ref: a6c5046f22be23435ab21c679e4bf14631f0e233
dir: /nok_pitch.c/
/*
* Long Term Prediction functions
*
* Copyright (c) 1999 M. Bakker
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2, or (at your option)
* any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; see the file COPYING. If not, write to
* the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
*/
/**************************************************************************
Version Control Information Method: CVS
Identifiers:
$Revision: 1.9 $
$Date: 2000/10/05 13:04:05 $ (check in)
$Author: menno $
*************************************************************************/
/**************************************************************************
External Objects Needed
*************************************************************************/
/*
Interface to related modules. */
#include "interface.h"
#include "nok_ltp_common.h"
#include "nok_ltp_enc.h"
/*************************************************************************
External Objects Provided
*************************************************************************/
#include "nok_pitch.h"
/**************************************************************************
Internal Objects
*************************************************************************/
#include "nok_ltp_common_internal.h"
void lnqgj (double (*a)[LPC + 1]);
void w_quantize (double *freq, int *ltp_idx);
/*
Purpose: Codebook for LTP weight coefficients. */
double codebook[CODESIZE] =
{
0.570829,
0.696616,
0.813004,
0.911304,
0.984900,
1.067894,
1.194601,
1.369533
};
/**************************************************************************
Title: snr_pred
Purpose: Determines is it feasible to employ long term prediction in
the encoder.
Usage: y = snr_pred(mdct_in, mdct_pred, sfb_flag, sfb_offset,
block_type, side_info, num_of_sfb)
Input: mdct_in - spectral coefficients
mdct_pred - predicted spectral coefficients
sfb_flag - an array of flags indicating whether LTP is
switched on (1) /off (0). One bit is reseved
for each sfb.
sfb_offset - scalefactor boundaries
block_type - window sequence type
side_info - LTP side information
num_of_sfb - number of scalefactor bands
Output: y - number of bits saved by using long term prediction
References: -
Explanation: -
Author(s): Juha Ojanpera, Lin Yin
*************************************************************************/
double
snr_pred (double *mdct_in,
double *mdct_pred,
int *sfb_flag,
int *sfb_offset,
enum WINDOW_TYPE block_type,
int side_info,
int num_of_sfb
)
{
int i, j, flen;
double snr_limit;
double num_bit, snr[NSFB_LONG];
double temp1, temp2;
double energy[BLOCK_LEN_LONG], snr_p[BLOCK_LEN_LONG];
if (block_type != ONLY_SHORT_WINDOW)
{
flen = BLOCK_LEN_LONG;
snr_limit = 1.e-30;
}
else
{
flen = BLOCK_LEN_SHORT;
snr_limit = 1.e-20;
}
for (i = 0; i < flen; i++)
{
energy[i] = mdct_in[i] * mdct_in[i];
snr_p[i] = (mdct_in[i] - mdct_pred[i]) * (mdct_in[i] - mdct_pred[i]);
}
num_bit = 0.0;
for (i = 0; i < num_of_sfb; i++)
{
temp1 = 0.0;
temp2 = 0.0;
for (j = sfb_offset[i]; j < sfb_offset[i + 1]; j++)
{
temp1 += energy[j];
temp2 += snr_p[j];
}
if (temp2 < snr_limit)
temp2 = snr_limit;
if (temp1 > /*1.e-20*/0.0)
snr[i] = -10. * log10 (temp2 / temp1);
else
snr[i] = 0.0;
sfb_flag[i] = 1;
if (block_type != ONLY_SHORT_WINDOW)
{
if (snr[i] <= 0.0)
{
sfb_flag[i] = 0;
for (j = sfb_offset[i]; j < sfb_offset[i + 1]; j++)
mdct_pred[j] = 0.0;
} else {
num_bit += snr[i] / 6. * (sfb_offset[i + 1] - sfb_offset[i]);
}
}
}
if (num_bit < side_info)
{
num_bit = 0.0;
for (j = 0; j < flen; j++)
mdct_pred[j] = 0.0;
for (i = 0; i < num_of_sfb; i++)
sfb_flag[i] = 0;
}
else
num_bit -= side_info;
return (num_bit);
}
/**************************************************************************
Title: prediction
Purpose: Predicts current frame from the past output samples.
