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"
42 ACTIVE_FRAME, ///< Active speech
43 SID_FRAME, ///< Silence Insertion Descriptor frame
53 * G723.1 unpacked data subframe
56 int ad_cb_lag; ///< adaptive codebook lag
66 * Pitch postfilter parameters
69 int index; ///< postfilter backward/forward lag
70 int16_t opt_gain; ///< optimal gain
71 int16_t sc_gain; ///< scaling gain
74 typedef struct g723_1_context {
78 G723_1_Subframe subframe[4];
79 enum FrameType cur_frame_type;
80 enum FrameType past_frame_type;
82 uint8_t lsp_index[LSP_BANDS];
86 int16_t prev_lsp[LPC_ORDER];
87 int16_t prev_excitation[PITCH_MAX];
88 int16_t excitation[PITCH_MAX + FRAME_LEN + 4];
89 int16_t synth_mem[LPC_ORDER];
90 int16_t fir_mem[LPC_ORDER];
91 int iir_mem[LPC_ORDER];
102 int16_t audio[FRAME_LEN + LPC_ORDER + PITCH_MAX];
105 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
107 G723_1_Context *p = avctx->priv_data;
109 avctx->channel_layout = AV_CH_LAYOUT_MONO;
110 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
112 avctx->sample_rate = 8000;
113 p->pf_gain = 1 << 12;
115 avcodec_get_frame_defaults(&p->frame);
116 avctx->coded_frame = &p->frame;
118 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
124 * Unpack the frame into parameters.
126 * @param p the context
127 * @param buf pointer to the input buffer
128 * @param buf_size size of the input buffer
130 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
135 int temp, info_bits, i;
137 init_get_bits(&gb, buf, buf_size * 8);
139 /* Extract frame type and rate info */
140 info_bits = get_bits(&gb, 2);
142 if (info_bits == 3) {
143 p->cur_frame_type = UNTRANSMITTED_FRAME;
147 /* Extract 24 bit lsp indices, 8 bit for each band */
148 p->lsp_index[2] = get_bits(&gb, 8);
149 p->lsp_index[1] = get_bits(&gb, 8);
150 p->lsp_index[0] = get_bits(&gb, 8);
152 if (info_bits == 2) {
153 p->cur_frame_type = SID_FRAME;
154 p->subframe[0].amp_index = get_bits(&gb, 6);
158 /* Extract the info common to both rates */
159 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
160 p->cur_frame_type = ACTIVE_FRAME;
162 p->pitch_lag[0] = get_bits(&gb, 7);
163 if (p->pitch_lag[0] > 123) /* test if forbidden code */
165 p->pitch_lag[0] += PITCH_MIN;
166 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
168 p->pitch_lag[1] = get_bits(&gb, 7);
169 if (p->pitch_lag[1] > 123)
171 p->pitch_lag[1] += PITCH_MIN;
172 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
173 p->subframe[0].ad_cb_lag = 1;
174 p->subframe[2].ad_cb_lag = 1;
176 for (i = 0; i < SUBFRAMES; i++) {
177 /* Extract combined gain */
178 temp = get_bits(&gb, 12);
180 p->subframe[i].dirac_train = 0;
181 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
182 p->subframe[i].dirac_train = temp >> 11;
186 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
187 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
188 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
195 p->subframe[0].grid_index = get_bits(&gb, 1);
196 p->subframe[1].grid_index = get_bits(&gb, 1);
197 p->subframe[2].grid_index = get_bits(&gb, 1);
198 p->subframe[3].grid_index = get_bits(&gb, 1);
200 if (p->cur_rate == RATE_6300) {
201 skip_bits(&gb, 1); /* skip reserved bit */
203 /* Compute pulse_pos index using the 13-bit combined position index */
204 temp = get_bits(&gb, 13);
205 p->subframe[0].pulse_pos = temp / 810;
207 temp -= p->subframe[0].pulse_pos * 810;
208 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
210 temp -= p->subframe[1].pulse_pos * 90;
211 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
212 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
214 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
216 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
218 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
220 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
223 p->subframe[0].pulse_sign = get_bits(&gb, 6);
224 p->subframe[1].pulse_sign = get_bits(&gb, 5);
225 p->subframe[2].pulse_sign = get_bits(&gb, 6);
226 p->subframe[3].pulse_sign = get_bits(&gb, 5);
227 } else { /* 5300 bps */
228 p->subframe[0].pulse_pos = get_bits(&gb, 12);
229 p->subframe[1].pulse_pos = get_bits(&gb, 12);
230 p->subframe[2].pulse_pos = get_bits(&gb, 12);
231 p->subframe[3].pulse_pos = get_bits(&gb, 12);
233 p->subframe[0].pulse_sign = get_bits(&gb, 4);
234 p->subframe[1].pulse_sign = get_bits(&gb, 4);
235 p->subframe[2].pulse_sign = get_bits(&gb, 4);
236 p->subframe[3].pulse_sign = get_bits(&gb, 4);
243 * Bitexact implementation of sqrt(val/2).
