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/channel_layout.h"
28 #include "libavutil/common.h"
29 #include "libavutil/float_dsp.h"
30 #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"
41 #define AMR_USE_16BIT_TABLES
44 #include "amrwbdata.h"
45 #include "mips/amrwbdec_mips.h"
48 AVFrame avframe; ///< AVFrame for decoded samples
49 AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream
50 enum Mode fr_cur_mode; ///< mode index of current frame
51 uint8_t fr_quality; ///< frame quality index (FQI)
52 float isf_cur[LP_ORDER]; ///< working ISF vector from current frame
53 float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame
54 float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame
55 double isp[4][LP_ORDER]; ///< ISP vectors from current frame
56 double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame
58 float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector
60 uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe
61 uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe
63 float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history
64 float *excitation; ///< points to current excitation in excitation_buf[]
66 float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe
67 float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe
69 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
70 float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes
71 float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes
73 float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe
75 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset"
76 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
77 float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold
79 float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz
80 float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling
81 float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz
83 float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters
84 float demph_mem[1]; ///< previous value in the de-emphasis filter
85 float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter
86 float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter
88 AVLFG prng; ///< random number generator for white noise excitation
89 uint8_t first_frame; ///< flag active during decoding of the first frame
90 ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs
91 ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs
92 CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs
93 CELPMContext celpm_ctx; ///< context for fixed point math operations
97 static av_cold int amrwb_decode_init(AVCodecContext *avctx)
99 AMRWBContext *ctx = avctx->priv_data;
102 if (avctx->channels > 1) {
103 av_log_missing_feature(avctx, "multi-channel AMR", 0);
104 return AVERROR_PATCHWELCOME;
108 avctx->channel_layout = AV_CH_LAYOUT_MONO;
109 if (!avctx->sample_rate)
110 avctx->sample_rate = 16000;
111 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
113 av_lfg_init(&ctx->prng, 1);
115 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1];
116 ctx->first_frame = 1;
118 for (i = 0; i < LP_ORDER; i++)
119 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
121 for (i = 0; i < 4; i++)
122 ctx->prediction_error[i] = MIN_ENERGY;
124 avcodec_get_frame_defaults(&ctx->avframe);
125 avctx->coded_frame = &ctx->avframe;
127 ff_acelp_filter_init(&ctx->acelpf_ctx);
128 ff_acelp_vectors_init(&ctx->acelpv_ctx);
129 ff_celp_filter_init(&ctx->celpf_ctx);
130 ff_celp_math_init(&ctx->celpm_ctx);
136 * Decode the frame header in the "MIME/storage" format. This format
137 * is simpler and does not carry the auxiliary frame information.
139 * @param[in] ctx The Context
140 * @param[in] buf Pointer to the input buffer
142 * @return The decoded header length in bytes
144 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
146 /* Decode frame header (1st octet) */
147 ctx->fr_cur_mode = buf[0] >> 3 & 0x0F;
148 ctx->fr_quality = (buf[0] & 0x4) == 0x4;
154 * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only).
156 * @param[in] ind Array of 5 indexes
157 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
160 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
164 for (i = 0; i < 9; i++)
165 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
167 for (i = 0; i < 7; i++)
168 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
170 for (i = 0; i < 5; i++)
171 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
173 for (i = 0; i < 4; i++)
174 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
176 for (i = 0; i < 7; i++)
177 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
181 * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode).
183 * @param[in] ind Array of 7 indexes
184 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
187 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
191 for (i = 0; i < 9; i++)
192 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
194 for (i = 0; i < 7; i++)
195 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
197 for (i = 0; i < 3; i++)
198 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
200 for (i = 0; i < 3; i++)
201 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
203 for (i = 0; i < 3; i++)
204 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
206 for (i = 0; i < 3; i++)
207 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
209 for (i = 0; i < 4; i++)
210 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
214 * Apply mean and past ISF values using the prediction factor.
215 * Updates past ISF vector.
217 * @param[in,out] isf_q Current quantized ISF
218 * @param[in,out] isf_past Past quantized ISF
221 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
226 for (i = 0; i < LP_ORDER; i++) {
228 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
229 isf_q[i] += PRED_FACTOR * isf_past[i];
235 * Interpolate the fourth ISP vector from current and past frames
236 * to obtain an ISP vector for each subframe.
