fly/flightgear
1
0
Fork 0
Code Activity
flightgear/3rdparty/hts_engine_API/lib/HTS_pstream.c
James Turner ebc10ee44f Update to newer version of HTS engine 
- version 1.1, released 25th December 2015

(note there is no corresponding update to Flite)
2015-12-30 17:05:25 -06:00

524 lines
21 KiB
C
Raw Blame History

/* ----------------------------------------------------------------- */
/* The HMM-Based Speech Synthesis Engine "hts_engine API" */
/* developed by HTS Working Group */
/* http://hts-engine.sourceforge.net/ */
/* ----------------------------------------------------------------- */
/* */
/* Copyright (c) 2001-2015 Nagoya Institute of Technology */
/* Department of Computer Science */
/* */
/* 2001-2008 Tokyo Institute of Technology */
/* Interdisciplinary Graduate School of */
/* Science and Engineering */
/* */
/* All rights reserved. */
/* */
/* Redistribution and use in source and binary forms, with or */
/* without modification, are permitted provided that the following */
/* conditions are met: */
/* */
/* - Redistributions of source code must retain the above copyright */
/* notice, this list of conditions and the following disclaimer. */
/* - Redistributions in binary form must reproduce the above */
/* copyright notice, this list of conditions and the following */
/* disclaimer in the documentation and/or other materials provided */
/* with the distribution. */
/* - Neither the name of the HTS working group nor the names of its */
/* contributors may be used to endorse or promote products derived */
/* from this software without specific prior written permission. */
/* */
/* THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND */
/* CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, */
/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
/* DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS */
/* BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, */
/* EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED */
/* TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, */
/* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON */
/* ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, */
/* OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY */
/* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
/* POSSIBILITY OF SUCH DAMAGE. */
/* ----------------------------------------------------------------- */
#ifndef HTS_PSTREAM_C
#define HTS_PSTREAM_C
#ifdef __cplusplus
#define HTS_PSTREAM_C_START extern "C" {
#define HTS_PSTREAM_C_END }
#else
#define HTS_PSTREAM_C_START
#define HTS_PSTREAM_C_END
#endif /* __CPLUSPLUS */
HTS_PSTREAM_C_START;
#include <math.h> /* for sqrt() */
/* hts_engine libraries */
#include "HTS_hidden.h"
/* HTS_finv: calculate 1.0/variance function */
static double HTS_finv(const double x)
{
if (x >= INFTY2)
return 0.0;
if (x <= -INFTY2)
return 0.0;
if (x <= INVINF2 && x >= 0)
return INFTY;
if (x >= -INVINF2 && x < 0)
return -INFTY;
return (1.0 / x);
}
/* HTS_PStream_calc_wuw_and_wum: calcurate W'U^{-1}W and W'U^{-1}M */
static void HTS_PStream_calc_wuw_and_wum(HTS_PStream * pst, size_t m)
{
size_t t, i, j;
int shift;
double wu;
for (t = 0; t < pst->length; t++) {
/* initialize */
pst->sm.wum[t] = 0.0;
for (i = 0; i < pst->width; i++)
pst->sm.wuw[t][i] = 0.0;
/* calc WUW & WUM */
for (i = 0; i < pst->win_size; i++)
for (shift = pst->win_l_width[i]; shift <= pst->win_r_width[i]; shift++)