245 static int16_t square_root(int val)
248 int16_t exp = 0x4000;
251 for (i = 0; i < 14; i ++) {
252 int res_exp = res + exp;
253 if (val >= res_exp * res_exp << 1)
261 * Calculate the number of left-shifts required for normalizing the input.
263 * @param num input number
264 * @param width width of the input, 16 bits(0) / 32 bits(1)
266 static int normalize_bits(int num, int width)
268 return width - av_log2(num) - 1;
272 * Scale vector contents based on the largest of their absolutes.
274 static int scale_vector(int16_t *dst, const int16_t *vector, int length)
280 for (i = 0; i < length; i++)
281 max |= FFABS(vector[i]);
283 max = FFMIN(max, 0x7FFF);
284 bits = normalize_bits(max, 15);
286 for (i = 0; i < length; i++)
287 dst[i] = vector[i] << bits >> 3;
293 * Perform inverse quantization of LSP frequencies.
295 * @param cur_lsp the current LSP vector
296 * @param prev_lsp the previous LSP vector
297 * @param lsp_index VQ indices
298 * @param bad_frame bad frame flag
300 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
301 uint8_t *lsp_index, int bad_frame)
304 int i, j, temp, stable;
306 /* Check for frame erasure */
313 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
316 /* Get the VQ table entry corresponding to the transmitted index */
317 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
318 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
319 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
320 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
321 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
322 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
323 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
324 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
325 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
326 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
328 /* Add predicted vector & DC component to the previously quantized vector */
329 for (i = 0; i < LPC_ORDER; i++) {
330 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
331 cur_lsp[i] += dc_lsp[i] + temp;
334 for (i = 0; i < LPC_ORDER; i++) {
335 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
336 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
338 /* Stability check */
339 for (j = 1; j < LPC_ORDER; j++) {
340 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
343 cur_lsp[j - 1] -= temp;
348 for (j = 1; j < LPC_ORDER; j++) {
349 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
359 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
363 * Bitexact implementation of 2ab scaled by 1/2^16.
365 * @param a 32 bit multiplicand
366 * @param b 16 bit multiplier
368 #define MULL2(a, b) \
369 ((((a) >> 16) * (b) << 1) + (((a) & 0xffff) * (b) >> 15))
372 * Convert LSP frequencies to LPC coefficients.
374 * @param lpc buffer for LPC coefficients
376 static void lsp2lpc(int16_t *lpc)
378 int f1[LPC_ORDER / 2 + 1];
379 int f2[LPC_ORDER / 2 + 1];
382 /* Calculate negative cosine */
383 for (j = 0; j < LPC_ORDER; j++) {
384 int index = lpc[j] >> 7;
385 int offset = lpc[j] & 0x7f;
386 int temp1 = cos_tab[index] << 16;
387 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
388 ((offset << 8) + 0x80) << 1;
390 lpc[j] = -(av_sat_dadd32(1 << 15, temp1 + temp2) >> 16);
394 * Compute sum and difference polynomial coefficients
395 * (bitexact alternative to lsp2poly() in lsp.c)
397 /* Initialize with values in Q28 */
399 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
400 f1[2] = lpc[0] * lpc[2] + (2 << 28);
403 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
404 f2[2] = lpc[1] * lpc[3] + (2 << 28);
407 * Calculate and scale the coefficients by 1/2 in
408 * each iteration for a final scaling factor of Q25
410 for (i = 2; i < LPC_ORDER / 2; i++) {
411 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
412 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
414 for (j = i; j >= 2; j--) {
415 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
416 (f1[j] >> 1) + (f1[j - 2] >> 1);
417 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
418 (f2[j] >> 1) + (f2[j - 2] >> 1);
423 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
424 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
427 /* Convert polynomial coefficients to LPC coefficients */
428 for (i = 0; i < LPC_ORDER / 2; i++) {
429 int64_t ff1 = f1[i + 1] + f1[i];
430 int64_t ff2 = f2[i + 1] - f2[i];
432 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
433 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
439 * Quantize LSP frequencies by interpolation and convert them to
440 * the corresponding LPC coefficients.