Usage: prediction(buffer, predicted_samples, weight, delay, flen)
Input: buffer - past reconstructed output samples
weight - LTP gain (scaling factor)
delay - LTP lag (optimal delay)
flen - length of the frame
Output: predicted_samples - predicted samples
References: -
Explanation: -
Author(s): Juha Ojanpera, Lin Yin
*************************************************************************/
void
prediction (short *buffer,
double *predicted_samples,
double *weight,
int delay,
int flen
)
{
int i, j;
int offset;
for (i = 0; i < flen; i++)
{
offset = NOK_LT_BLEN - flen - delay + i - (LPC - 1) / 2;
predicted_samples[i] = 0.0;
for (j = 0; j < LPC; j++)
predicted_samples[i] += weight[j] * buffer[offset + j];
}
}
/**************************************************************************
Title: estimate_delay
Purpose: Estimates optimal delay between current frame and past
reconstructed output samples. The lag between 0...DELAY-1
with the maximum correlation becomes the LTP lag (or delay).
Usage: y = estimate_delay(sb_samples, x_buffer, flen)
Input: sb_samples - current frame
x_buffer - past reconstructed output samples
flen - length of the frame
Output: y - LTP lag
References: -
Explanation: -
Author(s): Juha Ojanpera, Lin Yin
*************************************************************************/
int estimate_delay (double *sb_samples,
short *x_buffer
// ,int flen
)
{
int i, j;
int delay;
double corr[DELAY],corrtmp;
double p_max, energy;
p_max = 0.0;
delay = 0;
energy = 0.0;
corr[0] = 0.0;
for (j = 1; j < 2049; j++)
{
corr[0] += x_buffer[NOK_LT_BLEN - j] * sb_samples[2048 - j];
energy += x_buffer[NOK_LT_BLEN - j] * x_buffer[NOK_LT_BLEN - j];
}
corrtmp=corr[0];
if (energy != 0.0)
corr[0] = corr[0] / sqrt(energy);
else
corr[0] = 0.0;
if (p_max < corr[0])
{
p_max = corr[0];
delay = 0;
}
/* Used to look like this:
for (i = 1; i < DELAY; i+=16)
Because the new code by Oxygene2000 is so fast, we can now look for the
delay in steps of 1, instead of 16. Thus giving a more accurate delay estimation
*/
for (i = 1; i < DELAY; i++)
{
//NOK_LT_BLEN=4096
energy -= x_buffer[NOK_LT_BLEN - i] * x_buffer[NOK_LT_BLEN - i];
energy += x_buffer[NOK_LT_BLEN - i - 2048] * x_buffer[NOK_LT_BLEN - i - 2048]; //2048=j_max
corr[i] = corrtmp;
corr[i] -= x_buffer[NOK_LT_BLEN - i] * sb_samples[2047];
corr[i] += x_buffer[NOK_LT_BLEN - i - 2048] * sb_samples[0];
corrtmp=corr[i];
//flen=2048
// for (j = 1; j < 2049; j++) //2049=flen+1
// {
// corr[i] += x_buffer[NOK_LT_BLEN - i - j] * sb_samples[2048 - j];
// energy += x_buffer[NOK_LT_BLEN - i - j] * x_buffer[NOK_LT_BLEN - i - j];
// }
if (energy != 0.0)
corr[i] = corr[i] / sqrt(energy);
else
corr[i] = 0.0;
if (p_max < corr[i])
{
p_max = corr[i];
delay = i;
}
}
// if (delay < (LPC - 1) / 2)
// delay = (LPC - 1) / 2;
if (delay<0) delay=0; //for LPC=1
/*
fprintf(stdout, "delay : %d ... ", delay);
*/
return delay;
}
/**************************************************************************
Title: pitch
Purpose: Calculates LTP gains, quantizes them and finally scales
the past output samples (from 'delay' to 'delay'+'flen')
to get the predicted frame.
