/* HTS_Engine_refresh: free model per one time synthesis */
void HTS_Engine_refresh(HTS_Engine * engine)
{
/* free generated parameter stream set */
HTS_GStreamSet_clear(&engine->gss);
/* free parameter stream set */
HTS_PStreamSet_clear(&engine->pss);
/* free state stream set */
HTS_SStreamSet_clear(&engine->sss);
}
/* HTS_Engine_refresh: free model per one time synthesis */
void HTS_Engine_refresh(HTS_Engine * engine)
{
/* free generated parameter stream set */
HTS_GStreamSet_clear(&engine->gss);
/* free parameter stream set */
HTS_PStreamSet_clear(&engine->pss);
/* free state stream set */
HTS_SStreamSet_clear(&engine->sss);
/* free label list */
HTS_Label_clear(&engine->label);
/* stop flag */
engine->condition.stop = FALSE;
}
/* HTS_Engine_synthesize_from_strings: synthesize speech from strings */
HTS_Boolean HTS_Engine_synthesize_from_strings(HTS_Engine * engine, char **lines, size_t num_lines)
{
size_t i;
double f;
HTS_Engine_refresh(engine);
// Generate state sequence
if (HTS_SStreamSet_create(&engine->sss, &engine->ms, lines, num_lines, engine->condition.speed) != TRUE) {
HTS_Engine_refresh(engine);
return FALSE;
}
if (engine->condition.additional_half_tone != 0.0) {
for (i = 0; i < HTS_SStreamSet_get_total_state(&engine->sss); i++) {
f = HTS_SStreamSet_get_mean(&engine->sss, 1, i, 0);
f += engine->condition.additional_half_tone * HALF_TONE;
if (f < MIN_LF0)
f = MIN_LF0;
else if (f > MAX_LF0)
f = MAX_LF0;
HTS_SStreamSet_set_mean(&engine->sss, 1, i, 0, f);
}
}
// Generate parameter sequence
if (HTS_PStreamSet_create(&engine->pss, &engine->sss, engine->condition.msd_threshold) != TRUE) {
HTS_Engine_refresh(engine);
return FALSE;
}
// Free state sequence
HTS_SStreamSet_clear(&engine->sss);
// Generate sound sample sequence
if (HTS_GStreamSet_create(&engine->gss, &engine->pss, engine->condition.sampling_frequency,
engine->condition.fperiod, engine->condition.alpha,
engine->condition.beta) != TRUE) {
HTS_Engine_refresh(engine);
return FALSE;
}
return TRUE;
}
/* perform conversion */
int vc(const GMM * gmm, const DELTAWINDOW * window, const size_t total_frame,
const size_t source_vlen, const size_t target_vlen,
const double *gv_mean, const double *gv_vari,
const double *source, double *target)
{
size_t t, i, j, k, max_num_mix = 0,
src_vlen_dyn = source_vlen * window->win_size,
tgt_vlen_dyn = target_vlen * window->win_size;
int m, l, shift;
double max_post_mix = 0.0, logoutp = LZERO, *input = NULL,
*src_with_dyn = NULL, *logwgd = NULL,
**cov_xx_inv = NULL, ***cov_yx_xx = NULL, *gv_weight = NULL,
***cond_mean = NULL, ***cond_vari = NULL, **cond_post_mix = NULL;
HTS_SStreamSet sss;
HTS_PStreamSet pss;
/* append dynamic feature */
src_with_dyn = dgetmem(total_frame * src_vlen_dyn);
for (t = 0; t < total_frame; t++) {
for (i = 0; i < window->win_size; i++) {
j = window->win_size * source_vlen * t + source_vlen * i;