442 * @param lpc buffer for LPC coefficients
443 * @param cur_lsp the current LSP vector
444 * @param prev_lsp the previous LSP vector
446 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
449 int16_t *lpc_ptr = lpc;
451 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
452 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
453 4096, 12288, 1 << 13, 14, LPC_ORDER);
454 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
455 8192, 8192, 1 << 13, 14, LPC_ORDER);
456 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
457 12288, 4096, 1 << 13, 14, LPC_ORDER);
458 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
460 for (i = 0; i < SUBFRAMES; i++) {
462 lpc_ptr += LPC_ORDER;
467 * Generate a train of dirac functions with period as pitch lag.
469 static void gen_dirac_train(int16_t *buf, int pitch_lag)
471 int16_t vector[SUBFRAME_LEN];
474 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
475 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
476 for (j = 0; j < SUBFRAME_LEN - i; j++)
477 buf[i + j] += vector[j];
482 * Generate fixed codebook excitation vector.
484 * @param vector decoded excitation vector
485 * @param subfrm current subframe
486 * @param cur_rate current bitrate
487 * @param pitch_lag closed loop pitch lag
488 * @param index current subframe index
490 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
491 enum Rate cur_rate, int pitch_lag, int index)
495 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
497 if (cur_rate == RATE_6300) {
498 if (subfrm->pulse_pos >= max_pos[index])
501 /* Decode amplitudes and positions */
502 j = PULSE_MAX - pulses[index];
503 temp = subfrm->pulse_pos;
504 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
505 temp -= combinatorial_table[j][i];
508 temp += combinatorial_table[j++][i];
509 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
510 vector[subfrm->grid_index + GRID_SIZE * i] =
511 -fixed_cb_gain[subfrm->amp_index];
513 vector[subfrm->grid_index + GRID_SIZE * i] =
514 fixed_cb_gain[subfrm->amp_index];
519 if (subfrm->dirac_train == 1)
520 gen_dirac_train(vector, pitch_lag);
521 } else { /* 5300 bps */
522 int cb_gain = fixed_cb_gain[subfrm->amp_index];
523 int cb_shift = subfrm->grid_index;
524 int cb_sign = subfrm->pulse_sign;
525 int cb_pos = subfrm->pulse_pos;
526 int offset, beta, lag;
528 for (i = 0; i < 8; i += 2) {
529 offset = ((cb_pos & 7) << 3) + cb_shift + i;
530 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
535 /* Enhance harmonic components */
536 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
537 subfrm->ad_cb_lag - 1;
538 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
540 if (lag < SUBFRAME_LEN - 2) {
541 for (i = lag; i < SUBFRAME_LEN; i++)
542 vector[i] += beta * vector[i - lag] >> 15;
548 * Get delayed contribution from the previous excitation vector.
550 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
552 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
555 residual[0] = prev_excitation[offset];
556 residual[1] = prev_excitation[offset + 1];
559 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
560 residual[i] = prev_excitation[offset + (i - 2) % lag];
563 static int dot_product(const int16_t *a, const int16_t *b, int length)
567 for (i = 0; i < length; i++) {
568 int prod = a[i] * b[i];
569 sum = av_sat_dadd32(sum, prod);
575 * Generate adaptive codebook excitation.
577 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
578 int pitch_lag, G723_1_Subframe *subfrm,
581 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
582 const int16_t *cb_ptr;
583 int lag = pitch_lag + subfrm->ad_cb_lag - 1;
588 get_residual(residual, prev_excitation, lag);
590 /* Select quantization table */
591 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2)
592 cb_ptr = adaptive_cb_gain85;
594 cb_ptr = adaptive_cb_gain170;
596 /* Calculate adaptive vector */
597 cb_ptr += subfrm->ad_cb_gain * 20;
598 for (i = 0; i < SUBFRAME_LEN; i++) {
599 sum = dot_product(residual + i, cb_ptr, PITCH_ORDER);
600 vector[i] = av_sat_dadd32(1 << 15, sum) >> 16;
605 * Estimate maximum auto-correlation around pitch lag.