238 * @param[in,out] isp_q ISPs for each subframe
239 * @param[in] isp4_past Past ISP for subframe 4
241 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
245 for (k = 0; k < 3; k++) {
246 float c = isfp_inter[k];
247 for (i = 0; i < LP_ORDER; i++)
248 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
253 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes).
254 * Calculate integer lag and fractional lag always using 1/4 resolution.
255 * In 1st and 3rd subframes the index is relative to last subframe integer lag.
257 * @param[out] lag_int Decoded integer pitch lag
258 * @param[out] lag_frac Decoded fractional pitch lag
259 * @param[in] pitch_index Adaptive codebook pitch index
260 * @param[in,out] base_lag_int Base integer lag used in relative subframes
261 * @param[in] subframe Current subframe index (0 to 3)
263 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
264 uint8_t *base_lag_int, int subframe)
266 if (subframe == 0 || subframe == 2) {
267 if (pitch_index < 376) {
268 *lag_int = (pitch_index + 137) >> 2;
269 *lag_frac = pitch_index - (*lag_int << 2) + 136;
270 } else if (pitch_index < 440) {
271 *lag_int = (pitch_index + 257 - 376) >> 1;
272 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;
273 /* the actual resolution is 1/2 but expressed as 1/4 */
275 *lag_int = pitch_index - 280;
278 /* minimum lag for next subframe */
279 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
280 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
281 // XXX: the spec states clearly that *base_lag_int should be
282 // the nearest integer to *lag_int (minus 8), but the ref code
283 // actually always uses its floor, I'm following the latter
285 *lag_int = (pitch_index + 1) >> 2;
286 *lag_frac = pitch_index - (*lag_int << 2);
287 *lag_int += *base_lag_int;
292 * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes.
293 * The description is analogous to decode_pitch_lag_high, but in 6k60 the
294 * relative index is used for all subframes except the first.
296 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
297 uint8_t *base_lag_int, int subframe, enum Mode mode)
299 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
300 if (pitch_index < 116) {
301 *lag_int = (pitch_index + 69) >> 1;
302 *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;
304 *lag_int = pitch_index - 24;
307 // XXX: same problem as before
308 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
309 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
311 *lag_int = (pitch_index + 1) >> 1;
312 *lag_frac = (pitch_index - (*lag_int << 1)) << 1;
313 *lag_int += *base_lag_int;
318 * Find the pitch vector by interpolating the past excitation at the
319 * pitch delay, which is obtained in this function.
321 * @param[in,out] ctx The context
322 * @param[in] amr_subframe Current subframe data
323 * @param[in] subframe Current subframe index (0 to 3)
325 static void decode_pitch_vector(AMRWBContext *ctx,
326 const AMRWBSubFrame *amr_subframe,
329 int pitch_lag_int, pitch_lag_frac;
331 float *exc = ctx->excitation;
332 enum Mode mode = ctx->fr_cur_mode;
334 if (mode <= MODE_8k85) {
335 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
336 &ctx->base_pitch_lag, subframe, mode);
338 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
339 &ctx->base_pitch_lag, subframe);
341 ctx->pitch_lag_int = pitch_lag_int;
342 pitch_lag_int += pitch_lag_frac > 0;
344 /* Calculate the pitch vector by interpolating the past excitation at the
345 pitch lag using a hamming windowed sinc function */
346 ctx->acelpf_ctx.acelp_interpolatef(exc,
347 exc + 1 - pitch_lag_int,
349 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
350 LP_ORDER, AMRWB_SFR_SIZE + 1);
352 /* Check which pitch signal path should be used
353 * 6k60 and 8k85 modes have the ltp flag set to 0 */
354 if (amr_subframe->ltp) {
355 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
357 for (i = 0; i < AMRWB_SFR_SIZE; i++)
358 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
360 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
364 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */
365 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))
367 /** Get the bit at specified position */
368 #define BIT_POS(x, p) (((x) >> (p)) & 1)
371 * The next six functions decode_[i]p_track decode exactly i pulses
372 * positions and amplitudes (-1 or 1) in a subframe track using
373 * an encoded pulse indexing (TS 26.190 section 5.8.2).
375 * The results are given in out[], in which a negative number means
376 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ).
