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/audioconvert.h"
30 #include "libavutil/lzo.h"
31 #include "libavutil/opt.h"
34 #include "acelp_vectors.h"
35 #include "celp_filters.h"
36 #include "g723_1_data.h"
38 #define CNG_RANDOM_SEED 12345
44 ACTIVE_FRAME, ///< Active speech
45 SID_FRAME, ///< Silence Insertion Descriptor frame
55 * G723.1 unpacked data subframe
58 int ad_cb_lag; ///< adaptive codebook lag
68 * Pitch postfilter parameters
71 int index; ///< postfilter backward/forward lag
72 int16_t opt_gain; ///< optimal gain
73 int16_t sc_gain; ///< scaling gain
76 typedef struct g723_1_context {
80 G723_1_Subframe subframe[4];
81 enum FrameType cur_frame_type;
82 enum FrameType past_frame_type;
84 uint8_t lsp_index[LSP_BANDS];
88 int16_t prev_lsp[LPC_ORDER];
89 int16_t sid_lsp[LPC_ORDER];
90 int16_t prev_excitation[PITCH_MAX];
91 int16_t excitation[PITCH_MAX + FRAME_LEN + 4];
92 int16_t synth_mem[LPC_ORDER];
93 int16_t fir_mem[LPC_ORDER];
94 int iir_mem[LPC_ORDER];
106 int16_t audio[FRAME_LEN + LPC_ORDER + PITCH_MAX + 4];
109 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
111 G723_1_Context *p = avctx->priv_data;
113 avctx->channel_layout = AV_CH_LAYOUT_MONO;
114 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
116 avctx->sample_rate = 8000;
117 p->pf_gain = 1 << 12;
119 avcodec_get_frame_defaults(&p->frame);
120 avctx->coded_frame = &p->frame;
122 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
123 memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
125 p->cng_random_seed = CNG_RANDOM_SEED;
126 p->past_frame_type = SID_FRAME;
132 * Unpack the frame into parameters.
134 * @param p the context
135 * @param buf pointer to the input buffer
136 * @param buf_size size of the input buffer
138 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
143 int temp, info_bits, i;
145 init_get_bits(&gb, buf, buf_size * 8);
147 /* Extract frame type and rate info */
148 info_bits = get_bits(&gb, 2);
150 if (info_bits == 3) {
151 p->cur_frame_type = UNTRANSMITTED_FRAME;
155 /* Extract 24 bit lsp indices, 8 bit for each band */
156 p->lsp_index[2] = get_bits(&gb, 8);
157 p->lsp_index[1] = get_bits(&gb, 8);
158 p->lsp_index[0] = get_bits(&gb, 8);
160 if (info_bits == 2) {
161 p->cur_frame_type = SID_FRAME;
162 p->subframe[0].amp_index = get_bits(&gb, 6);
166 /* Extract the info common to both rates */
167 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
168 p->cur_frame_type = ACTIVE_FRAME;
170 p->pitch_lag[0] = get_bits(&gb, 7);
171 if (p->pitch_lag[0] > 123) /* test if forbidden code */
173 p->pitch_lag[0] += PITCH_MIN;
174 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
176 p->pitch_lag[1] = get_bits(&gb, 7);
177 if (p->pitch_lag[1] > 123)
179 p->pitch_lag[1] += PITCH_MIN;
180 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
181 p->subframe[0].ad_cb_lag = 1;
182 p->subframe[2].ad_cb_lag = 1;
184 for (i = 0; i < SUBFRAMES; i++) {
185 /* Extract combined gain */
186 temp = get_bits(&gb, 12);
188 p->subframe[i].dirac_train = 0;
189 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
190 p->subframe[i].dirac_train = temp >> 11;
194 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
195 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
196 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
203 p->subframe[0].grid_index = get_bits(&gb, 1);
204 p->subframe[1].grid_index = get_bits(&gb, 1);
205 p->subframe[2].grid_index = get_bits(&gb, 1);
206 p->subframe[3].grid_index = get_bits(&gb, 1);
208 if (p->cur_rate == RATE_6300) {
209 skip_bits(&gb, 1); /* skip reserved bit */
211 /* Compute pulse_pos index using the 13-bit combined position index */
212 temp = get_bits(&gb, 13);
213 p->subframe[0].pulse_pos = temp / 810;
215 temp -= p->subframe[0].pulse_pos * 810;
216 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
