2 * G.729, G729 Annex D postfilter
3 * Copyright (c) 2008 Vladimir Voroshilov
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 #include "acelp_pitch_delay.h"
27 #include "g729postfilter.h"
28 #include "celp_math.h"
29 #include "acelp_filters.h"
30 #include "acelp_vectors.h"
31 #include "celp_filters.h"
37 * short interpolation filter (of length 33, according to spec)
38 * for computing signal with non-integer delay
40 static const int16_t ff_g729_interp_filt_short[(ANALYZED_FRAC_DELAYS+1)*SHORT_INT_FILT_LEN] = {
41 0, 31650, 28469, 23705, 18050, 12266, 7041, 2873,
42 0, -1597, -2147, -1992, -1492, -933, -484, -188,
46 * long interpolation filter (of length 129, according to spec)
47 * for computing signal with non-integer delay
49 static const int16_t ff_g729_interp_filt_long[(ANALYZED_FRAC_DELAYS+1)*LONG_INT_FILT_LEN] = {
50 0, 31915, 29436, 25569, 20676, 15206, 9639, 4439,
51 0, -3390, -5579, -6549, -6414, -5392, -3773, -1874,
52 0, 1595, 2727, 3303, 3319, 2850, 2030, 1023,
53 0, -887, -1527, -1860, -1876, -1614, -1150, -579,
54 0, 501, 859, 1041, 1044, 892, 631, 315,
55 0, -266, -453, -543, -538, -455, -317, -156,
56 0, 130, 218, 258, 253, 212, 147, 72,
57 0, -59, -101, -122, -123, -106, -77, -40,
61 * formant_pp_factor_num_pow[i] = FORMANT_PP_FACTOR_NUM^(i+1)
63 static const int16_t formant_pp_factor_num_pow[10]= {
65 18022, 9912, 5451, 2998, 1649, 907, 499, 274, 151, 83
69 * formant_pp_factor_den_pow[i] = FORMANT_PP_FACTOR_DEN^(i+1)
71 static const int16_t formant_pp_factor_den_pow[10] = {
73 22938, 16057, 11240, 7868, 5508, 3856, 2699, 1889, 1322, 925
77 * \brief Residual signal calculation (4.2.1 if G.729)
78 * \param out [out] output data filtered through A(z/FORMANT_PP_FACTOR_NUM)
79 * \param filter_coeffs (3.12) A(z/FORMANT_PP_FACTOR_NUM) filter coefficients
80 * \param in input speech data to process
81 * \param subframe_size size of one subframe
83 * \note in buffer must contain 10 items of previous speech data before top of the buffer
84 * \remark It is safe to pass the same buffer for input and output.
86 static void residual_filter(int16_t* out, const int16_t* filter_coeffs, const int16_t* in,
91 for (n = subframe_size - 1; n >= 0; n--) {
93 for (i = 0; i < 10; i++)
94 sum += filter_coeffs[i] * in[n - i - 1];
96 out[n] = in[n] + (sum >> 12);
101 * \brief long-term postfilter (4.2.1)
102 * \param dsp initialized DSP context
103 * \param pitch_delay_int integer part of the pitch delay in the first subframe
104 * \param residual filtering input data
105 * \param residual_filt [out] speech signal with applied A(z/FORMANT_PP_FACTOR_NUM) filter
106 * \param subframe_size size of subframe
108 * \return 0 if long-term prediction gain is less than 3dB, 1 - otherwise
110 static int16_t long_term_filter(AudioDSPContext *adsp, int pitch_delay_int,
111 const int16_t* residual, int16_t *residual_filt,
121 int corr_int_num, corr_int_den;
126 int16_t gain_num,gain_den; //selected signal's gain numerator and denominator
127 int16_t sh_gain_num, sh_gain_den;
130 int16_t gain_long_num,gain_long_den; //filtered through long interpolation filter signal's gain numerator and denominator
131 int16_t sh_gain_long_num, sh_gain_long_den;
133 int16_t best_delay_int, best_delay_frac;
135 int16_t delayed_signal_offset;
136 int lt_filt_factor_a, lt_filt_factor_b;
138 int16_t * selected_signal;
139 const int16_t * selected_signal_const; //Necessary to avoid compiler warning
141 int16_t sig_scaled[SUBFRAME_SIZE + RES_PREV_DATA_SIZE];
142 int16_t delayed_signal[ANALYZED_FRAC_DELAYS][SUBFRAME_SIZE+1];
143 int corr_den[ANALYZED_FRAC_DELAYS][2];
146 for(i=0; i<subframe_size + RES_PREV_DATA_SIZE; i++)
147 tmp |= FFABS(residual[i]);
152 shift = av_log2(tmp) - 11;
155 for (i = 0; i < subframe_size + RES_PREV_DATA_SIZE; i++)
156 sig_scaled[i] = residual[i] >> shift;
158 for (i = 0; i < subframe_size + RES_PREV_DATA_SIZE; i++)
159 sig_scaled[i] = residual[i] << -shift;
161 /* Start of best delay searching code */
164 ener = adsp->scalarproduct_int16(sig_scaled + RES_PREV_DATA_SIZE,
165 sig_scaled + RES_PREV_DATA_SIZE,
168 sh_ener = FFMAX(av_log2(ener) - 14, 0);
170 /* Search for best pitch delay.
