cst_utterance *default_tokenization(cst_utterance *u)
{
const char *text,*token;
cst_tokenstream *fd;
cst_item *t;
cst_relation *r;
text = utt_input_text(u);
r = utt_relation_create(u,"Token");
fd = ts_open_string(text,
get_param_string(u->features,"text_whitespace",NULL),
get_param_string(u->features,"text_singlecharsymbols",NULL),
get_param_string(u->features,"text_prepunctuation",NULL),
get_param_string(u->features,"text_postpunctuation",NULL));
while(!ts_eof(fd))
{
token = ts_get(fd);
if (cst_strlen(token) > 0)
{
t = relation_append(r,NULL);
item_set_string(t,"name",token);
item_set_string(t,"whitespace",fd->whitespace);
item_set_string(t,"prepunctuation",fd->prepunctuation);
item_set_string(t,"punc",fd->postpunctuation);
item_set_int(t,"file_pos",fd->file_pos);
item_set_int(t,"line_number",fd->line_number);
}
}
ts_close(fd);
return u;
}
static int bbb_relation_load(cst_relation *r,const char *filename)
{
const char *token;
cst_item *item;
cst_tokenstream *fd;
fd = ts_open(filename);
if (fd == 0)
return 0;
while (!ts_eof(fd))
{
token = ts_get(fd);
if (cst_streq(token,""))
continue;
item = relation_append(r,NULL);
item_set_string(item,"name",token);
item_set_string(item,"whitespace",fd->whitespace);
item_set_string(item,"prepunctuation",fd->prepunctuation);
item_set_string(item,"punc",fd->postpunctuation);
item_set_int(item,"file_pos",fd->file_pos);
item_set_int(item,"line_number",fd->line_number);
}
ts_close(fd);
return 1;
}
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_val* word_to_phones(const cst_item *word)
{
cst_val*phones=NULL;
const char *name=item_feat_string(word, "name");
ustring32_t letters=ustring32_alloc(0);
if(letters==NULL) return NULL;
ustring32_assign8(letters,(const uint8_t*)name);
if(ustring32_empty(letters))
{
ustring32_free(letters);
return NULL;
}
unsigned int flags=classify_characters(ustring32_str(letters),ustring32_length(letters));
int variant=item_feat_int(item_parent(item_as(word,"Token")),"variant");
if((flags&cs_lc)&&cst_streq(ffeature_string(word,"gpos"),"content"))
{
if(variant==variant_pseudo_english)
phones=ustring32_lts_apply(letters,&en_consonants_lts);
else phones=ustring32_lts_apply(letters,&ru_consonants_lts);
item_set_int(word,"no_vr",1);
}
else if((variant==variant_pseudo_english)&&(flags&cs_en))
{
cst_val *en_phones=lex_lookup(en_lex,name,(cst_streq(name,"a")?"n":NULL));
if(en_phones)
{
phones=ru_lts_apply(en_phones,&ru_en_lts);
delete_val(en_phones);
}
item_set_int(word,"no_pl",1);
}
else
{
const ru_dict_entry *e=bsearch(name,ru_dict,ru_dict_size,sizeof(ru_dict_entry),compare_entries);
if(e!=NULL)
{
if(e->stress > 0)
ustring32_set(letters,e->stress-1,1105);
else
item_set_int(word,"stressed_syl_num",e->stress);
}
phones=ustring32_lts_apply(letters,&ru_lts);
}
ustring32_free(letters);
return phones;
}
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 *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;
}
float flite_file_to_speech(const char *filename,
cst_voice *voice,
const char *outtype)
{
cst_utterance *utt;
cst_tokenstream *ts;
const char *token;
cst_item *t;
cst_relation *tokrel;
float d, durs = 0;
int num_tokens;
cst_breakfunc breakfunc = default_utt_break;
if ((ts = ts_open(filename,
get_param_string(voice->features,"text_whitespace",NULL),
get_param_string(voice->features,"text_singlecharsymbols",NULL),