218 temp -= p->subframe[1].pulse_pos * 90;
219 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
220 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
222 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
224 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
226 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
228 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
231 p->subframe[0].pulse_sign = get_bits(&gb, 6);
232 p->subframe[1].pulse_sign = get_bits(&gb, 5);
233 p->subframe[2].pulse_sign = get_bits(&gb, 6);
234 p->subframe[3].pulse_sign = get_bits(&gb, 5);
235 } else { /* 5300 bps */
236 p->subframe[0].pulse_pos = get_bits(&gb, 12);
237 p->subframe[1].pulse_pos = get_bits(&gb, 12);
238 p->subframe[2].pulse_pos = get_bits(&gb, 12);
239 p->subframe[3].pulse_pos = get_bits(&gb, 12);
241 p->subframe[0].pulse_sign = get_bits(&gb, 4);
242 p->subframe[1].pulse_sign = get_bits(&gb, 4);
243 p->subframe[2].pulse_sign = get_bits(&gb, 4);
244 p->subframe[3].pulse_sign = get_bits(&gb, 4);
251 * Bitexact implementation of sqrt(val/2).
253 static int16_t square_root(int val)
256 int16_t exp = 0x4000;
259 for (i = 0; i < 14; i ++) {
260 int res_exp = res + exp;
261 if (val >= res_exp * res_exp << 1)
269 * Calculate the number of left-shifts required for normalizing the input.
271 * @param num input number
272 * @param width width of the input, 16 bits(0) / 32 bits(1)
274 static int normalize_bits(int num, int width)
276 return width - av_log2(num) - 1;
280 * Scale vector contents based on the largest of their absolutes.
282 static int scale_vector(int16_t *dst, const int16_t *vector, int length)
288 for (i = 0; i < length; i++)
289 max |= FFABS(vector[i]);
291 max = FFMIN(max, 0x7FFF);
292 bits = normalize_bits(max, 15);
294 for (i = 0; i < length; i++)
295 dst[i] = vector[i] << bits >> 3;
301 * Perform inverse quantization of LSP frequencies.
303 * @param cur_lsp the current LSP vector
304 * @param prev_lsp the previous LSP vector
305 * @param lsp_index VQ indices
306 * @param bad_frame bad frame flag
308 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
309 uint8_t *lsp_index, int bad_frame)
312 int i, j, temp, stable;
314 /* Check for frame erasure */
321 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
324 /* Get the VQ table entry corresponding to the transmitted index */
325 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
326 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
327 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
328 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
329 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
330 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
331 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
332 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
333 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
334 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
336 /* Add predicted vector & DC component to the previously quantized vector */
337 for (i = 0; i < LPC_ORDER; i++) {
338 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
339 cur_lsp[i] += dc_lsp[i] + temp;
342 for (i = 0; i < LPC_ORDER; i++) {
343 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
344 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
346 /* Stability check */
347 for (j = 1; j < LPC_ORDER; j++) {
348 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
351 cur_lsp[j - 1] -= temp;
356 for (j = 1; j < LPC_ORDER; j++) {
357 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
367 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
371 * Bitexact implementation of 2ab scaled by 1/2^16.
373 * @param a 32 bit multiplicand
374 * @param b 16 bit multiplier
376 #define MULL2(a, b) \
377 ((((a) >> 16) * (b) << 1) + (((a) & 0xffff) * (b) >> 15))
380 * Convert LSP frequencies to LPC coefficients.