172 sum{ r(n) * r(k,n) ] }^2
173 R'(k)^2 := -------------------------
174 sum{ r(k,n) * r(k,n) }
177 R(T) := sum{ r(n) * r(n-T) ] }
181 r(n-T) is integer delayed signal with delay T
182 r(k,n) is non-integer delayed signal with integer delay best_delay
183 and fractional delay k */
185 /* Find integer delay best_delay which maximizes correlation R(T).
187 This is also equals to numerator of R'(0),
188 since the fine search (second step) is done with 1/8
189 precision around best_delay. */
191 best_delay_int = pitch_delay_int - 1;
192 for (i = pitch_delay_int - 1; i <= pitch_delay_int + 1; i++) {
193 sum = adsp->scalarproduct_int16(sig_scaled + RES_PREV_DATA_SIZE,
194 sig_scaled + RES_PREV_DATA_SIZE - i,
196 if (sum > corr_int_num) {
202 /* Compute denominator of pseudo-normalized correlation R'(0). */
203 corr_int_den = adsp->scalarproduct_int16(sig_scaled - best_delay_int + RES_PREV_DATA_SIZE,
204 sig_scaled - best_delay_int + RES_PREV_DATA_SIZE,
207 /* Compute signals with non-integer delay k (with 1/8 precision),
208 where k is in [0;6] range.
209 Entire delay is qual to best_delay+(k+1)/8
210 This is archieved by applying an interpolation filter of
211 legth 33 to source signal. */
212 for (k = 0; k < ANALYZED_FRAC_DELAYS; k++) {
213 ff_acelp_interpolate(&delayed_signal[k][0],
214 &sig_scaled[RES_PREV_DATA_SIZE - best_delay_int],
215 ff_g729_interp_filt_short,
216 ANALYZED_FRAC_DELAYS+1,
222 /* Compute denominator of pseudo-normalized correlation R'(k).
224 corr_den[k][0] is square root of R'(k) denominator, for int(T) == int(T0)
225 corr_den[k][1] is square root of R'(k) denominator, for int(T) == int(T0)+1
227 Also compute maximum value of above denominators over all k. */
229 for (k = 0; k < ANALYZED_FRAC_DELAYS; k++) {
230 sum = adsp->scalarproduct_int16(&delayed_signal[k][1],
231 &delayed_signal[k][1],
233 corr_den[k][0] = sum + delayed_signal[k][0 ] * delayed_signal[k][0 ];
234 corr_den[k][1] = sum + delayed_signal[k][subframe_size] * delayed_signal[k][subframe_size];
236 tmp = FFMAX3(tmp, corr_den[k][0], corr_den[k][1]);
239 sh_gain_den = av_log2(tmp) - 14;
240 if (sh_gain_den >= 0) {
242 sh_gain_num = FFMAX(sh_gain_den, sh_ener);
243 /* Loop through all k and find delay that maximizes
245 Search is done in [int(T0)-1; intT(0)+1] range
246 with 1/8 precision. */
247 delayed_signal_offset = 1;
249 gain_den = corr_int_den >> sh_gain_den;
250 gain_num = corr_int_num >> sh_gain_num;
251 gain_num_square = gain_num * gain_num;
252 for (k = 0; k < ANALYZED_FRAC_DELAYS; k++) {
253 for (i = 0; i < 2; i++) {
254 int16_t gain_num_short, gain_den_short;
255 int gain_num_short_square;
256 /* Compute numerator of pseudo-normalized
257 correlation R'(k). */
258 sum = adsp->scalarproduct_int16(&delayed_signal[k][i],
259 sig_scaled + RES_PREV_DATA_SIZE,
261 gain_num_short = FFMAX(sum >> sh_gain_num, 0);
264 gain_num_short_square gain_num_square
265 R'(T)^2 = -----------------------, max R'(T)^2= --------------
268 gain_num_short_square = gain_num_short * gain_num_short;
269 gain_den_short = corr_den[k][i] >> sh_gain_den;
271 tmp = MULL(gain_num_short_square, gain_den, FRAC_BITS);
272 tmp2 = MULL(gain_num_square, gain_den_short, FRAC_BITS);
274 // R'(T)^2 > max R'(T)^2