Usage: pitch(sb_samples, sb_samples_pred, x_buffer, ltp_coef,
delay, flen)
Input: sb_samples - current frame
x_buffer - past reconstructed output samples
delay - LTP lag
flen - length of the frame
Output: sb_samples_pred - predicted frame
ltp_coef - indices of the LTP gains
References: 1.) lnqgj
2.) w_quantize
3.) prediction
Explanation: -
Author(s): Juha Ojanpera, Lin Yin
*************************************************************************/
void
pitch (double *sb_samples,
double *sb_samples_pred,
short *x_buffer,
int *ltp_coef,
int delay,
int flen
)
{
int i, k, j;
int offset1, offset2;
double weight[LPC];
double r[LPC][LPC + 1];
for (i = 0; i < LPC; i++)
for (j = 0; j < LPC; j++)
{
offset1 = NOK_LT_BLEN - flen - delay + i - (LPC - 1) / 2;
offset2 = NOK_LT_BLEN - flen - delay + j - (LPC - 1) / 2;
r[i][j] = 0.0;
for (k = 0; k < flen; k++)
r[i][j] += x_buffer[offset1 + k] * x_buffer[offset2 + k];
}
for (i = 0; i < LPC; i++)
r[i][i] = 1.010 * r[i][i];
for (i = 0; i < LPC; i++)
{
offset1 = NOK_LT_BLEN - flen - delay + i - (LPC - 1) / 2;
r[i][LPC] = 0.0;
for (k = 0; k < flen; k++)
r[i][LPC] += x_buffer[offset1 + k] * sb_samples[k];
}
for (i = 0; i < LPC; i++)
weight[i] = 0.0;
if (r[(LPC - 1) / 2][(LPC - 1) / 2] != 0.0)
weight[(LPC - 1) / 2] = r[(LPC - 1) / 2][LPC] / r[(LPC - 1) / 2][(LPC - 1) / 2];
lnqgj (r);
for (i = 0; i < LPC; i++)
weight[i] = r[i][LPC];
/*
fprintf(stdout, "unquantized weight : %f ... ", weight[0]);
*/
w_quantize (weight, ltp_coef);
/*
fprintf(stdout, "quantized weight : %f\n", weight[0]);
*/
prediction (x_buffer, sb_samples_pred, weight, delay, flen);
}
/**************************************************************************
Title: lnqgj
Purpose: Calculates LTP gains.
Usage: lnqgj(a)
Input: a - auto-correlation matrix
Output: a - LTP gains (in the last column)
References: -
Explanation: -
Author(s): Juha Ojanpera, Lin Yin
*************************************************************************/
void
lnqgj (double (*a)[LPC + 1])
{
int nn, nnpls1, ip, i, j;
double p, w;
for (nn = 0; nn < LPC; nn++)
{
p = 0.0;
nnpls1 = nn + 1;
for (i = nn; i < LPC; i++)
if (p - fabs (a[i][nn]) < 0)
{
p = fabs (a[i][nn]);
ip = i;
}
if (p - 1.e-10 <= 0)
return;
if (p - 1.e-10 > 0)
{
for (j = nn; j <= LPC; j++)
{
w = a[nn][j];
a[nn][j] = a[ip][j];
a[ip][j] = w;
}
for (j = nnpls1; j <= LPC; j++)
a[nn][j] = a[nn][j] / a[nn][nn];
for (i = 0; i < LPC; i++)
if ((i - nn) != 0)
for (j = nnpls1; j <= LPC; j++)
a[i][j] = a[i][j] - a[i][nn] * a[nn][j];
}
}
}
/**************************************************************************
Title: w_quantize
Purpose: Quantizes LTP gain(s).
Usage: w_quantize(freq, ltp_idx)
Input: freq - original LTP gain(s)
Output: freq - LTP gain(s) selected from the codebook
ltp_idx - corresponding indices of the LTP gain(s)
References: -
Explanation: The closest value from the codebook is selected to be the
quantized LTP gain.
Author(s): Juha Ojanpera, Lin Yin
*************************************************************************/
void
w_quantize (double *freq, int *ltp_idx)
{
int i, j;
double dist, low;
low = 1.0e+10;
for (i = 0; i < LPC; i++)
{
// dist = 0.0;
for (j = 0; j < CODESIZE; j++)
{
dist = (freq[i] - codebook[j]) * (freq[i] - codebook[j]);
if (dist < low)
{
low = dist;
ltp_idx[i] = j;
}
}
}
for (j = 0; j < LPC; j++)
freq[j] = codebook[ltp_idx[j]];
}