382 * @param lpc buffer for LPC coefficients
384 static void lsp2lpc(int16_t *lpc)
386 int f1[LPC_ORDER / 2 + 1];
387 int f2[LPC_ORDER / 2 + 1];
390 /* Calculate negative cosine */
391 for (j = 0; j < LPC_ORDER; j++) {
392 int index = lpc[j] >> 7;
393 int offset = lpc[j] & 0x7f;
394 int temp1 = cos_tab[index] << 16;
395 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
396 ((offset << 8) + 0x80) << 1;
398 lpc[j] = -(av_sat_dadd32(1 << 15, temp1 + temp2) >> 16);
402 * Compute sum and difference polynomial coefficients
403 * (bitexact alternative to lsp2poly() in lsp.c)
405 /* Initialize with values in Q28 */
407 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
408 f1[2] = lpc[0] * lpc[2] + (2 << 28);
411 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
412 f2[2] = lpc[1] * lpc[3] + (2 << 28);
415 * Calculate and scale the coefficients by 1/2 in
416 * each iteration for a final scaling factor of Q25
418 for (i = 2; i < LPC_ORDER / 2; i++) {
419 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
420 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
422 for (j = i; j >= 2; j--) {
423 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
424 (f1[j] >> 1) + (f1[j - 2] >> 1);
425 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
426 (f2[j] >> 1) + (f2[j - 2] >> 1);
431 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
432 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
435 /* Convert polynomial coefficients to LPC coefficients */
436 for (i = 0; i < LPC_ORDER / 2; i++) {
437 int64_t ff1 = f1[i + 1] + f1[i];
438 int64_t ff2 = f2[i + 1] - f2[i];
440 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
441 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
447 * Quantize LSP frequencies by interpolation and convert them to
448 * the corresponding LPC coefficients.
450 * @param lpc buffer for LPC coefficients
451 * @param cur_lsp the current LSP vector
452 * @param prev_lsp the previous LSP vector
454 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
457 int16_t *lpc_ptr = lpc;
459 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
460 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
461 4096, 12288, 1 << 13, 14, LPC_ORDER);
462 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
463 8192, 8192, 1 << 13, 14, LPC_ORDER);
464 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
465 12288, 4096, 1 << 13, 14, LPC_ORDER);
466 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
468 for (i = 0; i < SUBFRAMES; i++) {
470 lpc_ptr += LPC_ORDER;
475 * Generate a train of dirac functions with period as pitch lag.
477 static void gen_dirac_train(int16_t *buf, int pitch_lag)
479 int16_t vector[SUBFRAME_LEN];
482 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
483 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
484 for (j = 0; j < SUBFRAME_LEN - i; j++)
485 buf[i + j] += vector[j];
490 * Generate fixed codebook excitation vector.
492 * @param vector decoded excitation vector
493 * @param subfrm current subframe
494 * @param cur_rate current bitrate
495 * @param pitch_lag closed loop pitch lag
496 * @param index current subframe index
498 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
499 enum Rate cur_rate, int pitch_lag, int index)
503 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
505 if (cur_rate == RATE_6300) {
506 if (subfrm->pulse_pos >= max_pos[index])
509 /* Decode amplitudes and positions */
510 j = PULSE_MAX - pulses[index];
511 temp = subfrm->pulse_pos;
512 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
513 temp -= combinatorial_table[j][i];
516 temp += combinatorial_table[j++][i];
517 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
518 vector[subfrm->grid_index + GRID_SIZE * i] =
519 -fixed_cb_gain[subfrm->amp_index];
521 vector[subfrm->grid_index + GRID_SIZE * i] =
522 fixed_cb_gain[subfrm->amp_index];
527 if (subfrm->dirac_train == 1)
528 gen_dirac_train(vector, pitch_lag);
529 } else { /* 5300 bps */
530 int cb_gain = fixed_cb_gain[subfrm->amp_index];
531 int cb_shift = subfrm->grid_index;
532 int cb_sign = subfrm->pulse_sign;
533 int cb_pos = subfrm->pulse_pos;
534 int offset, beta, lag;
536 for (i = 0; i < 8; i += 2) {
537 offset = ((cb_pos & 7) << 3) + cb_shift + i;
538 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
543 /* Enhance harmonic components */
544 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
545 subfrm->ad_cb_lag - 1;
546 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
548 if (lag < SUBFRAME_LEN - 2) {
549 for (i = lag; i < SUBFRAME_LEN; i++)
550 vector[i] += beta * vector[i - lag] >> 15;
556 * Get delayed contribution from the previous excitation vector.