276 gain_num = gain_num_short;
277 gain_den = gain_den_short;
278 gain_num_square = gain_num_short_square;
279 delayed_signal_offset = i;
280 best_delay_frac = k + 1;
290 L64_temp0 = (int64_t)gain_num_square << ((sh_gain_num << 1) + 1);
291 L64_temp1 = ((int64_t)gain_den * ener) << (sh_gain_den + sh_ener);
292 if (L64_temp0 < L64_temp1)
294 } // if(sh_gain_den >= 0)
295 } // if(corr_int_num)
297 /* End of best delay searching code */
300 memcpy(residual_filt, residual + RES_PREV_DATA_SIZE, subframe_size * sizeof(int16_t));
302 /* Long-term prediction gain is less than 3dB. Long-term postfilter is disabled. */
305 if (best_delay_frac) {
306 /* Recompute delayed signal with an interpolation filter of length 129. */
307 ff_acelp_interpolate(residual_filt,
308 &sig_scaled[RES_PREV_DATA_SIZE - best_delay_int + delayed_signal_offset],
309 ff_g729_interp_filt_long,
310 ANALYZED_FRAC_DELAYS + 1,
314 /* Compute R'(k) correlation's numerator. */
315 sum = adsp->scalarproduct_int16(residual_filt,
316 sig_scaled + RES_PREV_DATA_SIZE,
321 sh_gain_long_num = 0;
323 tmp = FFMAX(av_log2(sum) - 14, 0);
326 sh_gain_long_num = tmp;
329 /* Compute R'(k) correlation's denominator. */
330 sum = adsp->scalarproduct_int16(residual_filt, residual_filt, subframe_size);
332 tmp = FFMAX(av_log2(sum) - 14, 0);
335 sh_gain_long_den = tmp;
337 /* Select between original and delayed signal.
338 Delayed signal will be selected if it increases R'(k)
340 L_temp0 = gain_num * gain_num;
341 L_temp0 = MULL(L_temp0, gain_long_den, FRAC_BITS);
343 L_temp1 = gain_long_num * gain_long_num;
344 L_temp1 = MULL(L_temp1, gain_den, FRAC_BITS);
346 tmp = ((sh_gain_long_num - sh_gain_num) << 1) - (sh_gain_long_den - sh_gain_den);
352 /* Check if longer filter increases the values of R'(k). */
353 if (L_temp1 > L_temp0) {
354 /* Select long filter. */
355 selected_signal = residual_filt;
356 gain_num = gain_long_num;
357 gain_den = gain_long_den;
358 sh_gain_num = sh_gain_long_num;
359 sh_gain_den = sh_gain_long_den;
361 /* Select short filter. */
362 selected_signal = &delayed_signal[best_delay_frac-1][delayed_signal_offset];
364 /* Rescale selected signal to original value. */
366 for (i = 0; i < subframe_size; i++)
367 selected_signal[i] <<= shift;
369 for (i = 0; i < subframe_size; i++)
370 selected_signal[i] >>= -shift;
372 /* necessary to avoid compiler warning */
373 selected_signal_const = selected_signal;
374 } // if(best_delay_frac)
376 selected_signal_const = residual + RES_PREV_DATA_SIZE - (best_delay_int + 1 - delayed_signal_offset);
378 tmp = sh_gain_num - sh_gain_den;
384 if (gain_num > gain_den)
385 lt_filt_factor_a = MIN_LT_FILT_FACTOR_A;
389 lt_filt_factor_a = (gain_den << 15) / (gain_den + gain_num);
392 L64_temp0 = (((int64_t)gain_num) << sh_gain_num) >> 1;
393 L64_temp1 = ((int64_t)gain_den) << sh_gain_den;
394 lt_filt_factor_a = FFMAX((L64_temp1 << 15) / (L64_temp1 + L64_temp0), MIN_LT_FILT_FACTOR_A);
397 /* Filter through selected filter. */
398 lt_filt_factor_b = 32767 - lt_filt_factor_a + 1;
400 ff_acelp_weighted_vector_sum(residual_filt, residual + RES_PREV_DATA_SIZE,
401 selected_signal_const,
402 lt_filt_factor_a, lt_filt_factor_b,
403 1<<14, 15, subframe_size);
405 // Long-term prediction gain is larger than 3dB.
