2 * G.723.1 compatible decoder
3 * Copyright (c) 2006 Benjamin Larsson
4 * Copyright (c) 2010 Mohamed Naufal Basheer
6 * This file is part of Libav.
8 * Libav is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * Libav is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with Libav; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * G.723.1 compatible decoder
28 #define BITSTREAM_READER_LE
29 #include "libavutil/channel_layout.h"
30 #include "libavutil/mem.h"
31 #include "libavutil/opt.h"
34 #include "acelp_vectors.h"
35 #include "celp_filters.h"
36 #include "g723_1_data.h"
39 #define CNG_RANDOM_SEED 12345
45 ACTIVE_FRAME, ///< Active speech
46 SID_FRAME, ///< Silence Insertion Descriptor frame
56 * G723.1 unpacked data subframe
59 int ad_cb_lag; ///< adaptive codebook lag
69 * Pitch postfilter parameters
72 int index; ///< postfilter backward/forward lag
73 int16_t opt_gain; ///< optimal gain
74 int16_t sc_gain; ///< scaling gain
77 typedef struct g723_1_context {
81 G723_1_Subframe subframe[4];
82 enum FrameType cur_frame_type;
83 enum FrameType past_frame_type;
85 uint8_t lsp_index[LSP_BANDS];
89 int16_t prev_lsp[LPC_ORDER];
90 int16_t sid_lsp[LPC_ORDER];
91 int16_t prev_excitation[PITCH_MAX];
92 int16_t excitation[PITCH_MAX + FRAME_LEN + 4];
93 int16_t synth_mem[LPC_ORDER];
94 int16_t fir_mem[LPC_ORDER];
95 int iir_mem[LPC_ORDER];
107 int16_t audio[FRAME_LEN + LPC_ORDER + PITCH_MAX + 4];
110 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
112 G723_1_Context *p = avctx->priv_data;
114 avctx->channel_layout = AV_CH_LAYOUT_MONO;
115 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
117 avctx->sample_rate = 8000;
118 p->pf_gain = 1 << 12;
120 avcodec_get_frame_defaults(&p->frame);
121 avctx->coded_frame = &p->frame;
123 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
124 memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
126 p->cng_random_seed = CNG_RANDOM_SEED;
127 p->past_frame_type = SID_FRAME;
133 * Unpack the frame into parameters.
135 * @param p the context
136 * @param buf pointer to the input buffer
137 * @param buf_size size of the input buffer
139 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
144 int temp, info_bits, i;
146 init_get_bits(&gb, buf, buf_size * 8);
148 /* Extract frame type and rate info */
149 info_bits = get_bits(&gb, 2);
151 if (info_bits == 3) {
152 p->cur_frame_type = UNTRANSMITTED_FRAME;
156 /* Extract 24 bit lsp indices, 8 bit for each band */
157 p->lsp_index[2] = get_bits(&gb, 8);
158 p->lsp_index[1] = get_bits(&gb, 8);
159 p->lsp_index[0] = get_bits(&gb, 8);
161 if (info_bits == 2) {
162 p->cur_frame_type = SID_FRAME;
163 p->subframe[0].amp_index = get_bits(&gb, 6);
167 /* Extract the info common to both rates */
168 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
169 p->cur_frame_type = ACTIVE_FRAME;
171 p->pitch_lag[0] = get_bits(&gb, 7);
172 if (p->pitch_lag[0] > 123) /* test if forbidden code */
174 p->pitch_lag[0] += PITCH_MIN;
175 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
177 p->pitch_lag[1] = get_bits(&gb, 7);
178 if (p->pitch_lag[1] > 123)
180 p->pitch_lag[1] += PITCH_MIN;
181 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
182 p->subframe[0].ad_cb_lag = 1;
183 p->subframe[2].ad_cb_lag = 1;
185 for (i = 0; i < SUBFRAMES; i++) {
186 /* Extract combined gain */
187 temp = get_bits(&gb, 12);
189 p->subframe[i].dirac_train = 0;
190 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
191 p->subframe[i].dirac_train = temp >> 11;
195 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
196 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
197 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
204 p->subframe[0].grid_index = get_bits(&gb, 1);
205 p->subframe[1].grid_index = get_bits(&gb, 1);
206 p->subframe[2].grid_index = get_bits(&gb, 1);
207 p->subframe[3].grid_index = get_bits(&gb, 1);
209 if (p->cur_rate == RATE_6300) {
210 skip_bits(&gb, 1); /* skip reserved bit */
212 /* Compute pulse_pos index using the 13-bit combined position index */