558 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
560 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
563 residual[0] = prev_excitation[offset];
564 residual[1] = prev_excitation[offset + 1];
567 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
568 residual[i] = prev_excitation[offset + (i - 2) % lag];
571 static int dot_product(const int16_t *a, const int16_t *b, int length)
575 for (i = 0; i < length; i++) {
576 int prod = a[i] * b[i];
577 sum = av_sat_dadd32(sum, prod);
583 * Generate adaptive codebook excitation.
585 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
586 int pitch_lag, G723_1_Subframe *subfrm,
589 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
590 const int16_t *cb_ptr;
591 int lag = pitch_lag + subfrm->ad_cb_lag - 1;
596 get_residual(residual, prev_excitation, lag);
598 /* Select quantization table */
599 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2)
600 cb_ptr = adaptive_cb_gain85;
602 cb_ptr = adaptive_cb_gain170;
604 /* Calculate adaptive vector */
605 cb_ptr += subfrm->ad_cb_gain * 20;
606 for (i = 0; i < SUBFRAME_LEN; i++) {
607 sum = dot_product(residual + i, cb_ptr, PITCH_ORDER);
608 vector[i] = av_sat_dadd32(1 << 15, sum) >> 16;
613 * Estimate maximum auto-correlation around pitch lag.
615 * @param buf buffer with offset applied
616 * @param offset offset of the excitation vector
617 * @param ccr_max pointer to the maximum auto-correlation
618 * @param pitch_lag decoded pitch lag
619 * @param length length of autocorrelation
620 * @param dir forward lag(1) / backward lag(-1)
622 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
623 int pitch_lag, int length, int dir)
625 int limit, ccr, lag = 0;
628 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
630 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
632 limit = pitch_lag + 3;
634 for (i = pitch_lag - 3; i <= limit; i++) {
635 ccr = dot_product(buf, buf + dir * i, length);
637 if (ccr > *ccr_max) {
646 * Calculate pitch postfilter optimal and scaling gains.
648 * @param lag pitch postfilter forward/backward lag
649 * @param ppf pitch postfilter parameters
650 * @param cur_rate current bitrate
651 * @param tgt_eng target energy
652 * @param ccr cross-correlation
653 * @param res_eng residual energy
655 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
656 int tgt_eng, int ccr, int res_eng)
658 int pf_residual; /* square of postfiltered residual */
663 temp1 = tgt_eng * res_eng >> 1;
664 temp2 = ccr * ccr << 1;
667 if (ccr >= res_eng) {
668 ppf->opt_gain = ppf_gain_weight[cur_rate];
670 ppf->opt_gain = (ccr << 15) / res_eng *
671 ppf_gain_weight[cur_rate] >> 15;
673 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
674 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
675 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
676 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
678 if (tgt_eng >= pf_residual << 1) {
681 temp1 = (tgt_eng << 14) / pf_residual;
684 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
685 ppf->sc_gain = square_root(temp1 << 16);
688 ppf->sc_gain = 0x7fff;
691 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
695 * Calculate pitch postfilter parameters.