378 * @param[out] out Output buffer (writes i elements)
379 * @param[in] code Pulse index (no. of bits varies, see below)
380 * @param[in] m (log2) Number of potential positions
381 * @param[in] off Offset for decoded positions
383 static inline void decode_1p_track(int *out, int code, int m, int off)
385 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits
387 out[0] = BIT_POS(code, m) ? -pos : pos;
390 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits
392 int pos0 = BIT_STR(code, m, m) + off;
393 int pos1 = BIT_STR(code, 0, m) + off;
395 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
396 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
397 out[1] = pos0 > pos1 ? -out[1] : out[1];
400 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits
402 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
404 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
405 m - 1, off + half_2p);
406 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
409 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits
411 int half_4p, subhalf_2p;
412 int b_offset = 1 << (m - 1);
414 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
415 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
416 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
417 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
419 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
420 m - 2, off + half_4p + subhalf_2p);
421 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
422 m - 1, off + half_4p);
424 case 1: /* 1 pulse in A, 3 pulses in B */
425 decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
427 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
428 m - 1, off + b_offset);
430 case 2: /* 2 pulses in each half */
431 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
433 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
434 m - 1, off + b_offset);
436 case 3: /* 3 pulses in A, 1 pulse in B */
437 decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
439 decode_1p_track(out + 3, BIT_STR(code, 0, m),
440 m - 1, off + b_offset);
445 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits
447 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
449 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
450 m - 1, off + half_3p);
452 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
455 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits
457 int b_offset = 1 << (m - 1);
458 /* which half has more pulses in cases 0 to 2 */
459 int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
460 int half_other = b_offset - half_more;
462 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
463 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
464 decode_1p_track(out, BIT_STR(code, 0, m),
465 m - 1, off + half_more);
466 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
467 m - 1, off + half_more);
469 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
470 decode_1p_track(out, BIT_STR(code, 0, m),
471 m - 1, off + half_other);
472 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
473 m - 1, off + half_more);
475 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
476 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
477 m - 1, off + half_other);
478 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
479 m - 1, off + half_more);
481 case 3: /* 3 pulses in A, 3 pulses in B */
482 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
484 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
485 m - 1, off + b_offset);
491 * Decode the algebraic codebook index to pulse positions and signs,
492 * then construct the algebraic codebook vector.
494 * @param[out] fixed_vector Buffer for the fixed codebook excitation
495 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)
496 * @param[in] pulse_lo LSBs part of the pulse index array
497 * @param[in] mode Mode of the current frame
499 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
500 const uint16_t *pulse_lo, const enum Mode mode)
502 /* sig_pos stores for each track the decoded pulse position indexes
503 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
505 int spacing = (mode == MODE_6k60) ? 2 : 4;
510 for (i = 0; i < 2; i++)
511 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
514 for (i = 0; i < 4; i++)
515 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
518 for (i = 0; i < 4; i++)
519 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
522 for (i = 0; i < 2; i++)
523 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
524 for (i = 2; i < 4; i++)
525 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
528 for (i = 0; i < 4; i++)
529 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
532 for (i = 0; i < 4; i++)
533 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
534 ((int) pulse_hi[i] << 14), 4, 1);
537 for (i = 0; i < 2; i++)
538 decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
539 ((int) pulse_hi[i] << 10), 4, 1);
540 for (i = 2; i < 4; i++)
541 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
542 ((int) pulse_hi[i] << 14), 4, 1);
546 for (i = 0; i < 4; i++)
547 decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
548 ((int) pulse_hi[i] << 11), 4, 1);
552 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
554 for (i = 0; i < 4; i++)
555 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
556 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
558 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
563 * Decode pitch gain and fixed gain correction factor.
565 * @param[in] vq_gain Vector-quantized index for gains
566 * @param[in] mode Mode of the current frame
567 * @param[out] fixed_gain_factor Decoded fixed gain correction factor
568 * @param[out] pitch_gain Decoded pitch gain
570 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
571 float *fixed_gain_factor, float *pitch_gain)
573 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
574 qua_gain_7b[vq_gain]);
576 *pitch_gain = gains[0] * (1.0f / (1 << 14));
577 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
581 * Apply pitch sharpening filters to the fixed codebook vector.