213 temp = get_bits(&gb, 13);
214 p->subframe[0].pulse_pos = temp / 810;
216 temp -= p->subframe[0].pulse_pos * 810;
217 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
219 temp -= p->subframe[1].pulse_pos * 90;
220 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
221 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
223 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
225 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
227 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
229 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
232 p->subframe[0].pulse_sign = get_bits(&gb, 6);
233 p->subframe[1].pulse_sign = get_bits(&gb, 5);
234 p->subframe[2].pulse_sign = get_bits(&gb, 6);
235 p->subframe[3].pulse_sign = get_bits(&gb, 5);
236 } else { /* 5300 bps */
237 p->subframe[0].pulse_pos = get_bits(&gb, 12);
238 p->subframe[1].pulse_pos = get_bits(&gb, 12);
239 p->subframe[2].pulse_pos = get_bits(&gb, 12);
240 p->subframe[3].pulse_pos = get_bits(&gb, 12);
242 p->subframe[0].pulse_sign = get_bits(&gb, 4);
243 p->subframe[1].pulse_sign = get_bits(&gb, 4);
244 p->subframe[2].pulse_sign = get_bits(&gb, 4);
245 p->subframe[3].pulse_sign = get_bits(&gb, 4);
252 * Bitexact implementation of sqrt(val/2).
254 static int16_t square_root(int val)
257 int16_t exp = 0x4000;
260 for (i = 0; i < 14; i ++) {
261 int res_exp = res + exp;
262 if (val >= res_exp * res_exp << 1)
270 * Calculate the number of left-shifts required for normalizing the input.
272 * @param num input number
273 * @param width width of the input, 16 bits(0) / 32 bits(1)
275 static int normalize_bits(int num, int width)
277 return width - av_log2(num) - 1;
281 * Scale vector contents based on the largest of their absolutes.
283 static int scale_vector(int16_t *dst, const int16_t *vector, int length)
289 for (i = 0; i < length; i++)
290 max |= FFABS(vector[i]);
292 max = FFMIN(max, 0x7FFF);
293 bits = normalize_bits(max, 15);
295 for (i = 0; i < length; i++)
296 dst[i] = vector[i] << bits >> 3;
302 * Perform inverse quantization of LSP frequencies.
304 * @param cur_lsp the current LSP vector
305 * @param prev_lsp the previous LSP vector
306 * @param lsp_index VQ indices
307 * @param bad_frame bad frame flag
309 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
310 uint8_t *lsp_index, int bad_frame)
313 int i, j, temp, stable;
315 /* Check for frame erasure */
322 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
325 /* Get the VQ table entry corresponding to the transmitted index */
326 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
327 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
328 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
329 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
330 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
331 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
332 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
333 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
334 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
335 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
337 /* Add predicted vector & DC component to the previously quantized vector */
338 for (i = 0; i < LPC_ORDER; i++) {
339 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
340 cur_lsp[i] += dc_lsp[i] + temp;
343 for (i = 0; i < LPC_ORDER; i++) {
344 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
345 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
347 /* Stability check */
348 for (j = 1; j < LPC_ORDER; j++) {
349 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
352 cur_lsp[j - 1] -= temp;
357 for (j = 1; j < LPC_ORDER; j++) {
358 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
368 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
372 * Bitexact implementation of 2ab scaled by 1/2^16.
374 * @param a 32 bit multiplicand
375 * @param b 16 bit multiplier
377 #define MULL2(a, b) \
378 ((((a) >> 16) * (b) << 1) + (((a) & 0xffff) * (b) >> 15))
381 * Convert LSP frequencies to LPC coefficients.