if (((int) t + shift >= 0) && ((int) t + shift < pst->length) && (pst->win_coefficient[i][-shift] != 0.0)) {
wu = pst->win_coefficient[i][-shift] * pst->sm.ivar[t + shift][i * pst->vector_length + m];
pst->sm.wum[t] += wu * pst->sm.mean[t + shift][i * pst->vector_length + m];
for (j = 0; (j < pst->width) && (t + j < pst->length); j++)
if (((int) j <= pst->win_r_width[i] + shift) && (pst->win_coefficient[i][j - shift] != 0.0))
pst->sm.wuw[t][j] += wu * pst->win_coefficient[i][j - shift];
}
}
}
/* HTS_PStream_ldl_factorization: Factorize W'*U^{-1}*W to L*D*L' (L: lower triangular, D: diagonal) */
static void HTS_PStream_ldl_factorization(HTS_PStream * pst)
{
size_t t, i, j;
for (t = 0; t < pst->length; t++) {
for (i = 1; (i < pst->width) && (t >= i); i++)
pst->sm.wuw[t][0] -= pst->sm.wuw[t - i][i] * pst->sm.wuw[t - i][i] * pst->sm.wuw[t - i][0];
for (i = 1; i < pst->width; i++) {
for (j = 1; (i + j < pst->width) && (t >= j); j++)
pst->sm.wuw[t][i] -= pst->sm.wuw[t - j][j] * pst->sm.wuw[t - j][i + j] * pst->sm.wuw[t - j][0];
pst->sm.wuw[t][i] /= pst->sm.wuw[t][0];
}
}
}
/* HTS_PStream_forward_substitution: forward subtitution for mlpg */
static void HTS_PStream_forward_substitution(HTS_PStream * pst)
{
size_t t, i;
for (t = 0; t < pst->length; t++) {
pst->sm.g[t] = pst->sm.wum[t];
for (i = 1; (i < pst->width) && (t >= i); i++)
pst->sm.g[t] -= pst->sm.wuw[t - i][i] * pst->sm.g[t - i];
}
}
/* HTS_PStream_backward_substitution: backward subtitution for mlpg */
static void HTS_PStream_backward_substitution(HTS_PStream * pst, size_t m)
{
size_t rev, t, i;
for (rev = 0; rev < pst->length; rev++) {
t = pst->length - 1 - rev;
pst->par[t][m] = pst->sm.g[t] / pst->sm.wuw[t][0];
for (i = 1; (i < pst->width) && (t + i < pst->length); i++)
pst->par[t][m] -= pst->sm.wuw[t][i] * pst->par[t + i][m];
}
}
/* HTS_PStream_calc_gv: subfunction for mlpg using GV */
static void HTS_PStream_calc_gv(HTS_PStream * pst, size_t m, double *mean, double *vari)
{
size_t t;
*mean = 0.0;
for (t = 0; t < pst->length; t++)
if (pst->gv_switch[t])
*mean += pst->par[t][m];
*mean /= pst->gv_length;
*vari = 0.0;
for (t = 0; t < pst->length; t++)
if (pst->gv_switch[t])
*vari += (pst->par[t][m] - *mean) * (pst->par[t][m] - *mean);
*vari /= pst->gv_length;
}
/* HTS_PStream_conv_gv: subfunction for mlpg using GV */
static void HTS_PStream_conv_gv(HTS_PStream * pst, size_t m)
{
size_t t;
double ratio;
double mean;
double vari;
HTS_PStream_calc_gv(pst, m, &mean, &vari);
ratio = sqrt(pst->gv_mean[m] / vari);
for (t = 0; t < pst->length; t++)
if (pst->gv_switch[t])
pst->par[t][m] = ratio * (pst->par[t][m] - mean) + mean;
}
/* HTS_PStream_calc_derivative: subfunction for mlpg using GV */
static double HTS_PStream_calc_derivative(HTS_PStream * pst, size_t m)
{
size_t t, i;
double mean;
double vari;
double dv;
double h;
double gvobj;
double hmmobj;
double w = 1.0 / (pst->win_size * pst->length);
HTS_PStream_calc_gv(pst, m, &mean, &vari);
gvobj = -0.5 * W2 * vari * pst->gv_vari[m] * (vari - 2.0 * pst->gv_mean[m]);
dv = -2.0 * pst->gv_vari[m] * (vari - pst->gv_mean[m]) / pst->length;
for (t = 0; t < pst->length; t++) {
pst->sm.g[t] = pst->sm.wuw[t][0] * pst->par[t][m];
for (i = 1; i < pst->width; i++) {
if (t + i < pst->length)
pst->sm.g[t] += pst->sm.wuw[t][i] * pst->par[t + i][m];
if (t + 1 > i)
pst->sm.g[t] += pst->sm.wuw[t - i][i] * pst->par[t - i][m];
}
}
for (t = 0, hmmobj = 0.0; t < pst->length; t++) {
hmmobj += W1 * w * pst->par[t][m] * (pst->sm.wum[t] - 0.5 * pst->sm.g[t]);
