3 * Copyright (c) 2010 Marcelo Galvao Povoa
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A particular PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * AMR wideband decoder
27 #include "libavutil/audioconvert.h"
28 #include "libavutil/common.h"
29 #include "libavutil/lfg.h"
34 #include "celp_filters.h"
35 #include "celp_math.h"
36 #include "acelp_filters.h"
37 #include "acelp_vectors.h"
38 #include "acelp_pitch_delay.h"
40 #define AMR_USE_16BIT_TABLES
43 #include "amrwbdata.h"
44 #include "mips/amrwbdec_mips.h"
47 AVFrame avframe; ///< AVFrame for decoded samples
48 AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream
49 enum Mode fr_cur_mode; ///< mode index of current frame
50 uint8_t fr_quality; ///< frame quality index (FQI)
51 float isf_cur[LP_ORDER]; ///< working ISF vector from current frame
52 float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame
53 float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame
54 double isp[4][LP_ORDER]; ///< ISP vectors from current frame
55 double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame
57 float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector
59 uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe
60 uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe
62 float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history
63 float *excitation; ///< points to current excitation in excitation_buf[]
65 float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe
66 float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe
68 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
69 float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes
70 float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes
72 float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe
74 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset"
75 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
76 float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold
78 float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz
79 float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling
80 float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz
82 float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters
83 float demph_mem[1]; ///< previous value in the de-emphasis filter
84 float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter
85 float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter
87 AVLFG prng; ///< random number generator for white noise excitation
88 uint8_t first_frame; ///< flag active during decoding of the first frame
89 ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs
90 ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs
91 CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs
92 CELPMContext celpm_ctx; ///< context for fixed point math operations
96 static av_cold int amrwb_decode_init(AVCodecContext *avctx)
98 AMRWBContext *ctx = avctx->priv_data;
101 if (avctx->channels > 1) {
102 av_log_missing_feature(avctx, "multi-channel AMR", 0);
103 return AVERROR_PATCHWELCOME;
107 avctx->channel_layout = AV_CH_LAYOUT_MONO;
108 if (!avctx->sample_rate)
109 avctx->sample_rate = 16000;
110 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
112 av_lfg_init(&ctx->prng, 1);
114 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1];
115 ctx->first_frame = 1;
117 for (i = 0; i < LP_ORDER; i++)
118 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
120 for (i = 0; i < 4; i++)
121 ctx->prediction_error[i] = MIN_ENERGY;
123 avcodec_get_frame_defaults(&ctx->avframe);
124 avctx->coded_frame = &ctx->avframe;
126 ff_acelp_filter_init(&ctx->acelpf_ctx);
127 ff_acelp_vectors_init(&ctx->acelpv_ctx);
128 ff_celp_filter_init(&ctx->celpf_ctx);
129 ff_celp_math_init(&ctx->celpm_ctx);
135 * Decode the frame header in the "MIME/storage" format. This format
136 * is simpler and does not carry the auxiliary frame information.
138 * @param[in] ctx The Context
139 * @param[in] buf Pointer to the input buffer
141 * @return The decoded header length in bytes
143 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
145 /* Decode frame header (1st octet) */
146 ctx->fr_cur_mode = buf[0] >> 3 & 0x0F;
147 ctx->fr_quality = (buf[0] & 0x4) == 0x4;
153 * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only).
155 * @param[in] ind Array of 5 indexes
156 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
159 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
163 for (i = 0; i < 9; i++)
164 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
166 for (i = 0; i < 7; i++)
167 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
169 for (i = 0; i < 5; i++)
170 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
172 for (i = 0; i < 4; i++)
173 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
175 for (i = 0; i < 7; i++)
176 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
180 * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode).
182 * @param[in] ind Array of 7 indexes
183 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
186 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
190 for (i = 0; i < 9; i++)
191 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
193 for (i = 0; i < 7; i++)
194 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
196 for (i = 0; i < 3; i++)
197 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
199 for (i = 0; i < 3; i++)
200 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
202 for (i = 0; i < 3; i++)
203 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
205 for (i = 0; i < 3; i++)
206 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
208 for (i = 0; i < 4; i++)
209 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
213 * Apply mean and past ISF values using the prediction factor.
214 * Updates past ISF vector.