383 * @param lpc buffer for LPC coefficients
385 static void lsp2lpc(int16_t *lpc)
387 int f1[LPC_ORDER / 2 + 1];
388 int f2[LPC_ORDER / 2 + 1];
391 /* Calculate negative cosine */
392 for (j = 0; j < LPC_ORDER; j++) {
393 int index = lpc[j] >> 7;
394 int offset = lpc[j] & 0x7f;
395 int temp1 = cos_tab[index] << 16;
396 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
397 ((offset << 8) + 0x80) << 1;
399 lpc[j] = -(av_sat_dadd32(1 << 15, temp1 + temp2) >> 16);
403 * Compute sum and difference polynomial coefficients
404 * (bitexact alternative to lsp2poly() in lsp.c)
406 /* Initialize with values in Q28 */
408 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
409 f1[2] = lpc[0] * lpc[2] + (2 << 28);
412 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
413 f2[2] = lpc[1] * lpc[3] + (2 << 28);
416 * Calculate and scale the coefficients by 1/2 in
417 * each iteration for a final scaling factor of Q25
419 for (i = 2; i < LPC_ORDER / 2; i++) {
420 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
421 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
423 for (j = i; j >= 2; j--) {
424 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
425 (f1[j] >> 1) + (f1[j - 2] >> 1);
426 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
427 (f2[j] >> 1) + (f2[j - 2] >> 1);
432 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
433 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
436 /* Convert polynomial coefficients to LPC coefficients */
437 for (i = 0; i < LPC_ORDER / 2; i++) {
438 int64_t ff1 = f1[i + 1] + f1[i];
439 int64_t ff2 = f2[i + 1] - f2[i];
441 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
442 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
448 * Quantize LSP frequencies by interpolation and convert them to
449 * the corresponding LPC coefficients.
451 * @param lpc buffer for LPC coefficients
452 * @param cur_lsp the current LSP vector
453 * @param prev_lsp the previous LSP vector
455 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
458 int16_t *lpc_ptr = lpc;
460 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
461 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
462 4096, 12288, 1 << 13, 14, LPC_ORDER);
463 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
464 8192, 8192, 1 << 13, 14, LPC_ORDER);
465 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
466 12288, 4096, 1 << 13, 14, LPC_ORDER);
467 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
469 for (i = 0; i < SUBFRAMES; i++) {
471 lpc_ptr += LPC_ORDER;
476 * Generate a train of dirac functions with period as pitch lag.
478 static void gen_dirac_train(int16_t *buf, int pitch_lag)
480 int16_t vector[SUBFRAME_LEN];
483 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
484 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
485 for (j = 0; j < SUBFRAME_LEN - i; j++)
486 buf[i + j] += vector[j];
491 * Generate fixed codebook excitation vector.
493 * @param vector decoded excitation vector
494 * @param subfrm current subframe
495 * @param cur_rate current bitrate
496 * @param pitch_lag closed loop pitch lag
497 * @param index current subframe index
499 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
500 enum Rate cur_rate, int pitch_lag, int index)
504 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
506 if (cur_rate == RATE_6300) {
507 if (subfrm->pulse_pos >= max_pos[index])
510 /* Decode amplitudes and positions */
511 j = PULSE_MAX - pulses[index];
512 temp = subfrm->pulse_pos;
513 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
514 temp -= combinatorial_table[j][i];
517 temp += combinatorial_table[j++][i];
518 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
519 vector[subfrm->grid_index + GRID_SIZE * i] =
520 -fixed_cb_gain[subfrm->amp_index];
522 vector[subfrm->grid_index + GRID_SIZE * i] =
523 fixed_cb_gain[subfrm->amp_index];
528 if (subfrm->dirac_train == 1)
529 gen_dirac_train(vector, pitch_lag);
530 } else { /* 5300 bps */
531 int cb_gain = fixed_cb_gain[subfrm->amp_index];
532 int cb_shift = subfrm->grid_index;
533 int cb_sign = subfrm->pulse_sign;
534 int cb_pos = subfrm->pulse_pos;
535 int offset, beta, lag;
537 for (i = 0; i < 8; i += 2) {
538 offset = ((cb_pos & 7) << 3) + cb_shift + i;
539 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
544 /* Enhance harmonic components */
545 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
546 subfrm->ad_cb_lag - 1;
547 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
549 if (lag < SUBFRAME_LEN - 2) {
550 for (i = lag; i < SUBFRAME_LEN; i++)
551 vector[i] += beta * vector[i - lag] >> 15;
557 * Get delayed contribution from the previous excitation vector.