h = -W1 * w * pst->sm.wuw[t][1 - 1] - W2 * 2.0 / (pst->length * pst->length) * ((pst->length - 1) * pst->gv_vari[m] * (vari - pst->gv_mean[m]) + 2.0 * pst->gv_vari[m] * (pst->par[t][m] - mean) * (pst->par[t][m] - mean));
if (pst->gv_switch[t])
pst->sm.g[t] = 1.0 / h * (W1 * w * (-pst->sm.g[t] + pst->sm.wum[t]) + W2 * dv * (pst->par[t][m] - mean));
else
pst->sm.g[t] = 1.0 / h * (W1 * w * (-pst->sm.g[t] + pst->sm.wum[t]));
}
return (-(hmmobj + gvobj));
}
/* HTS_PStream_gv_parmgen: function for mlpg using GV */
static void HTS_PStream_gv_parmgen(HTS_PStream * pst, size_t m)
{
size_t t, i;
double step = STEPINIT;
double prev = 0.0;
double obj;
if (pst->gv_length == 0)
return;
HTS_PStream_conv_gv(pst, m);
if (GV_MAX_ITERATION > 0) {
HTS_PStream_calc_wuw_and_wum(pst, m);
for (i = 1; i <= GV_MAX_ITERATION; i++) {
obj = HTS_PStream_calc_derivative(pst, m);
if (i > 1) {
if (obj > prev)
step *= STEPDEC;
if (obj < prev)
step *= STEPINC;
}
for (t = 0; t < pst->length; t++) {
if (pst->gv_switch[t])
pst->par[t][m] += step * pst->sm.g[t];
}
prev = obj;
}
}
}
/* HTS_PStream_mlpg: generate sequence of speech parameter vector maximizing its output probability for given pdf sequence */
static void HTS_PStream_mlpg(HTS_PStream * pst)
{
size_t m;
if (pst->length == 0)
return;
for (m = 0; m < pst->vector_length; m++) {
HTS_PStream_calc_wuw_and_wum(pst, m);
HTS_PStream_ldl_factorization(pst); /* LDL factorization */
HTS_PStream_forward_substitution(pst); /* forward substitution */
HTS_PStream_backward_substitution(pst, m); /* backward substitution */
if (pst->gv_length > 0)
HTS_PStream_gv_parmgen(pst, m);
}
}
/* HTS_PStreamSet_initialize: initialize parameter stream set */
void HTS_PStreamSet_initialize(HTS_PStreamSet * pss)
{
pss->pstream = NULL;
pss->nstream = 0;
pss->total_frame = 0;
}
/* HTS_PStreamSet_create: parameter generation using GV weight */
HTS_Boolean HTS_PStreamSet_create(HTS_PStreamSet * pss, HTS_SStreamSet * sss, double *msd_threshold, double *gv_weight)
{
size_t i, j, k, l, m;
int shift;
size_t frame, msd_frame, state;
HTS_PStream *pst;
HTS_Boolean not_bound;
if (pss->nstream != 0) {
HTS_error(1, "HTS_PstreamSet_create: HTS_PStreamSet should be clear.\n");
return FALSE;
}
/* initialize */
pss->nstream = HTS_SStreamSet_get_nstream(sss);
pss->pstream = (HTS_PStream *) HTS_calloc(pss->nstream, sizeof(HTS_PStream));
pss->total_frame = HTS_SStreamSet_get_total_frame(sss);
/* create */
for (i = 0; i < pss->nstream; i++) {
pst = &pss->pstream[i];
if (HTS_SStreamSet_is_msd(sss, i) == TRUE) { /* for MSD */
pst->length = 0;
for (state = 0; state < HTS_SStreamSet_get_total_state(sss); state++)
if (HTS_SStreamSet_get_msd(sss, i, state) > msd_threshold[i])
pst->length += HTS_SStreamSet_get_duration(sss, state);
pst->msd_flag = (HTS_Boolean *) HTS_calloc(pss->total_frame, sizeof(HTS_Boolean));
for (state = 0, frame = 0; state < HTS_SStreamSet_get_total_state(sss); state++) {
if (HTS_SStreamSet_get_msd(sss, i, state) > msd_threshold[i]) {
for (j = 0; j < HTS_SStreamSet_get_duration(sss, state); j++) {
pst->msd_flag[frame] = TRUE;
frame++;
}
} else {
for (j = 0; j < HTS_SStreamSet_get_duration(sss, state); j++) {
pst->msd_flag[frame] = FALSE;
frame++;
}
}
}
} else { /* for non MSD */
pst->length = pss->total_frame;
pst->msd_flag = NULL;
}
pst->vector_length = HTS_SStreamSet_get_vector_length(sss, i);
pst->width = HTS_SStreamSet_get_window_max_width(sss, i) * 2 + 1; /* band width of R */
pst->win_size = HTS_SStreamSet_get_window_size(sss, i);