216 * @param[in,out] isf_q Current quantized ISF
217 * @param[in,out] isf_past Past quantized ISF
220 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
225 for (i = 0; i < LP_ORDER; i++) {
227 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
228 isf_q[i] += PRED_FACTOR * isf_past[i];
234 * Interpolate the fourth ISP vector from current and past frames
235 * to obtain an ISP vector for each subframe.
237 * @param[in,out] isp_q ISPs for each subframe
238 * @param[in] isp4_past Past ISP for subframe 4
240 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
244 for (k = 0; k < 3; k++) {
245 float c = isfp_inter[k];
246 for (i = 0; i < LP_ORDER; i++)
247 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
252 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes).
253 * Calculate integer lag and fractional lag always using 1/4 resolution.
254 * In 1st and 3rd subframes the index is relative to last subframe integer lag.
256 * @param[out] lag_int Decoded integer pitch lag
257 * @param[out] lag_frac Decoded fractional pitch lag
258 * @param[in] pitch_index Adaptive codebook pitch index
259 * @param[in,out] base_lag_int Base integer lag used in relative subframes
260 * @param[in] subframe Current subframe index (0 to 3)
262 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
263 uint8_t *base_lag_int, int subframe)
265 if (subframe == 0 || subframe == 2) {
266 if (pitch_index < 376) {
267 *lag_int = (pitch_index + 137) >> 2;
268 *lag_frac = pitch_index - (*lag_int << 2) + 136;
269 } else if (pitch_index < 440) {
270 *lag_int = (pitch_index + 257 - 376) >> 1;
271 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;
272 /* the actual resolution is 1/2 but expressed as 1/4 */
274 *lag_int = pitch_index - 280;
277 /* minimum lag for next subframe */
278 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
279 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
280 // XXX: the spec states clearly that *base_lag_int should be
281 // the nearest integer to *lag_int (minus 8), but the ref code
282 // actually always uses its floor, I'm following the latter
284 *lag_int = (pitch_index + 1) >> 2;
285 *lag_frac = pitch_index - (*lag_int << 2);
286 *lag_int += *base_lag_int;
291 * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes.
292 * The description is analogous to decode_pitch_lag_high, but in 6k60 the
293 * relative index is used for all subframes except the first.
295 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
296 uint8_t *base_lag_int, int subframe, enum Mode mode)
298 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
299 if (pitch_index < 116) {
300 *lag_int = (pitch_index + 69) >> 1;
301 *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;
303 *lag_int = pitch_index - 24;
306 // XXX: same problem as before
307 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
308 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
310 *lag_int = (pitch_index + 1) >> 1;
311 *lag_frac = (pitch_index - (*lag_int << 1)) << 1;
312 *lag_int += *base_lag_int;
317 * Find the pitch vector by interpolating the past excitation at the
318 * pitch delay, which is obtained in this function.
320 * @param[in,out] ctx The context
321 * @param[in] amr_subframe Current subframe data
322 * @param[in] subframe Current subframe index (0 to 3)
324 static void decode_pitch_vector(AMRWBContext *ctx,
325 const AMRWBSubFrame *amr_subframe,
328 int pitch_lag_int, pitch_lag_frac;
330 float *exc = ctx->excitation;
331 enum Mode mode = ctx->fr_cur_mode;
333 if (mode <= MODE_8k85) {
334 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
335 &ctx->base_pitch_lag, subframe, mode);
337 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
338 &ctx->base_pitch_lag, subframe);
340 ctx->pitch_lag_int = pitch_lag_int;
341 pitch_lag_int += pitch_lag_frac > 0;
343 /* Calculate the pitch vector by interpolating the past excitation at the
344 pitch lag using a hamming windowed sinc function */
345 ctx->acelpf_ctx.acelp_interpolatef(exc,
346 exc + 1 - pitch_lag_int,
348 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
349 LP_ORDER, AMRWB_SFR_SIZE + 1);
351 /* Check which pitch signal path should be used
352 * 6k60 and 8k85 modes have the ltp flag set to 0 */
353 if (amr_subframe->ltp) {
354 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
356 for (i = 0; i < AMRWB_SFR_SIZE; i++)
357 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
359 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
363 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */
364 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))
366 /** Get the bit at specified position */
367 #define BIT_POS(x, p) (((x) >> (p)) & 1)
370 * The next six functions decode_[i]p_track decode exactly i pulses
371 * positions and amplitudes (-1 or 1) in a subframe track using
372 * an encoded pulse indexing (TS 26.190 section 5.8.2).