559 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
561 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
564 residual[0] = prev_excitation[offset];
565 residual[1] = prev_excitation[offset + 1];
568 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
569 residual[i] = prev_excitation[offset + (i - 2) % lag];
572 static int dot_product(const int16_t *a, const int16_t *b, int length)
576 for (i = 0; i < length; i++) {
577 int prod = a[i] * b[i];
578 sum = av_sat_dadd32(sum, prod);
584 * Generate adaptive codebook excitation.
586 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
587 int pitch_lag, G723_1_Subframe *subfrm,
590 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
591 const int16_t *cb_ptr;
592 int lag = pitch_lag + subfrm->ad_cb_lag - 1;
597 get_residual(residual, prev_excitation, lag);
599 /* Select quantization table */
600 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2)
601 cb_ptr = adaptive_cb_gain85;
603 cb_ptr = adaptive_cb_gain170;
605 /* Calculate adaptive vector */
606 cb_ptr += subfrm->ad_cb_gain * 20;
607 for (i = 0; i < SUBFRAME_LEN; i++) {
608 sum = dot_product(residual + i, cb_ptr, PITCH_ORDER);
609 vector[i] = av_sat_dadd32(1 << 15, sum) >> 16;
614 * Estimate maximum auto-correlation around pitch lag.
616 * @param buf buffer with offset applied
617 * @param offset offset of the excitation vector
618 * @param ccr_max pointer to the maximum auto-correlation
619 * @param pitch_lag decoded pitch lag
620 * @param length length of autocorrelation
621 * @param dir forward lag(1) / backward lag(-1)
623 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
624 int pitch_lag, int length, int dir)
626 int limit, ccr, lag = 0;
629 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
631 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
633 limit = pitch_lag + 3;
635 for (i = pitch_lag - 3; i <= limit; i++) {
636 ccr = dot_product(buf, buf + dir * i, length);
638 if (ccr > *ccr_max) {
647 * Calculate pitch postfilter optimal and scaling gains.
649 * @param lag pitch postfilter forward/backward lag
650 * @param ppf pitch postfilter parameters
651 * @param cur_rate current bitrate
652 * @param tgt_eng target energy
653 * @param ccr cross-correlation
654 * @param res_eng residual energy
656 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
657 int tgt_eng, int ccr, int res_eng)
659 int pf_residual; /* square of postfiltered residual */
664 temp1 = tgt_eng * res_eng >> 1;
665 temp2 = ccr * ccr << 1;
668 if (ccr >= res_eng) {
669 ppf->opt_gain = ppf_gain_weight[cur_rate];
671 ppf->opt_gain = (ccr << 15) / res_eng *
672 ppf_gain_weight[cur_rate] >> 15;
674 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
675 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
676 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
677 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
679 if (tgt_eng >= pf_residual << 1) {
682 temp1 = (tgt_eng << 14) / pf_residual;
685 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
686 ppf->sc_gain = square_root(temp1 << 16);
689 ppf->sc_gain = 0x7fff;
692 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
696 * Calculate pitch postfilter parameters.