for (shift = window->win_l_width[i];
shift <= window->win_r_width[i]; shift++) {
l = t + shift;
if (l < 0) {
l = 0;
}
if (!(l < (int) total_frame)) {
l = total_frame - 1;
}
for (k = 0; k < source_vlen; k++) {
src_with_dyn[j + k] += window->win_coefficient[i][shift]
* source[source_vlen * l + k];
}
}
}
}
/* calculate mean and covariace of conditional distribution
given source feature and mixture component */
cond_post_mix = ddgetmem(total_frame, gmm->nmix);
cond_mean = (double ***) getmem(gmm->nmix, sizeof(*(cond_mean)));
for (m = 0; m < gmm->nmix; m++) {
cond_mean[m] = ddgetmem(total_frame, tgt_vlen_dyn);
}
cond_vari = (double ***) getmem(gmm->nmix, sizeof(*(cond_vari)));
for (m = 0; m < gmm->nmix; m++) {
cond_vari[m] = ddgetmem(tgt_vlen_dyn, tgt_vlen_dyn);
}
cov_xx_inv = ddgetmem(src_vlen_dyn, src_vlen_dyn);
cov_yx_xx = (double ***) getmem(gmm->nmix, sizeof(*(cov_yx_xx)));
for (m = 0; m < gmm->nmix; m++) {
cov_yx_xx[m] = ddgetmem(tgt_vlen_dyn, src_vlen_dyn);
}
for (m = 0; m < gmm->nmix; m++) {
invert(gmm->gauss[m].cov, cov_xx_inv, src_vlen_dyn);
for (i = 0; i < tgt_vlen_dyn; i++) {
for (j = 0; j < src_vlen_dyn; j++) {
for (k = 0; k < src_vlen_dyn; k++) {
cov_yx_xx[m][i][j] += gmm->gauss[m].cov[src_vlen_dyn + i][k]
* cov_xx_inv[k][j];
}
}
}
}
logwgd = dgetmem(gmm->nmix);
input = dgetmem(src_vlen_dyn);
for (t = 0; t < total_frame; t++) {
for (i = 0; i < src_vlen_dyn; i++) {
input[i] = src_with_dyn[t * src_vlen_dyn + i];
}
for (m = 0, logoutp = LZERO; m < gmm->nmix; m++) {
logwgd[m] = log_wgd(gmm, m, src_vlen_dyn, input);
logoutp = log_add(logoutp, logwgd[m]);
}
for (m = 0; m < gmm->nmix; m++) {
/* posterior probability of mixture component given source feature */
cond_post_mix[t][m] = exp(logwgd[m] - logoutp);
for (i = 0; i < tgt_vlen_dyn; i++) {
for (j = 0; j < src_vlen_dyn; j++) {
cond_mean[m][t][i] += cov_yx_xx[m][i][j]
* (input[j] - gmm->gauss[m].mean[j]);
}
cond_mean[m][t][i] += gmm->gauss[m].mean[src_vlen_dyn + i];
}
}
}
for (m = 0; m < gmm->nmix; m++) {
for (i = 0; i < tgt_vlen_dyn; i++) {
for (j = 0; j < tgt_vlen_dyn; j++) {
for (k = 0; k < src_vlen_dyn; k++) {
cond_vari[m][i][j] += cov_yx_xx[m][i][k]
* gmm->gauss[m].cov[k][src_vlen_dyn + j];
}
cond_vari[m][i][j] =
gmm->gauss[m].cov[src_vlen_dyn + i][src_vlen_dyn + j]
- cond_vari[m][i][j];
}
}
}
/* initialize parameter set of hts_engine */
HTS_PStreamSet_initialize(&pss);
sss.nstream = 1;
sss.total_state = total_frame;
sss.total_frame = total_frame;
sss.duration = (size_t *) getmem(total_frame, sizeof(size_t));
for (i = 0; i < total_frame; i++) {
sss.duration[i] = 1;
}
sss.sstream = (HTS_SStream *) getmem(1, sizeof(HTS_SStream));
sss.sstream->vector_length = target_vlen;
sss.sstream->mean =
(double **) getmem(sss.total_state, sizeof(*(sss.sstream->mean)));
sss.sstream->vari =
(double **) getmem(sss.total_state, sizeof(*(sss.sstream->vari)));
for (i = 0; i < sss.total_state; i++) {