697 * @param p the context
698 * @param offset offset of the excitation vector
699 * @param pitch_lag decoded pitch lag
700 * @param ppf pitch postfilter parameters
701 * @param cur_rate current bitrate
703 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
704 PPFParam *ppf, enum Rate cur_rate)
713 * 1 - forward cross-correlation
714 * 2 - forward residual energy
715 * 3 - backward cross-correlation
716 * 4 - backward residual energy
718 int energy[5] = {0, 0, 0, 0, 0};
719 int16_t *buf = p->audio + LPC_ORDER + offset;
720 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
722 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
727 ppf->sc_gain = 0x7fff;
729 /* Case 0, Section 3.6 */
730 if (!back_lag && !fwd_lag)
733 /* Compute target energy */
734 energy[0] = dot_product(buf, buf, SUBFRAME_LEN);
736 /* Compute forward residual energy */
738 energy[2] = dot_product(buf + fwd_lag, buf + fwd_lag, SUBFRAME_LEN);
740 /* Compute backward residual energy */
742 energy[4] = dot_product(buf - back_lag, buf - back_lag, SUBFRAME_LEN);
744 /* Normalize and shorten */
746 for (i = 0; i < 5; i++)
747 temp1 = FFMAX(energy[i], temp1);
749 scale = normalize_bits(temp1, 31);
750 for (i = 0; i < 5; i++)
751 energy[i] = (energy[i] << scale) >> 16;
753 if (fwd_lag && !back_lag) { /* Case 1 */
754 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
756 } else if (!fwd_lag) { /* Case 2 */
757 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
759 } else { /* Case 3 */
762 * Select the largest of energy[1]^2/energy[2]
763 * and energy[3]^2/energy[4]
765 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
766 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
767 if (temp1 >= temp2) {
768 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
771 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
778 * Classify frames as voiced/unvoiced.
780 * @param p the context
781 * @param pitch_lag decoded pitch_lag
782 * @param exc_eng excitation energy estimation
783 * @param scale scaling factor of exc_eng
785 * @return residual interpolation index if voiced, 0 otherwise
787 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
788 int *exc_eng, int *scale)
790 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
791 int16_t *buf = p->audio + LPC_ORDER;
793 int index, ccr, tgt_eng, best_eng, temp;
795 *scale = scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
798 /* Compute maximum backward cross-correlation */
800 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
801 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
803 /* Compute target energy */
804 tgt_eng = dot_product(buf, buf, SUBFRAME_LEN * 2);
805 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
810 /* Compute best energy */
811 best_eng = dot_product(buf - index, buf - index, SUBFRAME_LEN * 2);
812 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
814 temp = best_eng * *exc_eng >> 3;
816 if (temp < ccr * ccr)
823 * Peform residual interpolation based on frame classification.
825 * @param buf decoded excitation vector
826 * @param out output vector
827 * @param lag decoded pitch lag
828 * @param gain interpolated gain
829 * @param rseed seed for random number generator
831 static void residual_interp(int16_t *buf, int16_t *out, int lag,
832 int gain, int *rseed)
835 if (lag) { /* Voiced */
836 int16_t *vector_ptr = buf + PITCH_MAX;
838 for (i = 0; i < lag; i++)
839 out[i] = vector_ptr[i - lag] * 3 >> 2;
840 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
841 (FRAME_LEN - lag) * sizeof(*out));
842 } else { /* Unvoiced */
843 for (i = 0; i < FRAME_LEN; i++) {
844 *rseed = *rseed * 521 + 259;
845 out[i] = gain * *rseed >> 15;
847 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
852 * Perform IIR filtering.
854 * @param fir_coef FIR coefficients
855 * @param iir_coef IIR coefficients
856 * @param src source vector
857 * @param dest destination vector
859 static inline void iir_filter(int16_t *fir_coef, int16_t *iir_coef,
860 int16_t *src, int *dest)
864 for (m = 0; m < SUBFRAME_LEN; m++) {
866 for (n = 1; n <= LPC_ORDER; n++) {
867 filter -= fir_coef[n - 1] * src[m - n] -
868 iir_coef[n - 1] * (dest[m - n] >> 16);
871 dest[m] = av_clipl_int32((src[m] << 16) + (filter << 3) + (1 << 15));
876 * Adjust gain of postfiltered signal.