410 * \brief Calculate reflection coefficient for tilt compensation filter (4.2.3).
411 * \param dsp initialized DSP context
412 * \param lp_gn (3.12) coefficients of A(z/FORMANT_PP_FACTOR_NUM) filter
413 * \param lp_gd (3.12) coefficients of A(z/FORMANT_PP_FACTOR_DEN) filter
414 * \param speech speech to update
415 * \param subframe_size size of subframe
417 * \return (3.12) reflection coefficient
419 * \remark The routine also calculates the gain term for the short-term
420 * filter (gf) and multiplies the speech data by 1/gf.
422 * \note All members of lp_gn, except 10-19 must be equal to zero.
424 static int16_t get_tilt_comp(AudioDSPContext *adsp, int16_t *lp_gn,
425 const int16_t *lp_gd, int16_t* speech,
428 int rh1,rh0; // (3.12)
433 lp_gn[10] = 4096; //1.0 in (3.12)
435 /* Apply 1/A(z/FORMANT_PP_FACTOR_DEN) filter to hf. */
436 ff_celp_lp_synthesis_filter(lp_gn + 11, lp_gd + 1, lp_gn + 11, 22, 10, 0, 0, 0x800);
437 /* Now lp_gn (starting with 10) contains impulse response
438 of A(z/FORMANT_PP_FACTOR_NUM)/A(z/FORMANT_PP_FACTOR_DEN) filter. */
440 rh0 = adsp->scalarproduct_int16(lp_gn + 10, lp_gn + 10, 20);
441 rh1 = adsp->scalarproduct_int16(lp_gn + 10, lp_gn + 11, 20);
443 /* downscale to avoid overflow */
444 temp = av_log2(rh0) - 14;
450 if (FFABS(rh1) > rh0 || !rh0)
454 for (i = 0; i < 20; i++)
455 gain_term += FFABS(lp_gn[i + 10]);
456 gain_term >>= 2; // (3.12) -> (5.10)
458 if (gain_term > 0x400) { // 1.0 in (5.10)
459 temp = 0x2000000 / gain_term; // 1.0/gain_term in (0.15)
460 for (i = 0; i < subframe_size; i++)
461 speech[i] = (speech[i] * temp + 0x4000) >> 15;
464 return -(rh1 << 15) / rh0;
468 * \brief Apply tilt compensation filter (4.2.3).
469 * \param res_pst [in/out] residual signal (partially filtered)
470 * \param k1 (3.12) reflection coefficient
471 * \param subframe_size size of subframe
472 * \param ht_prev_data previous data for 4.2.3, equation 86
474 * \return new value for ht_prev_data
476 static int16_t apply_tilt_comp(int16_t* out, int16_t* res_pst, int refl_coeff,
477 int subframe_size, int16_t ht_prev_data)
484 if (refl_coeff > 0) {
485 gt = (refl_coeff * G729_TILT_FACTOR_PLUS + 0x4000) >> 15;
486 fact = 0x4000; // 0.5 in (0.15)
489 gt = (refl_coeff * G729_TILT_FACTOR_MINUS + 0x4000) >> 15;
490 fact = 0x800; // 0.5 in (3.12)
493 ga = (fact << 15) / av_clip_int16(32768 - FFABS(gt));
496 /* Apply tilt compensation filter to signal. */
497 tmp = res_pst[subframe_size - 1];
499 for (i = subframe_size - 1; i >= 1; i--) {
500 tmp2 = (res_pst[i] << 15) + ((gt * res_pst[i-1]) << 1);
501 tmp2 = (tmp2 + 0x4000) >> 15;
503 tmp2 = (tmp2 * ga * 2 + fact) >> sh_fact;
506 tmp2 = (res_pst[0] << 15) + ((gt * ht_prev_data) << 1);
507 tmp2 = (tmp2 + 0x4000) >> 15;
508 tmp2 = (tmp2 * ga * 2 + fact) >> sh_fact;
514 void ff_g729_postfilter(AudioDSPContext *adsp, int16_t* ht_prev_data, int* voicing,
515 const int16_t *lp_filter_coeffs, int pitch_delay_int,
516 int16_t* residual, int16_t* res_filter_data,
517 int16_t* pos_filter_data, int16_t *speech, int subframe_size)
519 int16_t residual_filt_buf[SUBFRAME_SIZE+11];
520 int16_t lp_gn[33]; // (3.12)
521 int16_t lp_gd[11]; // (3.12)