sss.sstream->mean[i] = dgetmem(tgt_vlen_dyn);
sss.sstream->vari[i] = dgetmem(tgt_vlen_dyn);
}
sss.sstream->msd = NULL; /* no MSD */
sss.sstream->win_size = window->win_size;
sss.sstream->win_l_width =
(int *) getmem(window->win_size, sizeof(*(sss.sstream->win_l_width)));
sss.sstream->win_r_width =
(int *) getmem(window->win_size, sizeof(*(sss.sstream->win_r_width)));
sss.sstream->win_coefficient =
(double **) getmem(window->win_size,
sizeof(*(sss.sstream->win_coefficient)));
for (i = 0; i < window->win_size; i++) {
sss.sstream->win_l_width[i] = window->win_l_width[i];
sss.sstream->win_r_width[i] = window->win_r_width[i];
if (sss.sstream->win_l_width[i] + sss.sstream->win_r_width[i] == 0) {
sss.sstream->win_coefficient[i] =
dgetmem(-2 * sss.sstream->win_l_width[i] + 1);
} else {
sss.sstream->win_coefficient[i] =
dgetmem(-2 * sss.sstream->win_l_width[i]);
}
sss.sstream->win_coefficient[i] -= sss.sstream->win_l_width[i];
for (shift = sss.sstream->win_l_width[i];
shift <= sss.sstream->win_r_width[i]; shift++) {
sss.sstream->win_coefficient[i][shift] =
window->win_coefficient[i][shift];
}
}
sss.sstream->win_max_width = window->win_max_width;
if ((gv_mean != NULL) && (gv_vari != NULL)) { /* set GV parameters */
sss.sstream->gv_mean = dgetmem(sss.sstream->vector_length);
sss.sstream->gv_vari = dgetmem(sss.sstream->vector_length);
for (i = 0; i < sss.sstream->vector_length; i++) {
sss.sstream->gv_mean[i] = gv_mean[i];
sss.sstream->gv_vari[i] = gv_vari[i];
}
} else {
sss.sstream->gv_mean = NULL;
sss.sstream->gv_vari = NULL;
}
sss.sstream->gv_switch =
(HTS_Boolean *) getmem(total_frame, sizeof(HTS_Boolean));
for (i = 0; i < total_frame; i++) {
sss.sstream->gv_switch[i] = TRUE;
}
gv_weight = dgetmem(tgt_vlen_dyn);
for (i = 0; i < tgt_vlen_dyn; i++) {
gv_weight[i] = 1.0;
}
/* initialize pdf sequence */
for (t = 0; t < total_frame; t++) {
max_post_mix = cond_post_mix[t][0];
max_num_mix = 0;
for (m = 1; m < gmm->nmix; m++) {
if (max_post_mix < cond_post_mix[t][m]) {
max_post_mix = cond_post_mix[t][m];
max_num_mix = m;
}
}
for (i = 0; i < tgt_vlen_dyn; i++) {
sss.sstream->mean[t][i] = cond_mean[max_num_mix][t][i];
sss.sstream->vari[t][i] = cond_vari[max_num_mix][i][i];
}
}
/* parameter generation by hts_engine API */
HTS_PStreamSet_create(&pss, &sss, NULL, gv_weight);
for (t = 0; t < total_frame; t++) {
k = t * target_vlen;
for (i = 0; i < target_vlen; i++) {
target[k + i] = pss.pstream->par[t][i];
}
}
/* release memory */
free(src_with_dyn);
free(input);
free(logwgd);
free(cov_xx_inv[0]);
free(cov_xx_inv);
for (m = 0; m < gmm->nmix; m++) {
free(cov_yx_xx[m][0]);
free(cov_yx_xx[m]);
free(cond_mean[m][0]);
free(cond_mean[m]);
free(cond_vari[m][0]);
free(cond_vari[m]);
}
free(cov_yx_xx);
free(cond_mean);
free(cond_vari);
free(cond_post_mix[0]);
free(cond_post_mix);
free(gv_weight);
HTS_PStreamSet_clear(&pss);
HTS_SStreamSet_clear(&sss);
return (0);
}