878 * @param p the context
879 * @param buf postfiltered output vector
880 * @param energy input energy coefficient
882 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
884 int num, denom, gain, bits1, bits2;
889 for (i = 0; i < SUBFRAME_LEN; i++) {
890 int temp = buf[i] >> 2;
892 denom = av_sat_dadd32(denom, temp);
896 bits1 = normalize_bits(num, 31);
897 bits2 = normalize_bits(denom, 31);
898 num = num << bits1 >> 1;
901 bits2 = 5 + bits1 - bits2;
902 bits2 = FFMAX(0, bits2);
904 gain = (num >> 1) / (denom >> 16);
905 gain = square_root(gain << 16 >> bits2);
910 for (i = 0; i < SUBFRAME_LEN; i++) {
911 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
912 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
918 * Perform formant filtering.
920 * @param p the context
921 * @param lpc quantized lpc coefficients
922 * @param buf input buffer
923 * @param dst output buffer
925 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
926 int16_t *buf, int16_t *dst)
928 int16_t filter_coef[2][LPC_ORDER];
929 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
932 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
933 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
935 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
936 for (k = 0; k < LPC_ORDER; k++) {
937 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
939 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
942 iir_filter(filter_coef[0], filter_coef[1], buf + i,
947 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(*p->fir_mem));
948 memcpy(p->iir_mem, filter_signal + FRAME_LEN,
949 LPC_ORDER * sizeof(*p->iir_mem));
952 signal_ptr = filter_signal + LPC_ORDER;
953 for (i = 0; i < SUBFRAMES; i++) {
959 scale = scale_vector(dst, buf, SUBFRAME_LEN);
961 /* Compute auto correlation coefficients */
962 auto_corr[0] = dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
963 auto_corr[1] = dot_product(dst, dst, SUBFRAME_LEN);
965 /* Compute reflection coefficient */
966 temp = auto_corr[1] >> 16;
968 temp = (auto_corr[0] >> 2) / temp;
970 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
971 temp = -p->reflection_coef >> 1 & ~3;
973 /* Compensation filter */
974 for (j = 0; j < SUBFRAME_LEN; j++) {
975 dst[j] = av_sat_dadd32(signal_ptr[j],
976 (signal_ptr[j - 1] >> 16) * temp) >> 16;
979 /* Compute normalized signal energy */
980 temp = 2 * scale + 4;
982 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
984 energy = auto_corr[1] >> temp;
986 gain_scale(p, dst, energy);
989 signal_ptr += SUBFRAME_LEN;
994 static int sid_gain_to_lsp_index(int gain)
998 else if (gain < 0x20)
999 return gain - 8 << 7;
1001 return gain - 20 << 8;
1004 static inline int cng_rand(int *state, int base)
1006 *state = (*state * 521 + 259) & 0xFFFF;
1007 return (*state & 0x7FFF) * base >> 15;
1010 static int estimate_sid_gain(G723_1_Context *p)
1012 int i, shift, seg, seg2, t, val, val_add, x, y;
1014 shift = 16 - p->cur_gain * 2;
1016 t = p->sid_gain << shift;
1018 t = p->sid_gain >> -shift;
1019 x = t * cng_filt[0] >> 16;
1021 if (x >= cng_bseg[2])
1024 if (x >= cng_bseg[1]) {
1029 seg = (x >= cng_bseg[0]);
1031 seg2 = FFMIN(seg, 3);
1035 for (i = 0; i < shift; i++) {
1036 t = seg * 32 + (val << seg2);
1045 t = seg * 32 + (val << seg2);
1048 t = seg * 32 + (val + 1 << seg2);
1050 val = (seg2 - 1 << 4) + val;
1054 t = seg * 32 + (val - 1 << seg2);
1056 val = (seg2 - 1 << 4) + val;
1064 static void generate_noise(G723_1_Context *p)
1068 int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
1069 int tmp[SUBFRAME_LEN * 2];
1070 int16_t *vector_ptr;
1072 int b0, c, delta, x, shift;
1074 p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
1075 p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
1077 for (i = 0; i < SUBFRAMES; i++) {
1078 p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
1079 p->subframe[i].ad_cb_lag = cng_adaptive_cb_lag[i];