374 * The results are given in out[], in which a negative number means
375 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ).
377 * @param[out] out Output buffer (writes i elements)
378 * @param[in] code Pulse index (no. of bits varies, see below)
379 * @param[in] m (log2) Number of potential positions
380 * @param[in] off Offset for decoded positions
382 static inline void decode_1p_track(int *out, int code, int m, int off)
384 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits
386 out[0] = BIT_POS(code, m) ? -pos : pos;
389 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits
391 int pos0 = BIT_STR(code, m, m) + off;
392 int pos1 = BIT_STR(code, 0, m) + off;
394 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
395 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
396 out[1] = pos0 > pos1 ? -out[1] : out[1];
399 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits
401 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
403 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
404 m - 1, off + half_2p);
405 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
408 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits
410 int half_4p, subhalf_2p;
411 int b_offset = 1 << (m - 1);
413 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
414 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
415 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
416 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
418 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
419 m - 2, off + half_4p + subhalf_2p);
420 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
421 m - 1, off + half_4p);
423 case 1: /* 1 pulse in A, 3 pulses in B */
424 decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
426 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
427 m - 1, off + b_offset);
429 case 2: /* 2 pulses in each half */
430 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
432 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
433 m - 1, off + b_offset);
435 case 3: /* 3 pulses in A, 1 pulse in B */
436 decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
438 decode_1p_track(out + 3, BIT_STR(code, 0, m),
439 m - 1, off + b_offset);
444 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits
446 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
448 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
449 m - 1, off + half_3p);
451 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
454 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits
456 int b_offset = 1 << (m - 1);
457 /* which half has more pulses in cases 0 to 2 */
458 int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
459 int half_other = b_offset - half_more;
461 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
462 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
463 decode_1p_track(out, BIT_STR(code, 0, m),
464 m - 1, off + half_more);
465 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
466 m - 1, off + half_more);
468 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
469 decode_1p_track(out, BIT_STR(code, 0, m),
470 m - 1, off + half_other);
471 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
472 m - 1, off + half_more);
474 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
475 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
476 m - 1, off + half_other);
477 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
478 m - 1, off + half_more);
480 case 3: /* 3 pulses in A, 3 pulses in B */
481 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
483 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
484 m - 1, off + b_offset);
490 * Decode the algebraic codebook index to pulse positions and signs,
491 * then construct the algebraic codebook vector.
493 * @param[out] fixed_vector Buffer for the fixed codebook excitation
494 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)
495 * @param[in] pulse_lo LSBs part of the pulse index array
496 * @param[in] mode Mode of the current frame
498 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
499 const uint16_t *pulse_lo, const enum Mode mode)
501 /* sig_pos stores for each track the decoded pulse position indexes
502 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
504 int spacing = (mode == MODE_6k60) ? 2 : 4;
509 for (i = 0; i < 2; i++)
510 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
513 for (i = 0; i < 4; i++)
514 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
517 for (i = 0; i < 4; i++)
518 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
521 for (i = 0; i < 2; i++)
522 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
523 for (i = 2; i < 4; i++)
524 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
527 for (i = 0; i < 4; i++)
528 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
531 for (i = 0; i < 4; i++)
532 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
533 ((int) pulse_hi[i] << 14), 4, 1);
536 for (i = 0; i < 2; i++)
537 decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
538 ((int) pulse_hi[i] << 10), 4, 1);
539 for (i = 2; i < 4; i++)
540 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
541 ((int) pulse_hi[i] << 14), 4, 1);
545 for (i = 0; i < 4; i++)
546 decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
547 ((int) pulse_hi[i] << 11), 4, 1);
551 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
553 for (i = 0; i < 4; i++)
554 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
555 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
557 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
562 * Decode pitch gain and fixed gain correction factor.
564 * @param[in] vq_gain Vector-quantized index for gains
565 * @param[in] mode Mode of the current frame
566 * @param[out] fixed_gain_factor Decoded fixed gain correction factor
567 * @param[out] pitch_gain Decoded pitch gain
569 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
570 float *fixed_gain_factor, float *pitch_gain)
572 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
573 qua_gain_7b[vq_gain]);
575 *pitch_gain = gains[0] * (1.0f / (1 << 14));
576 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
580 * Apply pitch sharpening filters to the fixed codebook vector.