607 * @param buf buffer with offset applied
608 * @param offset offset of the excitation vector
609 * @param ccr_max pointer to the maximum auto-correlation
610 * @param pitch_lag decoded pitch lag
611 * @param length length of autocorrelation
612 * @param dir forward lag(1) / backward lag(-1)
614 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
615 int pitch_lag, int length, int dir)
617 int limit, ccr, lag = 0;
620 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
622 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
624 limit = pitch_lag + 3;
626 for (i = pitch_lag - 3; i <= limit; i++) {
627 ccr = dot_product(buf, buf + dir * i, length);
629 if (ccr > *ccr_max) {
638 * Calculate pitch postfilter optimal and scaling gains.
640 * @param lag pitch postfilter forward/backward lag
641 * @param ppf pitch postfilter parameters
642 * @param cur_rate current bitrate
643 * @param tgt_eng target energy
644 * @param ccr cross-correlation
645 * @param res_eng residual energy
647 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
648 int tgt_eng, int ccr, int res_eng)
650 int pf_residual; /* square of postfiltered residual */
655 temp1 = tgt_eng * res_eng >> 1;
656 temp2 = ccr * ccr << 1;
659 if (ccr >= res_eng) {
660 ppf->opt_gain = ppf_gain_weight[cur_rate];
662 ppf->opt_gain = (ccr << 15) / res_eng *
663 ppf_gain_weight[cur_rate] >> 15;
665 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
666 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
667 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
668 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
670 if (tgt_eng >= pf_residual << 1) {
673 temp1 = (tgt_eng << 14) / pf_residual;
676 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
677 ppf->sc_gain = square_root(temp1 << 16);
680 ppf->sc_gain = 0x7fff;
683 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
687 * Calculate pitch postfilter parameters.
689 * @param p the context
690 * @param offset offset of the excitation vector
691 * @param pitch_lag decoded pitch lag
692 * @param ppf pitch postfilter parameters
693 * @param cur_rate current bitrate
695 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
696 PPFParam *ppf, enum Rate cur_rate)
705 * 1 - forward cross-correlation
706 * 2 - forward residual energy
707 * 3 - backward cross-correlation
708 * 4 - backward residual energy
710 int energy[5] = {0, 0, 0, 0, 0};
711 int16_t *buf = p->audio + LPC_ORDER + offset;
712 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
714 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
719 ppf->sc_gain = 0x7fff;
721 /* Case 0, Section 3.6 */
722 if (!back_lag && !fwd_lag)
725 /* Compute target energy */
726 energy[0] = dot_product(buf, buf, SUBFRAME_LEN);
728 /* Compute forward residual energy */
730 energy[2] = dot_product(buf + fwd_lag, buf + fwd_lag, SUBFRAME_LEN);
732 /* Compute backward residual energy */
734 energy[4] = dot_product(buf - back_lag, buf - back_lag, SUBFRAME_LEN);
736 /* Normalize and shorten */
738 for (i = 0; i < 5; i++)
739 temp1 = FFMAX(energy[i], temp1);
741 scale = normalize_bits(temp1, 31);
742 for (i = 0; i < 5; i++)
743 energy[i] = (energy[i] << scale) >> 16;
745 if (fwd_lag && !back_lag) { /* Case 1 */
746 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
748 } else if (!fwd_lag) { /* Case 2 */
749 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
751 } else { /* Case 3 */
754 * Select the largest of energy[1]^2/energy[2]
755 * and energy[3]^2/energy[4]
757 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
758 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
759 if (temp1 >= temp2) {
760 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
763 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
770 * Classify frames as voiced/unvoiced.
772 * @param p the context
773 * @param pitch_lag decoded pitch_lag
774 * @param exc_eng excitation energy estimation
775 * @param scale scaling factor of exc_eng
777 * @return residual interpolation index if voiced, 0 otherwise
779 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
780 int *exc_eng, int *scale)
782 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
783 int16_t *buf = p->audio + LPC_ORDER;
785 int index, ccr, tgt_eng, best_eng, temp;
787 *scale = scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
790 /* Compute maximum backward cross-correlation */
792 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
793 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
795 /* Compute target energy */
796 tgt_eng = dot_product(buf, buf, SUBFRAME_LEN * 2);
797 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
802 /* Compute best energy */
803 best_eng = dot_product(buf - index, buf - index, SUBFRAME_LEN * 2);
804 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
806 temp = best_eng * *exc_eng >> 3;
808 if (temp < ccr * ccr)
815 * Peform residual interpolation based on frame classification.