1082 for (i = 0; i < SUBFRAMES / 2; i++) {
1083 t = cng_rand(&p->cng_random_seed, 1 << 13);
1085 off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
1087 for (j = 0; j < 11; j++) {
1088 signs[i * 11 + j] = (t & 1) * 2 - 1 << 14;
1094 for (i = 0; i < SUBFRAMES; i++) {
1095 for (j = 0; j < SUBFRAME_LEN / 2; j++)
1097 t = SUBFRAME_LEN / 2;
1098 for (j = 0; j < pulses[i]; j++, idx++) {
1099 int idx2 = cng_rand(&p->cng_random_seed, t);
1101 pos[idx] = tmp[idx2] * 2 + off[i];
1102 tmp[idx2] = tmp[--t];
1106 vector_ptr = p->audio + LPC_ORDER;
1107 memcpy(vector_ptr, p->prev_excitation,
1108 PITCH_MAX * sizeof(*p->excitation));
1109 for (i = 0; i < SUBFRAMES; i += 2) {
1110 gen_acb_excitation(vector_ptr, vector_ptr,
1111 p->pitch_lag[i >> 1], &p->subframe[i],
1113 gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
1114 vector_ptr + SUBFRAME_LEN,
1115 p->pitch_lag[i >> 1], &p->subframe[i + 1],
1119 for (j = 0; j < SUBFRAME_LEN * 2; j++)
1120 t |= FFABS(vector_ptr[j]);
1121 t = FFMIN(t, 0x7FFF);
1125 shift = -10 + av_log2(t);
1131 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1132 t = vector_ptr[j] << -shift;
1137 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1138 t = vector_ptr[j] >> shift;
1145 for (j = 0; j < 11; j++)
1146 b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
1147 b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
1149 c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
1150 if (shift * 2 + 3 >= 0)
1151 c >>= shift * 2 + 3;
1153 c <<= -(shift * 2 + 3);
1154 c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
1156 delta = b0 * b0 * 2 - c;
1160 delta = square_root(delta);
1163 if (FFABS(t) < FFABS(x))
1171 x = av_clip(x, -10000, 10000);
1173 for (j = 0; j < 11; j++) {
1174 idx = (i / 2) * 11 + j;
1175 vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
1176 (x * signs[idx] >> 15));
1179 /* copy decoded data to serve as a history for the next decoded subframes */
1180 memcpy(vector_ptr + PITCH_MAX, vector_ptr,
1181 sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
1182 vector_ptr += SUBFRAME_LEN * 2;
1184 /* Save the excitation for the next frame */
1185 memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
1186 PITCH_MAX * sizeof(*p->excitation));
1189 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
1190 int *got_frame_ptr, AVPacket *avpkt)
1192 G723_1_Context *p = avctx->priv_data;
1193 const uint8_t *buf = avpkt->data;
1194 int buf_size = avpkt->size;
1195 int dec_mode = buf[0] & 3;
1197 PPFParam ppf[SUBFRAMES];
1198 int16_t cur_lsp[LPC_ORDER];
1199 int16_t lpc[SUBFRAMES * LPC_ORDER];
1200 int16_t acb_vector[SUBFRAME_LEN];
1202 int bad_frame = 0, i, j, ret;
1203 int16_t *audio = p->audio;
1205 if (buf_size < frame_size[dec_mode]) {
1207 av_log(avctx, AV_LOG_WARNING,
1208 "Expected %d bytes, got %d - skipping packet\n",
1209 frame_size[dec_mode], buf_size);
1214 if (unpack_bitstream(p, buf, buf_size) < 0) {
1216 if (p->past_frame_type == ACTIVE_FRAME)
1217 p->cur_frame_type = ACTIVE_FRAME;
1219 p->cur_frame_type = UNTRANSMITTED_FRAME;
1222 p->frame.nb_samples = FRAME_LEN;
1223 if ((ret = avctx->get_buffer(avctx, &p->frame)) < 0) {
1224 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1228 out = (int16_t *)p->frame.data[0];
1230 if (p->cur_frame_type == ACTIVE_FRAME) {
1232 p->erased_frames = 0;
1233 else if (p->erased_frames != 3)
1236 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1237 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1239 /* Save the lsp_vector for the next frame */
1240 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1242 /* Generate the excitation for the frame */
1243 memcpy(p->excitation, p->prev_excitation,
1244 PITCH_MAX * sizeof(*p->excitation));
1245 if (!p->erased_frames) {
1246 int16_t *vector_ptr = p->excitation + PITCH_MAX;
1248 /* Update interpolation gain memory */