698 * @param p the context
699 * @param offset offset of the excitation vector
700 * @param pitch_lag decoded pitch lag
701 * @param ppf pitch postfilter parameters
702 * @param cur_rate current bitrate
704 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
705 PPFParam *ppf, enum Rate cur_rate)
714 * 1 - forward cross-correlation
715 * 2 - forward residual energy
716 * 3 - backward cross-correlation
717 * 4 - backward residual energy
719 int energy[5] = {0, 0, 0, 0, 0};
720 int16_t *buf = p->audio + LPC_ORDER + offset;
721 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
723 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
728 ppf->sc_gain = 0x7fff;
730 /* Case 0, Section 3.6 */
731 if (!back_lag && !fwd_lag)
734 /* Compute target energy */
735 energy[0] = dot_product(buf, buf, SUBFRAME_LEN);
737 /* Compute forward residual energy */
739 energy[2] = dot_product(buf + fwd_lag, buf + fwd_lag, SUBFRAME_LEN);
741 /* Compute backward residual energy */
743 energy[4] = dot_product(buf - back_lag, buf - back_lag, SUBFRAME_LEN);
745 /* Normalize and shorten */
747 for (i = 0; i < 5; i++)
748 temp1 = FFMAX(energy[i], temp1);
750 scale = normalize_bits(temp1, 31);
751 for (i = 0; i < 5; i++)
752 energy[i] = (energy[i] << scale) >> 16;
754 if (fwd_lag && !back_lag) { /* Case 1 */
755 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
757 } else if (!fwd_lag) { /* Case 2 */
758 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
760 } else { /* Case 3 */
763 * Select the largest of energy[1]^2/energy[2]
764 * and energy[3]^2/energy[4]
766 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
767 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
768 if (temp1 >= temp2) {
769 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
772 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
779 * Classify frames as voiced/unvoiced.
781 * @param p the context
782 * @param pitch_lag decoded pitch_lag
783 * @param exc_eng excitation energy estimation
784 * @param scale scaling factor of exc_eng
786 * @return residual interpolation index if voiced, 0 otherwise
788 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
789 int *exc_eng, int *scale)
791 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
792 int16_t *buf = p->audio + LPC_ORDER;
794 int index, ccr, tgt_eng, best_eng, temp;
796 *scale = scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
799 /* Compute maximum backward cross-correlation */
801 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
802 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
804 /* Compute target energy */
805 tgt_eng = dot_product(buf, buf, SUBFRAME_LEN * 2);
806 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
811 /* Compute best energy */
812 best_eng = dot_product(buf - index, buf - index, SUBFRAME_LEN * 2);
813 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
815 temp = best_eng * *exc_eng >> 3;
817 if (temp < ccr * ccr)
824 * Peform residual interpolation based on frame classification.
826 * @param buf decoded excitation vector
827 * @param out output vector
828 * @param lag decoded pitch lag
829 * @param gain interpolated gain
830 * @param rseed seed for random number generator
832 static void residual_interp(int16_t *buf, int16_t *out, int lag,
833 int gain, int *rseed)
836 if (lag) { /* Voiced */
837 int16_t *vector_ptr = buf + PITCH_MAX;
839 for (i = 0; i < lag; i++)
840 out[i] = vector_ptr[i - lag] * 3 >> 2;
841 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
842 (FRAME_LEN - lag) * sizeof(*out));
843 } else { /* Unvoiced */
844 for (i = 0; i < FRAME_LEN; i++) {
845 *rseed = *rseed * 521 + 259;
846 out[i] = gain * *rseed >> 15;
848 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
853 * Perform IIR filtering.
855 * @param fir_coef FIR coefficients
856 * @param iir_coef IIR coefficients
857 * @param src source vector
858 * @param dest destination vector
860 static inline void iir_filter(int16_t *fir_coef, int16_t *iir_coef,
861 int16_t *src, int *dest)
865 for (m = 0; m < SUBFRAME_LEN; m++) {
867 for (n = 1; n <= LPC_ORDER; n++) {
868 filter -= fir_coef[n - 1] * src[m - n] -
869 iir_coef[n - 1] * (dest[m - n] >> 16);
872 dest[m] = av_clipl_int32((src[m] << 16) + (filter << 3) + (1 << 15));
877 * Adjust gain of postfiltered signal.