582 * @param[in] ctx The context
583 * @param[in,out] fixed_vector Fixed codebook excitation
585 // XXX: Spec states this procedure should be applied when the pitch
586 // lag is less than 64, but this checking seems absent in reference and AMR-NB
587 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
592 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
593 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
595 /* Periodicity enhancement part */
596 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
597 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
601 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced).
603 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors
604 * @param[in] p_gain, f_gain Pitch and fixed gains
605 * @param[in] ctx The context
607 // XXX: There is something wrong with the precision here! The magnitudes
608 // of the energies are not correct. Please check the reference code carefully
609 static float voice_factor(float *p_vector, float p_gain,
610 float *f_vector, float f_gain,
613 double p_ener = (double) ctx->dot_productf(p_vector, p_vector,
616 double f_ener = (double) ctx->dot_productf(f_vector, f_vector,
620 return (p_ener - f_ener) / (p_ener + f_ener);
624 * Reduce fixed vector sparseness by smoothing with one of three IR filters,
625 * also known as "adaptive phase dispersion".
627 * @param[in] ctx The context
628 * @param[in,out] fixed_vector Unfiltered fixed vector
629 * @param[out] buf Space for modified vector if necessary
631 * @return The potentially overwritten filtered fixed vector address
633 static float *anti_sparseness(AMRWBContext *ctx,
634 float *fixed_vector, float *buf)
638 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
641 if (ctx->pitch_gain[0] < 0.6) {
642 ir_filter_nr = 0; // strong filtering
643 } else if (ctx->pitch_gain[0] < 0.9) {
644 ir_filter_nr = 1; // medium filtering
646 ir_filter_nr = 2; // no filtering
649 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
650 if (ir_filter_nr < 2)
655 for (i = 0; i < 6; i++)
656 if (ctx->pitch_gain[i] < 0.6)
662 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
666 /* update ir filter strength history */
667 ctx->prev_ir_filter_nr = ir_filter_nr;
669 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
671 if (ir_filter_nr < 2) {
673 const float *coef = ir_filters_lookup[ir_filter_nr];
675 /* Circular convolution code in the reference
676 * decoder was modified to avoid using one
677 * extra array. The filtered vector is given by:
679 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
682 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
683 for (i = 0; i < AMRWB_SFR_SIZE; i++)
685 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
694 * Calculate a stability factor {teta} based on distance between
695 * current and past isf. A value of 1 shows maximum signal stability.
697 static float stability_factor(const float *isf, const float *isf_past)
702 for (i = 0; i < LP_ORDER - 1; i++)
703 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
705 // XXX: This part is not so clear from the reference code
706 // the result is more accurate changing the "/ 256" to "* 512"
707 return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
711 * Apply a non-linear fixed gain smoothing in order to reduce
712 * fluctuation in the energy of excitation.
714 * @param[in] fixed_gain Unsmoothed fixed gain
715 * @param[in,out] prev_tr_gain Previous threshold gain (updated)
716 * @param[in] voice_fac Frame voicing factor
717 * @param[in] stab_fac Frame stability factor
719 * @return The smoothed gain
721 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
722 float voice_fac, float stab_fac)
724 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
727 // XXX: the following fixed-point constants used to in(de)crement
728 // gain by 1.5dB were taken from the reference code, maybe it could
730 if (fixed_gain < *prev_tr_gain) {
731 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
732 (6226 * (1.0f / (1 << 15)))); // +1.5 dB
734 g0 = FFMAX(*prev_tr_gain, fixed_gain *
735 (27536 * (1.0f / (1 << 15)))); // -1.5 dB
737 *prev_tr_gain = g0; // update next frame threshold
739 return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
743 * Filter the fixed_vector to emphasize the higher frequencies.
745 * @param[in,out] fixed_vector Fixed codebook vector
746 * @param[in] voice_fac Frame voicing factor
748 static void pitch_enhancer(float *fixed_vector, float voice_fac)
751 float cpe = 0.125 * (1 + voice_fac);
752 float last = fixed_vector[0]; // holds c(i - 1)
754 fixed_vector[0] -= cpe * fixed_vector[1];
756 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
757 float cur = fixed_vector[i];
759 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
763 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
767 * Conduct 16th order linear predictive coding synthesis from excitation.