817 * @param buf decoded excitation vector
818 * @param out output vector
819 * @param lag decoded pitch lag
820 * @param gain interpolated gain
821 * @param rseed seed for random number generator
823 static void residual_interp(int16_t *buf, int16_t *out, int lag,
824 int gain, int *rseed)
827 if (lag) { /* Voiced */
828 int16_t *vector_ptr = buf + PITCH_MAX;
830 for (i = 0; i < lag; i++)
831 out[i] = vector_ptr[i - lag] * 3 >> 2;
832 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
833 (FRAME_LEN - lag) * sizeof(*out));
834 } else { /* Unvoiced */
835 for (i = 0; i < FRAME_LEN; i++) {
836 *rseed = *rseed * 521 + 259;
837 out[i] = gain * *rseed >> 15;
839 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
844 * Perform IIR filtering.
846 * @param fir_coef FIR coefficients
847 * @param iir_coef IIR coefficients
848 * @param src source vector
849 * @param dest destination vector
851 static inline void iir_filter(int16_t *fir_coef, int16_t *iir_coef,
852 int16_t *src, int *dest)
856 for (m = 0; m < SUBFRAME_LEN; m++) {
858 for (n = 1; n <= LPC_ORDER; n++) {
859 filter -= fir_coef[n - 1] * src[m - n] -
860 iir_coef[n - 1] * (dest[m - n] >> 16);
863 dest[m] = av_clipl_int32((src[m] << 16) + (filter << 3) + (1 << 15));
868 * Adjust gain of postfiltered signal.
870 * @param p the context
871 * @param buf postfiltered output vector
872 * @param energy input energy coefficient
874 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
876 int num, denom, gain, bits1, bits2;
881 for (i = 0; i < SUBFRAME_LEN; i++) {
882 int temp = buf[i] >> 2;
884 denom = av_sat_dadd32(denom, temp);
888 bits1 = normalize_bits(num, 31);
889 bits2 = normalize_bits(denom, 31);
890 num = num << bits1 >> 1;
893 bits2 = 5 + bits1 - bits2;
894 bits2 = FFMAX(0, bits2);
896 gain = (num >> 1) / (denom >> 16);
897 gain = square_root(gain << 16 >> bits2);
902 for (i = 0; i < SUBFRAME_LEN; i++) {
903 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
904 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
910 * Perform formant filtering.
912 * @param p the context
913 * @param lpc quantized lpc coefficients
914 * @param buf input buffer
915 * @param dst output buffer
917 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
918 int16_t *buf, int16_t *dst)
920 int16_t filter_coef[2][LPC_ORDER];
921 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
924 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
925 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
927 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
928 for (k = 0; k < LPC_ORDER; k++) {
929 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
931 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
934 iir_filter(filter_coef[0], filter_coef[1], buf + i,
939 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(*p->fir_mem));
940 memcpy(p->iir_mem, filter_signal + FRAME_LEN,
941 LPC_ORDER * sizeof(*p->iir_mem));
944 signal_ptr = filter_signal + LPC_ORDER;
945 for (i = 0; i < SUBFRAMES; i++) {
951 scale = scale_vector(dst, buf, SUBFRAME_LEN);
953 /* Compute auto correlation coefficients */
954 auto_corr[0] = dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
955 auto_corr[1] = dot_product(dst, dst, SUBFRAME_LEN);
957 /* Compute reflection coefficient */
958 temp = auto_corr[1] >> 16;
960 temp = (auto_corr[0] >> 2) / temp;
962 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
963 temp = -p->reflection_coef >> 1 & ~3;
965 /* Compensation filter */
966 for (j = 0; j < SUBFRAME_LEN; j++) {
967 dst[j] = av_sat_dadd32(signal_ptr[j],
968 (signal_ptr[j - 1] >> 16) * temp) >> 16;
971 /* Compute normalized signal energy */
972 temp = 2 * scale + 4;
974 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
976 energy = auto_corr[1] >> temp;
978 gain_scale(p, dst, energy);
981 signal_ptr += SUBFRAME_LEN;
986 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
987 int *got_frame_ptr, AVPacket *avpkt)
989 G723_1_Context *p = avctx->priv_data;
990 const uint8_t *buf = avpkt->data;
991 int buf_size = avpkt->size;
992 int dec_mode = buf[0] & 3;
994 PPFParam ppf[SUBFRAMES];
995 int16_t cur_lsp[LPC_ORDER];
996 int16_t lpc[SUBFRAMES * LPC_ORDER];
997 int16_t acb_vector[SUBFRAME_LEN];
999 int bad_frame = 0, i, j, ret;
1000 int16_t *audio = p->audio;
1002 if (buf_size < frame_size[dec_mode]) {
1004 av_log(avctx, AV_LOG_WARNING,
1005 "Expected %d bytes, got %d - skipping packet\n",
1006 frame_size[dec_mode], buf_size);