879 * @param p the context
880 * @param buf postfiltered output vector
881 * @param energy input energy coefficient
883 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
885 int num, denom, gain, bits1, bits2;
890 for (i = 0; i < SUBFRAME_LEN; i++) {
891 int temp = buf[i] >> 2;
893 denom = av_sat_dadd32(denom, temp);
897 bits1 = normalize_bits(num, 31);
898 bits2 = normalize_bits(denom, 31);
899 num = num << bits1 >> 1;
902 bits2 = 5 + bits1 - bits2;
903 bits2 = FFMAX(0, bits2);
905 gain = (num >> 1) / (denom >> 16);
906 gain = square_root(gain << 16 >> bits2);
911 for (i = 0; i < SUBFRAME_LEN; i++) {
912 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
913 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
919 * Perform formant filtering.
921 * @param p the context
922 * @param lpc quantized lpc coefficients
923 * @param buf input buffer
924 * @param dst output buffer
926 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
927 int16_t *buf, int16_t *dst)
929 int16_t filter_coef[2][LPC_ORDER];
930 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
933 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
934 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
936 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
937 for (k = 0; k < LPC_ORDER; k++) {
938 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
940 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
943 iir_filter(filter_coef[0], filter_coef[1], buf + i,
948 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(*p->fir_mem));
949 memcpy(p->iir_mem, filter_signal + FRAME_LEN,
950 LPC_ORDER * sizeof(*p->iir_mem));
953 signal_ptr = filter_signal + LPC_ORDER;
954 for (i = 0; i < SUBFRAMES; i++) {
960 scale = scale_vector(dst, buf, SUBFRAME_LEN);
962 /* Compute auto correlation coefficients */
963 auto_corr[0] = dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
964 auto_corr[1] = dot_product(dst, dst, SUBFRAME_LEN);
966 /* Compute reflection coefficient */
967 temp = auto_corr[1] >> 16;
969 temp = (auto_corr[0] >> 2) / temp;
971 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
972 temp = -p->reflection_coef >> 1 & ~3;
974 /* Compensation filter */
975 for (j = 0; j < SUBFRAME_LEN; j++) {
976 dst[j] = av_sat_dadd32(signal_ptr[j],
977 (signal_ptr[j - 1] >> 16) * temp) >> 16;
980 /* Compute normalized signal energy */
981 temp = 2 * scale + 4;
983 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
985 energy = auto_corr[1] >> temp;
987 gain_scale(p, dst, energy);
990 signal_ptr += SUBFRAME_LEN;
995 static int sid_gain_to_lsp_index(int gain)
999 else if (gain < 0x20)
1000 return gain - 8 << 7;
1002 return gain - 20 << 8;
1005 static inline int cng_rand(int *state, int base)
1007 *state = (*state * 521 + 259) & 0xFFFF;
1008 return (*state & 0x7FFF) * base >> 15;
1011 static int estimate_sid_gain(G723_1_Context *p)
1013 int i, shift, seg, seg2, t, val, val_add, x, y;
1015 shift = 16 - p->cur_gain * 2;
1017 t = p->sid_gain << shift;
1019 t = p->sid_gain >> -shift;
1020 x = t * cng_filt[0] >> 16;
1022 if (x >= cng_bseg[2])
1025 if (x >= cng_bseg[1]) {
1030 seg = (x >= cng_bseg[0]);
1032 seg2 = FFMIN(seg, 3);
1036 for (i = 0; i < shift; i++) {
1037 t = seg * 32 + (val << seg2);
1046 t = seg * 32 + (val << seg2);
1049 t = seg * 32 + (val + 1 << seg2);
1051 val = (seg2 - 1 << 4) + val;
1055 t = seg * 32 + (val - 1 << seg2);
1057 val = (seg2 - 1 << 4) + val;
1065 static void generate_noise(G723_1_Context *p)
1069 int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
1070 int tmp[SUBFRAME_LEN * 2];
1071 int16_t *vector_ptr;
1073 int b0, c, delta, x, shift;
1075 p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
1076 p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
1078 for (i = 0; i < SUBFRAMES; i++) {
1079 p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
1080 p->subframe[i].ad_cb_lag = cng_adaptive_cb_lag[i];