769 * @param[in] ctx Pointer to the AMRWBContext
770 * @param[in] lpc Pointer to the LPC coefficients
771 * @param[out] excitation Buffer for synthesis final excitation
772 * @param[in] fixed_gain Fixed codebook gain for synthesis
773 * @param[in] fixed_vector Algebraic codebook vector
774 * @param[in,out] samples Pointer to the output samples and memory
776 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
777 float fixed_gain, const float *fixed_vector,
780 ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
781 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
783 /* emphasize pitch vector contribution in low bitrate modes */
784 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
786 float energy = ctx->celpm_ctx.dot_productf(excitation, excitation,
789 // XXX: Weird part in both ref code and spec. A unknown parameter
790 // {beta} seems to be identical to the current pitch gain
791 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
793 for (i = 0; i < AMRWB_SFR_SIZE; i++)
794 excitation[i] += pitch_factor * ctx->pitch_vector[i];
796 ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
797 energy, AMRWB_SFR_SIZE);
800 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
801 AMRWB_SFR_SIZE, LP_ORDER);
805 * Apply to synthesis a de-emphasis filter of the form:
806 * H(z) = 1 / (1 - m * z^-1)
808 * @param[out] out Output buffer
809 * @param[in] in Input samples array with in[-1]
810 * @param[in] m Filter coefficient
811 * @param[in,out] mem State from last filtering
813 static void de_emphasis(float *out, float *in, float m, float mem[1])
817 out[0] = in[0] + m * mem[0];
819 for (i = 1; i < AMRWB_SFR_SIZE; i++)
820 out[i] = in[i] + out[i - 1] * m;
822 mem[0] = out[AMRWB_SFR_SIZE - 1];
826 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using
827 * a FIR interpolation filter. Uses past data from before *in address.
829 * @param[out] out Buffer for interpolated signal
830 * @param[in] in Current signal data (length 0.8*o_size)
831 * @param[in] o_size Output signal length
832 * @param[in] ctx The context
834 static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx)
836 const float *in0 = in - UPS_FIR_SIZE + 1;
838 int int_part = 0, frac_part;
841 for (j = 0; j < o_size / 5; j++) {
842 out[i] = in[int_part];
846 for (k = 1; k < 5; k++) {
847 out[i] = ctx->dot_productf(in0 + int_part,
848 upsample_fir[4 - frac_part],
858 * Calculate the high-band gain based on encoded index (23k85 mode) or
859 * on the low-band speech signal and the Voice Activity Detection flag.
861 * @param[in] ctx The context
862 * @param[in] synth LB speech synthesis at 12.8k
863 * @param[in] hb_idx Gain index for mode 23k85 only
864 * @param[in] vad VAD flag for the frame
866 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
867 uint16_t hb_idx, uint8_t vad)
872 if (ctx->fr_cur_mode == MODE_23k85)
873 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
875 tilt = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
876 ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE);
878 /* return gain bounded by [0.1, 1.0] */
879 return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
883 * Generate the high-band excitation with the same energy from the lower
884 * one and scaled by the given gain.
886 * @param[in] ctx The context
887 * @param[out] hb_exc Buffer for the excitation
888 * @param[in] synth_exc Low-band excitation used for synthesis
889 * @param[in] hb_gain Wanted excitation gain
891 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
892 const float *synth_exc, float hb_gain)
895 float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc, AMRWB_SFR_SIZE);
897 /* Generate a white-noise excitation */
898 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
899 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
901 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc,
902 energy * hb_gain * hb_gain,
907 * Calculate the auto-correlation for the ISF difference vector.
909 static float auto_correlation(float *diff_isf, float mean, int lag)
914 for (i = 7; i < LP_ORDER - 2; i++) {
915 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
922 * Extrapolate a ISF vector to the 16kHz range (20th order LP)
923 * used at mode 6k60 LP filter for the high frequency band.