if (pst->length > 0) {
pst->sm.mean = HTS_alloc_matrix(pst->length, pst->vector_length * pst->win_size);
pst->sm.ivar = HTS_alloc_matrix(pst->length, pst->vector_length * pst->win_size);
pst->sm.wum = (double *) HTS_calloc(pst->length, sizeof(double));
pst->sm.wuw = HTS_alloc_matrix(pst->length, pst->width);
pst->sm.g = (double *) HTS_calloc(pst->length, sizeof(double));
pst->par = HTS_alloc_matrix(pst->length, pst->vector_length);
}
/* copy dynamic window */
pst->win_l_width = (int *) HTS_calloc(pst->win_size, sizeof(int));
pst->win_r_width = (int *) HTS_calloc(pst->win_size, sizeof(int));
pst->win_coefficient = (double **) HTS_calloc(pst->win_size, sizeof(double));
for (j = 0; j < pst->win_size; j++) {
pst->win_l_width[j] = HTS_SStreamSet_get_window_left_width(sss, i, j);
pst->win_r_width[j] = HTS_SStreamSet_get_window_right_width(sss, i, j);
if (pst->win_l_width[j] + pst->win_r_width[j] == 0)
pst->win_coefficient[j] = (double *)
HTS_calloc(-2 * pst->win_l_width[j] + 1, sizeof(double));
else
pst->win_coefficient[j] = (double *)
HTS_calloc(-2 * pst->win_l_width[j], sizeof(double));
pst->win_coefficient[j] -= pst->win_l_width[j];
for (shift = pst->win_l_width[j]; shift <= pst->win_r_width[j]; shift++)
pst->win_coefficient[j][shift] = HTS_SStreamSet_get_window_coefficient(sss, i, j, shift);
}
/* copy GV */
if (HTS_SStreamSet_use_gv(sss, i)) {
pst->gv_mean = (double *) HTS_calloc(pst->vector_length, sizeof(double));
pst->gv_vari = (double *) HTS_calloc(pst->vector_length, sizeof(double));
for (j = 0; j < pst->vector_length; j++) {
pst->gv_mean[j] = HTS_SStreamSet_get_gv_mean(sss, i, j) * gv_weight[i];
pst->gv_vari[j] = HTS_SStreamSet_get_gv_vari(sss, i, j);
}
pst->gv_switch = (HTS_Boolean *) HTS_calloc(pst->length, sizeof(HTS_Boolean));
if (HTS_SStreamSet_is_msd(sss, i) == TRUE) { /* for MSD */
for (state = 0, frame = 0, msd_frame = 0; state < HTS_SStreamSet_get_total_state(sss); state++)
for (j = 0; j < HTS_SStreamSet_get_duration(sss, state); j++, frame++)
if (pst->msd_flag[frame] == TRUE)
pst->gv_switch[msd_frame++] = HTS_SStreamSet_get_gv_switch(sss, i, state);
} else { /* for non MSD */
for (state = 0, frame = 0; state < HTS_SStreamSet_get_total_state(sss); state++)
for (j = 0; j < HTS_SStreamSet_get_duration(sss, state); j++)
pst->gv_switch[frame++] = HTS_SStreamSet_get_gv_switch(sss, i, state);
}
for (j = 0, pst->gv_length = 0; j < pst->length; j++)
if (pst->gv_switch[j])
pst->gv_length++;
} else {
pst->gv_switch = NULL;
pst->gv_length = 0;
pst->gv_mean = NULL;
pst->gv_vari = NULL;
}
/* copy pdfs */
if (HTS_SStreamSet_is_msd(sss, i) == TRUE) { /* for MSD */
for (state = 0, frame = 0, msd_frame = 0; state < HTS_SStreamSet_get_total_state(sss); state++) {
for (j = 0; j < HTS_SStreamSet_get_duration(sss, state); j++) {
if (pst->msd_flag[frame] == TRUE) {
/* check current frame is MSD boundary or not */
for (k = 0; k < pst->win_size; k++) {
not_bound = TRUE;
for (shift = pst->win_l_width[k]; shift <= pst->win_r_width[k]; shift++)
if ((int) frame + shift < 0 || (int) pss->total_frame <= (int) frame + shift || pst->msd_flag[frame + shift] != TRUE) {
not_bound = FALSE;
break;
}
for (l = 0; l < pst->vector_length; l++) {
m = pst->vector_length * k + l;
pst->sm.mean[msd_frame][m] = HTS_SStreamSet_get_mean(sss, i, state, m);
if (not_bound || k == 0)
pst->sm.ivar[msd_frame][m] = HTS_finv(HTS_SStreamSet_get_vari(sss, i, state, m));
else
pst->sm.ivar[msd_frame][m] = 0.0;
}
}
msd_frame++;