1249 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1250 p->subframe[3].amp_index) >> 1];
1251 for (i = 0; i < SUBFRAMES; i++) {
1252 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
1253 p->pitch_lag[i >> 1], i);
1254 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1255 p->pitch_lag[i >> 1], &p->subframe[i],
1257 /* Get the total excitation */
1258 for (j = 0; j < SUBFRAME_LEN; j++) {
1259 int v = av_clip_int16(vector_ptr[j] << 1);
1260 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1262 vector_ptr += SUBFRAME_LEN;
1265 vector_ptr = p->excitation + PITCH_MAX;
1267 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1268 &p->sid_gain, &p->cur_gain);
1270 /* Peform pitch postfiltering */
1271 if (p->postfilter) {
1273 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1274 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1275 ppf + j, p->cur_rate);
1277 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1278 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1280 vector_ptr + i + ppf[j].index,
1283 1 << 14, 15, SUBFRAME_LEN);
1285 audio = vector_ptr - LPC_ORDER;
1288 /* Save the excitation for the next frame */
1289 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1290 PITCH_MAX * sizeof(*p->excitation));
1292 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1293 if (p->erased_frames == 3) {
1295 memset(p->excitation, 0,
1296 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1297 memset(p->prev_excitation, 0,
1298 PITCH_MAX * sizeof(*p->excitation));
1299 memset(p->frame.data[0], 0,
1300 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1302 int16_t *buf = p->audio + LPC_ORDER;
1304 /* Regenerate frame */
1305 residual_interp(p->excitation, buf, p->interp_index,
1306 p->interp_gain, &p->random_seed);
1308 /* Save the excitation for the next frame */
1309 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1310 PITCH_MAX * sizeof(*p->excitation));
1313 p->cng_random_seed = CNG_RANDOM_SEED;
1315 if (p->cur_frame_type == SID_FRAME) {
1316 p->sid_gain = sid_gain_to_lsp_index(p->subframe[0].amp_index);
1317 inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
1318 } else if (p->past_frame_type == ACTIVE_FRAME) {
1319 p->sid_gain = estimate_sid_gain(p);
1322 if (p->past_frame_type == ACTIVE_FRAME)
1323 p->cur_gain = p->sid_gain;
1325 p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
1327 lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
1328 /* Save the lsp_vector for the next frame */
1329 memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1332 p->past_frame_type = p->cur_frame_type;
1334 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1335 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1336 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1337 audio + i, SUBFRAME_LEN, LPC_ORDER,
1339 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1341 if (p->postfilter) {
1342 formant_postfilter(p, lpc, p->audio, out);
1343 } else { // if output is not postfiltered it should be scaled by 2
1344 for (i = 0; i < FRAME_LEN; i++)
1345 out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1349 *(AVFrame *)data = p->frame;
1351 return frame_size[dec_mode];
1354 #define OFFSET(x) offsetof(G723_1_Context, x)
1355 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1357 static const AVOption options[] = {
1358 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1359 { .i64 = 1 }, 0, 1, AD },
1364 static const AVClass g723_1dec_class = {
1365 .class_name = "G.723.1 decoder",
1366 .item_name = av_default_item_name,
1368 .version = LIBAVUTIL_VERSION_INT,
1371 AVCodec ff_g723_1_decoder = {
1373 .type = AVMEDIA_TYPE_AUDIO,
1374 .id = AV_CODEC_ID_G723_1,
1375 .priv_data_size = sizeof(G723_1_Context),
1376 .init = g723_1_decode_init,
1377 .decode = g723_1_decode_frame,
1378 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1379 .capabilities = CODEC_CAP_SUBFRAMES,
1380 .priv_class = &g723_1dec_class,