925 * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER
928 static void extrapolate_isf(float isf[LP_ORDER_16k])
930 float diff_isf[LP_ORDER - 2], diff_mean;
933 int i, j, i_max_corr;
935 isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
937 /* Calculate the difference vector */
938 for (i = 0; i < LP_ORDER - 2; i++)
939 diff_isf[i] = isf[i + 1] - isf[i];
942 for (i = 2; i < LP_ORDER - 2; i++)
943 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
945 /* Find which is the maximum autocorrelation */
947 for (i = 0; i < 3; i++) {
948 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
950 if (corr_lag[i] > corr_lag[i_max_corr])
955 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
956 isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
957 - isf[i - 2 - i_max_corr];
959 /* Calculate an estimate for ISF(18) and scale ISF based on the error */
960 est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
961 scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
962 (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
964 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
965 diff_isf[j] = scale * (isf[i] - isf[i - 1]);
967 /* Stability insurance */
968 for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++)
969 if (diff_isf[i] + diff_isf[i - 1] < 5.0) {
970 if (diff_isf[i] > diff_isf[i - 1]) {
971 diff_isf[i - 1] = 5.0 - diff_isf[i];
973 diff_isf[i] = 5.0 - diff_isf[i - 1];
976 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
977 isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15));
979 /* Scale the ISF vector for 16000 Hz */
980 for (i = 0; i < LP_ORDER_16k - 1; i++)
985 * Spectral expand the LP coefficients using the equation:
986 * y[i] = x[i] * (gamma ** i)
988 * @param[out] out Output buffer (may use input array)
989 * @param[in] lpc LP coefficients array
990 * @param[in] gamma Weighting factor
991 * @param[in] size LP array size
993 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
998 for (i = 0; i < size; i++) {
999 out[i] = lpc[i] * fac;
1005 * Conduct 20th order linear predictive coding synthesis for the high
1006 * frequency band excitation at 16kHz.
1008 * @param[in] ctx The context
1009 * @param[in] subframe Current subframe index (0 to 3)
1010 * @param[in,out] samples Pointer to the output speech samples
1011 * @param[in] exc Generated white-noise scaled excitation
1012 * @param[in] isf Current frame isf vector
1013 * @param[in] isf_past Past frame final isf vector
1015 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
1016 const float *exc, const float *isf, const float *isf_past)
1018 float hb_lpc[LP_ORDER_16k];
1019 enum Mode mode = ctx->fr_cur_mode;
1021 if (mode == MODE_6k60) {
1022 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1023 double e_isp[LP_ORDER_16k];
1025 ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1026 1.0 - isfp_inter[subframe], LP_ORDER);
1028 extrapolate_isf(e_isf);
1030 e_isf[LP_ORDER_16k - 1] *= 2.0;
1031 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1032 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1034 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1036 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1039 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1040 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1044 * Apply a 15th order filter to high-band samples.
1045 * The filter characteristic depends on the given coefficients.
1047 * @param[out] out Buffer for filtered output
1048 * @param[in] fir_coef Filter coefficients
1049 * @param[in,out] mem State from last filtering (updated)
1050 * @param[in] in Input speech data (high-band)
1052 * @remark It is safe to pass the same array in in and out parameters
1055 #ifndef hb_fir_filter
1056 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1057 float mem[HB_FIR_SIZE], const float *in)
1060 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1062 memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1063 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1065 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1067 for (j = 0; j <= HB_FIR_SIZE; j++)
1068 out[i] += data[i + j] * fir_coef[j];
1071 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1073 #endif /* hb_fir_filter */
1076 * Update context state before the next subframe.