1083 for (i = 0; i < SUBFRAMES / 2; i++) {
1084 t = cng_rand(&p->cng_random_seed, 1 << 13);
1086 off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
1088 for (j = 0; j < 11; j++) {
1089 signs[i * 11 + j] = (t & 1) * 2 - 1 << 14;
1095 for (i = 0; i < SUBFRAMES; i++) {
1096 for (j = 0; j < SUBFRAME_LEN / 2; j++)
1098 t = SUBFRAME_LEN / 2;
1099 for (j = 0; j < pulses[i]; j++, idx++) {
1100 int idx2 = cng_rand(&p->cng_random_seed, t);
1102 pos[idx] = tmp[idx2] * 2 + off[i];
1103 tmp[idx2] = tmp[--t];
1107 vector_ptr = p->audio + LPC_ORDER;
1108 memcpy(vector_ptr, p->prev_excitation,
1109 PITCH_MAX * sizeof(*p->excitation));
1110 for (i = 0; i < SUBFRAMES; i += 2) {
1111 gen_acb_excitation(vector_ptr, vector_ptr,
1112 p->pitch_lag[i >> 1], &p->subframe[i],
1114 gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
1115 vector_ptr + SUBFRAME_LEN,
1116 p->pitch_lag[i >> 1], &p->subframe[i + 1],
1120 for (j = 0; j < SUBFRAME_LEN * 2; j++)
1121 t |= FFABS(vector_ptr[j]);
1122 t = FFMIN(t, 0x7FFF);
1126 shift = -10 + av_log2(t);
1132 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1133 t = vector_ptr[j] << -shift;
1138 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1139 t = vector_ptr[j] >> shift;
1146 for (j = 0; j < 11; j++)
1147 b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
1148 b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
1150 c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
1151 if (shift * 2 + 3 >= 0)
1152 c >>= shift * 2 + 3;
1154 c <<= -(shift * 2 + 3);
1155 c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
1157 delta = b0 * b0 * 2 - c;
1161 delta = square_root(delta);
1164 if (FFABS(t) < FFABS(x))
1172 x = av_clip(x, -10000, 10000);
1174 for (j = 0; j < 11; j++) {
1175 idx = (i / 2) * 11 + j;
1176 vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
1177 (x * signs[idx] >> 15));
1180 /* copy decoded data to serve as a history for the next decoded subframes */
1181 memcpy(vector_ptr + PITCH_MAX, vector_ptr,
1182 sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
1183 vector_ptr += SUBFRAME_LEN * 2;
1185 /* Save the excitation for the next frame */
1186 memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
1187 PITCH_MAX * sizeof(*p->excitation));
1190 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
1191 int *got_frame_ptr, AVPacket *avpkt)
1193 G723_1_Context *p = avctx->priv_data;
1194 const uint8_t *buf = avpkt->data;
1195 int buf_size = avpkt->size;
1196 int dec_mode = buf[0] & 3;
1198 PPFParam ppf[SUBFRAMES];
1199 int16_t cur_lsp[LPC_ORDER];
1200 int16_t lpc[SUBFRAMES * LPC_ORDER];
1201 int16_t acb_vector[SUBFRAME_LEN];
1203 int bad_frame = 0, i, j, ret;
1204 int16_t *audio = p->audio;
1206 if (buf_size < frame_size[dec_mode]) {
1208 av_log(avctx, AV_LOG_WARNING,
1209 "Expected %d bytes, got %d - skipping packet\n",
1210 frame_size[dec_mode], buf_size);
1215 if (unpack_bitstream(p, buf, buf_size) < 0) {
1217 if (p->past_frame_type == ACTIVE_FRAME)
1218 p->cur_frame_type = ACTIVE_FRAME;
1220 p->cur_frame_type = UNTRANSMITTED_FRAME;
1223 p->frame.nb_samples = FRAME_LEN;
1224 if ((ret = ff_get_buffer(avctx, &p->frame)) < 0) {
1225 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1229 out = (int16_t *)p->frame.data[0];
1231 if (p->cur_frame_type == ACTIVE_FRAME) {
1233 p->erased_frames = 0;
1234 else if (p->erased_frames != 3)
1237 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1238 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1240 /* Save the lsp_vector for the next frame */
1241 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1243 /* Generate the excitation for the frame */
1244 memcpy(p->excitation, p->prev_excitation,
1245 PITCH_MAX * sizeof(*p->excitation));
1246 if (!p->erased_frames) {
1247 int16_t *vector_ptr = p->excitation + PITCH_MAX;
1249 /* Update interpolation gain memory */
1250 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1251 p->subframe[3].amp_index) >> 1];