1011 if (unpack_bitstream(p, buf, buf_size) < 0) {
1013 if (p->past_frame_type == ACTIVE_FRAME)
1014 p->cur_frame_type = ACTIVE_FRAME;
1016 p->cur_frame_type = UNTRANSMITTED_FRAME;
1019 p->frame.nb_samples = FRAME_LEN;
1020 if ((ret = avctx->get_buffer(avctx, &p->frame)) < 0) {
1021 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1025 out = (int16_t *)p->frame.data[0];
1027 if (p->cur_frame_type == ACTIVE_FRAME) {
1029 p->erased_frames = 0;
1030 else if (p->erased_frames != 3)
1033 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1034 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1036 /* Save the lsp_vector for the next frame */
1037 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1039 /* Generate the excitation for the frame */
1040 memcpy(p->excitation, p->prev_excitation,
1041 PITCH_MAX * sizeof(*p->excitation));
1042 if (!p->erased_frames) {
1043 int16_t *vector_ptr = p->excitation + PITCH_MAX;
1045 /* Update interpolation gain memory */
1046 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1047 p->subframe[3].amp_index) >> 1];
1048 for (i = 0; i < SUBFRAMES; i++) {
1049 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
1050 p->pitch_lag[i >> 1], i);
1051 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1052 p->pitch_lag[i >> 1], &p->subframe[i],
1054 /* Get the total excitation */
1055 for (j = 0; j < SUBFRAME_LEN; j++) {
1056 int v = av_clip_int16(vector_ptr[j] << 1);
1057 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1059 vector_ptr += SUBFRAME_LEN;
1062 vector_ptr = p->excitation + PITCH_MAX;
1064 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1065 &p->sid_gain, &p->cur_gain);
1067 /* Peform pitch postfiltering */
1068 if (p->postfilter) {
1070 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1071 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1072 ppf + j, p->cur_rate);
1074 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1075 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1077 vector_ptr + i + ppf[j].index,
1080 1 << 14, 15, SUBFRAME_LEN);
1082 audio = vector_ptr - LPC_ORDER;
1085 /* Save the excitation for the next frame */
1086 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1087 PITCH_MAX * sizeof(*p->excitation));
1089 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1090 if (p->erased_frames == 3) {
1092 memset(p->excitation, 0,
1093 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1094 memset(p->prev_excitation, 0,
1095 PITCH_MAX * sizeof(*p->excitation));
1096 memset(p->frame.data[0], 0,
1097 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1099 int16_t *buf = p->audio + LPC_ORDER;
1101 /* Regenerate frame */
1102 residual_interp(p->excitation, buf, p->interp_index,
1103 p->interp_gain, &p->random_seed);
1105 /* Save the excitation for the next frame */
1106 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1107 PITCH_MAX * sizeof(*p->excitation));
1111 memset(out, 0, FRAME_LEN * 2);
1112 av_log(avctx, AV_LOG_WARNING,
1113 "G.723.1: Comfort noise generation not supported yet\n");
1116 *(AVFrame *)data = p->frame;
1117 return frame_size[dec_mode];
1120 p->past_frame_type = p->cur_frame_type;
1122 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1123 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1124 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1125 audio + i, SUBFRAME_LEN, LPC_ORDER,
1127 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1129 if (p->postfilter) {
1130 formant_postfilter(p, lpc, p->audio, out);
1131 } else { // if output is not postfiltered it should be scaled by 2
1132 for (i = 0; i < FRAME_LEN; i++)
1133 out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1137 *(AVFrame *)data = p->frame;
1139 return frame_size[dec_mode];
1142 #define OFFSET(x) offsetof(G723_1_Context, x)
1143 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1145 static const AVOption options[] = {
1146 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1152 static const AVClass g723_1dec_class = {
1153 .class_name = "G.723.1 decoder",
1154 .item_name = av_default_item_name,
1156 .version = LIBAVUTIL_VERSION_INT,
1159 AVCodec ff_g723_1_decoder = {
1161 .type = AVMEDIA_TYPE_AUDIO,
1162 .id = AV_CODEC_ID_G723_1,
1163 .priv_data_size = sizeof(G723_1_Context),
1164 .init = g723_1_decode_init,
1165 .decode = g723_1_decode_frame,
1166 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1167 .capabilities = CODEC_CAP_SUBFRAMES,
1168 .priv_class = &g723_1dec_class,