525 /* Zero-filling is necessary for tilt-compensation filter. */
526 memset(lp_gn, 0, 33 * sizeof(int16_t));
528 /* Calculate A(z/FORMANT_PP_FACTOR_NUM) filter coefficients. */
529 for (i = 0; i < 10; i++)
530 lp_gn[i + 11] = (lp_filter_coeffs[i + 1] * formant_pp_factor_num_pow[i] + 0x4000) >> 15;
532 /* Calculate A(z/FORMANT_PP_FACTOR_DEN) filter coefficients. */
533 for (i = 0; i < 10; i++)
534 lp_gd[i + 1] = (lp_filter_coeffs[i + 1] * formant_pp_factor_den_pow[i] + 0x4000) >> 15;
536 /* residual signal calculation (one-half of short-term postfilter) */
537 memcpy(speech - 10, res_filter_data, 10 * sizeof(int16_t));
538 residual_filter(residual + RES_PREV_DATA_SIZE, lp_gn + 11, speech, subframe_size);
539 /* Save data to use it in the next subframe. */
540 memcpy(res_filter_data, speech + subframe_size - 10, 10 * sizeof(int16_t));
542 /* long-term filter. If long-term prediction gain is larger than 3dB (returned value is
543 nonzero) then declare current subframe as periodic. */
544 *voicing = FFMAX(*voicing, long_term_filter(adsp, pitch_delay_int,
545 residual, residual_filt_buf + 10,
548 /* shift residual for using in next subframe */
549 memmove(residual, residual + subframe_size, RES_PREV_DATA_SIZE * sizeof(int16_t));
551 /* short-term filter tilt compensation */
552 tilt_comp_coeff = get_tilt_comp(adsp, lp_gn, lp_gd, residual_filt_buf + 10, subframe_size);
554 /* Apply second half of short-term postfilter: 1/A(z/FORMANT_PP_FACTOR_DEN) */
555 ff_celp_lp_synthesis_filter(pos_filter_data + 10, lp_gd + 1,
556 residual_filt_buf + 10,
557 subframe_size, 10, 0, 0, 0x800);
558 memcpy(pos_filter_data, pos_filter_data + subframe_size, 10 * sizeof(int16_t));
560 *ht_prev_data = apply_tilt_comp(speech, pos_filter_data + 10, tilt_comp_coeff,
561 subframe_size, *ht_prev_data);
565 * \brief Adaptive gain control (4.2.4)
566 * \param gain_before gain of speech before applying postfilters
567 * \param gain_after gain of speech after applying postfilters
568 * \param speech [in/out] signal buffer
569 * \param subframe_size length of subframe
570 * \param gain_prev (3.12) previous value of gain coefficient
572 * \return (3.12) last value of gain coefficient
574 int16_t ff_g729_adaptive_gain_control(int gain_before, int gain_after, int16_t *speech,
575 int subframe_size, int16_t gain_prev)
579 int exp_before, exp_after;
581 if(!gain_after && gain_before)
586 exp_before = 14 - av_log2(gain_before);
587 gain_before = bidir_sal(gain_before, exp_before);
589 exp_after = 14 - av_log2(gain_after);
590 gain_after = bidir_sal(gain_after, exp_after);
592 if (gain_before < gain_after) {
593 gain = (gain_before << 15) / gain_after;
594 gain = bidir_sal(gain, exp_after - exp_before - 1);
596 gain = ((gain_before - gain_after) << 14) / gain_after + 0x4000;
597 gain = bidir_sal(gain, exp_after - exp_before);
599 gain = (gain * G729_AGC_FAC1 + 0x4000) >> 15; // gain * (1-0.9875)
603 for (n = 0; n < subframe_size; n++) {
604 // gain_prev = gain + 0.9875 * gain_prev
605 gain_prev = (G729_AGC_FACTOR * gain_prev + 0x4000) >> 15;
606 gain_prev = av_clip_int16(gain + gain_prev);
607 speech[n] = av_clip_int16((speech[n] * gain_prev + 0x2000) >> 14);