1252 for (i = 0; i < SUBFRAMES; i++) {
1253 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
1254 p->pitch_lag[i >> 1], i);
1255 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1256 p->pitch_lag[i >> 1], &p->subframe[i],
1258 /* Get the total excitation */
1259 for (j = 0; j < SUBFRAME_LEN; j++) {
1260 int v = av_clip_int16(vector_ptr[j] << 1);
1261 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1263 vector_ptr += SUBFRAME_LEN;
1266 vector_ptr = p->excitation + PITCH_MAX;
1268 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1269 &p->sid_gain, &p->cur_gain);
1271 /* Peform pitch postfiltering */
1272 if (p->postfilter) {
1274 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1275 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1276 ppf + j, p->cur_rate);
1278 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1279 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1281 vector_ptr + i + ppf[j].index,
1284 1 << 14, 15, SUBFRAME_LEN);
1286 audio = vector_ptr - LPC_ORDER;
1289 /* Save the excitation for the next frame */
1290 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1291 PITCH_MAX * sizeof(*p->excitation));
1293 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1294 if (p->erased_frames == 3) {
1296 memset(p->excitation, 0,
1297 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1298 memset(p->prev_excitation, 0,
1299 PITCH_MAX * sizeof(*p->excitation));
1300 memset(p->frame.data[0], 0,
1301 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1303 int16_t *buf = p->audio + LPC_ORDER;
1305 /* Regenerate frame */
1306 residual_interp(p->excitation, buf, p->interp_index,
1307 p->interp_gain, &p->random_seed);
1309 /* Save the excitation for the next frame */
1310 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1311 PITCH_MAX * sizeof(*p->excitation));
1314 p->cng_random_seed = CNG_RANDOM_SEED;
1316 if (p->cur_frame_type == SID_FRAME) {
1317 p->sid_gain = sid_gain_to_lsp_index(p->subframe[0].amp_index);
1318 inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
1319 } else if (p->past_frame_type == ACTIVE_FRAME) {
1320 p->sid_gain = estimate_sid_gain(p);
1323 if (p->past_frame_type == ACTIVE_FRAME)
1324 p->cur_gain = p->sid_gain;
1326 p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
1328 lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
1329 /* Save the lsp_vector for the next frame */
1330 memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1333 p->past_frame_type = p->cur_frame_type;
1335 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1336 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1337 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1338 audio + i, SUBFRAME_LEN, LPC_ORDER,
1340 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1342 if (p->postfilter) {
1343 formant_postfilter(p, lpc, p->audio, out);
1344 } else { // if output is not postfiltered it should be scaled by 2
1345 for (i = 0; i < FRAME_LEN; i++)
1346 out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1350 *(AVFrame *)data = p->frame;
1352 return frame_size[dec_mode];
1355 #define OFFSET(x) offsetof(G723_1_Context, x)
1356 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1358 static const AVOption options[] = {
1359 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1360 { .i64 = 1 }, 0, 1, AD },
1365 static const AVClass g723_1dec_class = {
1366 .class_name = "G.723.1 decoder",
1367 .item_name = av_default_item_name,
1369 .version = LIBAVUTIL_VERSION_INT,
1372 AVCodec ff_g723_1_decoder = {
1374 .type = AVMEDIA_TYPE_AUDIO,
1375 .id = AV_CODEC_ID_G723_1,
1376 .priv_data_size = sizeof(G723_1_Context),
1377 .init = g723_1_decode_init,
1378 .decode = g723_1_decode_frame,
1379 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1380 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1381 .priv_class = &g723_1dec_class,