}
frame++;
}
}
} else { /* for non MSD */
for (state = 0, frame = 0; state < HTS_SStreamSet_get_total_state(sss); state++) {
for (j = 0; j < HTS_SStreamSet_get_duration(sss, state); j++) {
for (k = 0; k < pst->win_size; k++) {
not_bound = TRUE;
for (shift = pst->win_l_width[k]; shift <= pst->win_r_width[k]; shift++)
if ((int) frame + shift < 0 || (int) pss->total_frame <= (int) frame + shift) {
not_bound = FALSE;
break;
}
for (l = 0; l < pst->vector_length; l++) {
m = pst->vector_length * k + l;
pst->sm.mean[frame][m] = HTS_SStreamSet_get_mean(sss, i, state, m);
if (not_bound || k == 0)
pst->sm.ivar[frame][m] = HTS_finv(HTS_SStreamSet_get_vari(sss, i, state, m));
else
pst->sm.ivar[frame][m] = 0.0;
}
}
frame++;
}
}
}
/* parameter generation */
HTS_PStream_mlpg(pst);
}
return TRUE;
}
/* HTS_PStreamSet_get_nstream: get number of stream */
size_t HTS_PStreamSet_get_nstream(HTS_PStreamSet * pss)
{
return pss->nstream;
}
/* HTS_PStreamSet_get_vector_length: get feature length */
size_t HTS_PStreamSet_get_vector_length(HTS_PStreamSet * pss, size_t stream_index)
{
return pss->pstream[stream_index].vector_length;
}
/* HTS_PStreamSet_get_total_frame: get total number of frame */
size_t HTS_PStreamSet_get_total_frame(HTS_PStreamSet * pss)
{
return pss->total_frame;
}
/* HTS_PStreamSet_get_parameter: get parameter */
double HTS_PStreamSet_get_parameter(HTS_PStreamSet * pss, size_t stream_index, size_t frame_index, size_t vector_index)
{
return pss->pstream[stream_index].par[frame_index][vector_index];
}
/* HTS_PStreamSet_get_parameter_vector: get parameter vector*/
double *HTS_PStreamSet_get_parameter_vector(HTS_PStreamSet * pss, size_t stream_index, size_t frame_index)
{
return pss->pstream[stream_index].par[frame_index];
}
/* HTS_PStreamSet_get_msd_flag: get generated MSD flag per frame */
HTS_Boolean HTS_PStreamSet_get_msd_flag(HTS_PStreamSet * pss, size_t stream_index, size_t frame_index)
{
return pss->pstream[stream_index].msd_flag[frame_index];
}
/* HTS_PStreamSet_is_msd: get MSD flag */
HTS_Boolean HTS_PStreamSet_is_msd(HTS_PStreamSet * pss, size_t stream_index)
{
return pss->pstream[stream_index].msd_flag ? TRUE : FALSE;
}
/* HTS_PStreamSet_clear: free parameter stream set */
void HTS_PStreamSet_clear(HTS_PStreamSet * pss)
{
size_t i, j;
HTS_PStream *pstream;
if (pss->pstream) {
for (i = 0; i < pss->nstream; i++) {
pstream = &pss->pstream[i];
if (pstream->sm.wum)
HTS_free(pstream->sm.wum);
if (pstream->sm.g)
HTS_free(pstream->sm.g);
if (pstream->sm.wuw)
HTS_free_matrix(pstream->sm.wuw, pstream->length);
if (pstream->sm.ivar)
HTS_free_matrix(pstream->sm.ivar, pstream->length);
if (pstream->sm.mean)
HTS_free_matrix(pstream->sm.mean, pstream->length);
if (pstream->par)
HTS_free_matrix(pstream->par, pstream->length);
if (pstream->msd_flag)
HTS_free(pstream->msd_flag);
if (pstream->win_coefficient) {
for (j = 0; j < pstream->win_size; j++) {
pstream->win_coefficient[j] += pstream->win_l_width[j];
HTS_free(pstream->win_coefficient[j]);
}
}
if (pstream->gv_mean)
HTS_free(pstream->gv_mean);
if (pstream->gv_vari)
HTS_free(pstream->gv_vari);
if (pstream->win_coefficient)
HTS_free(pstream->win_coefficient);
if (pstream->win_l_width)
HTS_free(pstream->win_l_width);
if (pstream->win_r_width)
HTS_free(pstream->win_r_width);
if (pstream->gv_switch)
HTS_free(pstream->gv_switch);
}
HTS_free(pss->pstream);
}
HTS_PStreamSet_initialize(pss);
}
HTS_PSTREAM_C_END;
#endif /* !HTS_PSTREAM_C */
View git blame Copy permalink