1078 static void update_sub_state(AMRWBContext *ctx)
1080 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1081 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1083 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1084 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1086 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1087 LP_ORDER * sizeof(float));
1088 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1089 UPS_MEM_SIZE * sizeof(float));
1090 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1091 LP_ORDER_16k * sizeof(float));
1094 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1095 int *got_frame_ptr, AVPacket *avpkt)
1097 AMRWBContext *ctx = avctx->priv_data;
1098 AMRWBFrame *cf = &ctx->frame;
1099 const uint8_t *buf = avpkt->data;
1100 int buf_size = avpkt->size;
1101 int expected_fr_size, header_size;
1103 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1104 float fixed_gain_factor; // fixed gain correction factor (gamma)
1105 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1106 float synth_fixed_gain; // the fixed gain that synthesis should use
1107 float voice_fac, stab_fac; // parameters used for gain smoothing
1108 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1109 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1110 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1114 /* get output buffer */
1115 ctx->avframe.nb_samples = 4 * AMRWB_SFR_SIZE_16k;
1116 if ((ret = avctx->get_buffer(avctx, &ctx->avframe)) < 0) {
1117 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1120 buf_out = (float *)ctx->avframe.data[0];
1122 header_size = decode_mime_header(ctx, buf);
1123 if (ctx->fr_cur_mode > MODE_SID) {
1124 av_log(avctx, AV_LOG_ERROR,
1125 "Invalid mode %d\n", ctx->fr_cur_mode);
1126 return AVERROR_INVALIDDATA;
1128 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1130 if (buf_size < expected_fr_size) {
1131 av_log(avctx, AV_LOG_ERROR,
1132 "Frame too small (%d bytes). Truncated file?\n", buf_size);
1134 return AVERROR_INVALIDDATA;
1137 if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1138 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1140 if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1141 av_log_missing_feature(avctx, "SID mode", 1);
1142 return AVERROR_PATCHWELCOME;
1145 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1146 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1148 /* Decode the quantized ISF vector */
1149 if (ctx->fr_cur_mode == MODE_6k60) {
1150 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur);
1152 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur);
1155 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past);
1156 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1);
1158 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1160 ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1161 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1163 /* Generate a ISP vector for each subframe */
1164 if (ctx->first_frame) {
1165 ctx->first_frame = 0;
1166 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1168 interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1170 for (sub = 0; sub < 4; sub++)
1171 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1173 for (sub = 0; sub < 4; sub++) {
1174 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1175 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1177 /* Decode adaptive codebook (pitch vector) */
1178 decode_pitch_vector(ctx, cur_subframe, sub);
1179 /* Decode innovative codebook (fixed vector) */
1180 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1181 cur_subframe->pul_il, ctx->fr_cur_mode);
1183 pitch_sharpening(ctx, ctx->fixed_vector);
1185 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1186 &fixed_gain_factor, &ctx->pitch_gain[0]);
1188 ctx->fixed_gain[0] =
1189 ff_amr_set_fixed_gain(fixed_gain_factor,
1190 ctx->celpm_ctx.dot_productf(ctx->fixed_vector,
1194 ctx->prediction_error,
1195 ENERGY_MEAN, energy_pred_fac);
1197 /* Calculate voice factor and store tilt for next subframe */
1198 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1199 ctx->fixed_vector, ctx->fixed_gain[0],
1201 ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1203 /* Construct current excitation */
1204 for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1205 ctx->excitation[i] *= ctx->pitch_gain[0];
1206 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1207 ctx->excitation[i] = truncf(ctx->excitation[i]);
1210 /* Post-processing of excitation elements */
1211 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1212 voice_fac, stab_fac);
1214 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1217 pitch_enhancer(synth_fixed_vector, voice_fac);
1219 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1220 synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1222 /* Synthesis speech post-processing */
1223 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],
1224 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1226 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
1227 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,
1228 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1230 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1231 AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx);
1233 /* High frequency band (6.4 - 7.0 kHz) generation part */
1234 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(hb_samples,
1235 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,
1236 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1238 hb_gain = find_hb_gain(ctx, hb_samples,
1239 cur_subframe->hb_gain, cf->vad);
1241 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1243 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1244 hb_exc, ctx->isf_cur, ctx->isf_past_final);
1246 /* High-band post-processing filters */
1247 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1248 &ctx->samples_hb[LP_ORDER_16k]);
1250 if (ctx->fr_cur_mode == MODE_23k85)
1251 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1254 /* Add the low and high frequency bands */
1255 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1256 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1258 /* Update buffers and history */
1259 update_sub_state(ctx);
1262 /* update state for next frame */
1263 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1264 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1267 *(AVFrame *)data = ctx->avframe;
1269 return expected_fr_size;
1272 AVCodec ff_amrwb_decoder = {
1274 .type = AVMEDIA_TYPE_AUDIO,
1275 .id = AV_CODEC_ID_AMR_WB,
1276 .priv_data_size = sizeof(AMRWBContext),
1277 .init = amrwb_decode_init,
1278 .decode = amrwb_decode_frame,
1279 .capabilities = CODEC_CAP_DR1,
1280 .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"),
1281 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1282 AV_SAMPLE_FMT_NONE },