get_param_string(voice->features,"text_prepunctuation",NULL),
get_param_string(voice->features,"text_postpunctuation",NULL)))
== NULL)
{
cst_errmsg("failed to open file \"%s\" for reading\n",
filename);
return 1;
}
if (feat_present(voice->features,"utt_break"))
breakfunc = val_breakfunc(feat_val(voice->features,"utt_break"));
/* If its a file to write to delete it as we're going to */
/* incrementally append to it */
if (!cst_streq(outtype,"play") && !cst_streq(outtype,"none"))
{
cst_wave *w;
w = new_wave();
cst_wave_resize(w,0,1);
cst_wave_set_sample_rate(w,16000);
cst_wave_save_riff(w,outtype); /* an empty wave */
delete_wave(w);
}
num_tokens = 0;
utt = new_utterance();
tokrel = utt_relation_create(utt, "Token");
while (!ts_eof(ts) || num_tokens > 0)
{
token = ts_get(ts);
if ((strlen(token) == 0) ||
(num_tokens > 500) || /* need an upper bound */
(relation_head(tokrel) &&
breakfunc(ts,token,tokrel)))
{
/* An end of utt */
d = flite_tokens_to_speech(utt,voice,outtype);
utt = NULL;
if (d < 0)
goto out;
durs += d;
if (ts_eof(ts))
goto out;
utt = new_utterance();
tokrel = utt_relation_create(utt, "Token");
num_tokens = 0;
}
num_tokens++;
t = relation_append(tokrel, NULL);
item_set_string(t,"name",token);
item_set_string(t,"whitespace",ts->whitespace);
item_set_string(t,"prepunctuation",ts->prepunctuation);
item_set_string(t,"punc",ts->postpunctuation);
item_set_int(t,"file_pos",ts->file_pos);
item_set_int(t,"line_number",ts->line_number);
}
out:
delete_utterance(utt);
ts_close(ts);
return durs;
}
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;
}
static float flite_ssml_to_speech_ts(cst_tokenstream *ts,
cst_voice *voice,
const char *outtype)
{
cst_features *ssml_feats, *ssml_word_feats;
cst_features *attributes;
const char *token;
char *tag;
cst_utterance *utt;
cst_relation *tokrel;
int num_tokens;
cst_breakfunc breakfunc = default_utt_break;
cst_uttfunc utt_user_callback = 0;
float durs = 0.0;
cst_item *t;
ssml_feats = new_features();
ssml_word_feats = new_features();
set_charclasses(ts,
" \t\n\r",
ssml_singlecharsymbols_general,
get_param_string(voice->features,"text_prepunctuation",""),
get_param_string(voice->features,"text_postpunctuation","")
);
if (feat_present(voice->features,"utt_break"))
breakfunc = val_breakfunc(feat_val(voice->features,"utt_break"));
if (feat_present(voice->features,"utt_user_callback"))
utt_user_callback = val_uttfunc(feat_val(voice->features,"utt_user_callback"));
num_tokens = 0;
utt = new_utterance();
tokrel = utt_relation_create(utt, "Token");
while (!ts_eof(ts) || num_tokens > 0)
{
token = ts_get(ts);
if (cst_streq("<",token))
{ /* A tag */
tag = cst_upcase(ts_get(ts));
if (cst_streq("/",tag)) /* an end tag */
{
tag = cst_upcase(ts_get(ts));
attributes = ssml_get_attributes(ts);
feat_set_string(attributes,"_type","end");
}
else
attributes = ssml_get_attributes(ts);
utt = ssml_apply_tag(tag,attributes,utt,ssml_word_feats);
cst_free(tag);
}
else if (cst_streq("&",token))
{ /* an escape sequence */
/* skip to ; and insert value in rawdata */
}
else
{
if ((cst_strlen(token) == 0) ||
(num_tokens > 500) || /* need an upper bound */
(relation_head(tokrel) &&
breakfunc(ts,token,tokrel)))
{
/* An end of utt, so synthesize it */
if (utt_user_callback)
utt = (utt_user_callback)(utt);
if (utt)
{
utt = flite_do_synth(utt,voice,utt_synth_tokens);
