
/* Just copy tokens into the Segment relation */
seg = utt_relation_create(u, "Segment");
syl = utt_relation_create(u, "Syllable");
word = utt_relation_create(u, "Word");
sylstructure = utt_relation_create(u, "SylStructure");
sssyl = sylitem = worditem = sylstructureitem = 0;
for (t = relation_head(utt_relation(u, "Token")); t; t = item_next(t))
{
cst_item *segitem = relation_append(seg, NULL);
char const *pname = item_feat_string(t, "name");
char *name = cst_strdup(pname);
if (worditem == 0)
{
worditem = relation_append(word,NULL);
item_set_string(worditem, "name", "phonestring");
sylstructureitem = relation_append(sylstructure,worditem);
}
if (sylitem == 0)
{
sylitem = relation_append(syl,NULL);
sssyl = item_add_daughter(sylstructureitem,sylitem);
}
if (name[cst_strlen(name)-1] == '1')
{
item_set_string(sssyl,"stress","1");
name[cst_strlen(name)-1] = '\0';
}
else if (name[cst_strlen(name)-1] == '0')
{
item_set_string(sssyl,"stress","0");
name[cst_strlen(name)-1] = '\0';
}
if (cst_streq(name,"-"))
{
sylitem = 0; /* syllable break */
}
else if (phone_id(ps, name) == -1)
{
cst_errmsg("Phone `%s' not in phoneset\n", pname);
cst_error();
}
else
{
item_add_daughter(sssyl,segitem);
item_set_string(segitem, "name", name);
}
cst_free(name);
}
return u;
}
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;
}