durs += flite_process_output(utt,outtype,TRUE);
delete_utterance(utt); utt = NULL;
}
else
break;
if (ts_eof(ts)) break;
utt = new_utterance();
tokrel = utt_relation_create(utt, "Token");
num_tokens = 0;
}
num_tokens++;
t = relation_append(tokrel, NULL);
item_set_string(t,"name",token);
item_set_string(t,"whitespace",ts->whitespace);
item_set_string(t,"prepunctuation",ts->prepunctuation);
item_set_string(t,"punc",ts->postpunctuation);
/* Mark it at the beginning of the token */
item_set_int(t,"file_pos",
ts->file_pos-(1+ /* as we are already on the next char */
cst_strlen(token)+
cst_strlen(ts->prepunctuation)+
cst_strlen(ts->postpunctuation)));
item_set_int(t,"line_number",ts->line_number);
}
}
delete_utterance(utt);
return durs;
}
float flite_file_to_speech(const char *filename,
cst_voice *voice,
const char *outtype)
{
cst_utterance *utt;
cst_tokenstream *ts;
const char *token;
cst_item *t;
cst_relation *tokrel;
float durs = 0;
int num_tokens;
cst_wave *w;
cst_breakfunc breakfunc = default_utt_break;
cst_uttfunc utt_user_callback = 0;
int fp;
if ((ts = ts_open(filename,
get_param_string(voice->features,"text_whitespace",NULL),
get_param_string(voice->features,"text_singlecharsymbols",NULL),
get_param_string(voice->features,"text_prepunctuation",NULL),
get_param_string(voice->features,"text_postpunctuation",NULL)))
== NULL)
{
cst_errmsg("failed to open file \"%s\" for reading\n",
filename);
return 1;
}
fp = get_param_int(voice->features,"file_start_position",0);
if (fp > 0)
ts_set_stream_pos(ts,fp);
if (feat_present(voice->features,"utt_break"))
breakfunc = val_breakfunc(feat_val(voice->features,"utt_break"));
if (feat_present(voice->features,"utt_user_callback"))
utt_user_callback = val_uttfunc(feat_val(voice->features,"utt_user_callback"));
/* If its a file to write to, create and save an empty wave file */
/* as we are going to incrementally append to it */
if (!cst_streq(outtype,"play") &&
!cst_streq(outtype,"none") &&
!cst_streq(outtype,"stream"))
{
w = new_wave();
cst_wave_resize(w,0,1);
cst_wave_set_sample_rate(w,16000);
cst_wave_save_riff(w,outtype); /* an empty wave */
delete_wave(w);
}
num_tokens = 0;
utt = new_utterance();
tokrel = utt_relation_create(utt, "Token");
while (!ts_eof(ts) || num_tokens > 0)
{
token = ts_get(ts);
if ((cst_strlen(token) == 0) ||
(num_tokens > 500) || /* need an upper bound */
(relation_head(tokrel) &&
breakfunc(ts,token,tokrel)))
{
/* An end of utt, so synthesize it */
if (utt_user_callback)
utt = (utt_user_callback)(utt);
if (utt)
{
utt = flite_do_synth(utt,voice,utt_synth_tokens);
durs += flite_process_output(utt,outtype,TRUE);
delete_utterance(utt); utt = NULL;
}
else
break;
if (ts_eof(ts)) break;
utt = new_utterance();
tokrel = utt_relation_create(utt, "Token");
num_tokens = 0;
}
num_tokens++;
t = relation_append(tokrel, NULL);
item_set_string(t,"name",token);
item_set_string(t,"whitespace",ts->whitespace);
item_set_string(t,"prepunctuation",ts->prepunctuation);
item_set_string(t,"punc",ts->postpunctuation);
/* Mark it at the beginning of the token */
item_set_int(t,"file_pos",
ts->file_pos-(1+ /* as we are already on the next char */
cst_strlen(token)+
cst_strlen(ts->prepunctuation)+
cst_strlen(ts->postpunctuation)));
item_set_int(t,"line_number",ts->line_number);
}
delete_utterance(utt);
ts_close(ts);
return durs;
}