583 * @param[in] ctx The context
584 * @param[in,out] fixed_vector Fixed codebook excitation
586 // XXX: Spec states this procedure should be applied when the pitch
587 // lag is less than 64, but this checking seems absent in reference and AMR-NB
588 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
593 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
594 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
596 /* Periodicity enhancement part */
597 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
598 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
602 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced).
604 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors
605 * @param[in] p_gain, f_gain Pitch and fixed gains
606 * @param[in] ctx The context
608 // XXX: There is something wrong with the precision here! The magnitudes
609 // of the energies are not correct. Please check the reference code carefully
610 static float voice_factor(float *p_vector, float p_gain,
611 float *f_vector, float f_gain,
614 double p_ener = (double) ctx->dot_productf(p_vector, p_vector,
617 double f_ener = (double) ctx->dot_productf(f_vector, f_vector,
621 return (p_ener - f_ener) / (p_ener + f_ener);
625 * Reduce fixed vector sparseness by smoothing with one of three IR filters,
626 * also known as "adaptive phase dispersion".
628 * @param[in] ctx The context
629 * @param[in,out] fixed_vector Unfiltered fixed vector
630 * @param[out] buf Space for modified vector if necessary
632 * @return The potentially overwritten filtered fixed vector address
634 static float *anti_sparseness(AMRWBContext *ctx,
635 float *fixed_vector, float *buf)
639 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
642 if (ctx->pitch_gain[0] < 0.6) {
643 ir_filter_nr = 0; // strong filtering
644 } else if (ctx->pitch_gain[0] < 0.9) {
645 ir_filter_nr = 1; // medium filtering
647 ir_filter_nr = 2; // no filtering
650 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
651 if (ir_filter_nr < 2)
656 for (i = 0; i < 6; i++)
657 if (ctx->pitch_gain[i] < 0.6)
663 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
667 /* update ir filter strength history */
668 ctx->prev_ir_filter_nr = ir_filter_nr;
670 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
672 if (ir_filter_nr < 2) {
674 const float *coef = ir_filters_lookup[ir_filter_nr];
676 /* Circular convolution code in the reference
677 * decoder was modified to avoid using one
678 * extra array. The filtered vector is given by:
680 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
683 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
684 for (i = 0; i < AMRWB_SFR_SIZE; i++)
686 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
695 * Calculate a stability factor {teta} based on distance between
696 * current and past isf. A value of 1 shows maximum signal stability.
698 static float stability_factor(const float *isf, const float *isf_past)
703 for (i = 0; i < LP_ORDER - 1; i++)
704 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
706 // XXX: This part is not so clear from the reference code
707 // the result is more accurate changing the "/ 256" to "* 512"
708 return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
712 * Apply a non-linear fixed gain smoothing in order to reduce
713 * fluctuation in the energy of excitation.
715 * @param[in] fixed_gain Unsmoothed fixed gain
716 * @param[in,out] prev_tr_gain Previous threshold gain (updated)
717 * @param[in] voice_fac Frame voicing factor
718 * @param[in] stab_fac Frame stability factor
720 * @return The smoothed gain
722 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
723 float voice_fac, float stab_fac)
725 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
728 // XXX: the following fixed-point constants used to in(de)crement
729 // gain by 1.5dB were taken from the reference code, maybe it could
731 if (fixed_gain < *prev_tr_gain) {
732 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
733 (6226 * (1.0f / (1 << 15)))); // +1.5 dB
735 g0 = FFMAX(*prev_tr_gain, fixed_gain *
736 (27536 * (1.0f / (1 << 15)))); // -1.5 dB
738 *prev_tr_gain = g0; // update next frame threshold
740 return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
744 * Filter the fixed_vector to emphasize the higher frequencies.
746 * @param[in,out] fixed_vector Fixed codebook vector
747 * @param[in] voice_fac Frame voicing factor
749 static void pitch_enhancer(float *fixed_vector, float voice_fac)
752 float cpe = 0.125 * (1 + voice_fac);
753 float last = fixed_vector[0]; // holds c(i - 1)
755 fixed_vector[0] -= cpe * fixed_vector[1];
757 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
758 float cur = fixed_vector[i];
760 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
764 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
768 * Conduct 16th order linear predictive coding synthesis from excitation.
770 * @param[in] ctx Pointer to the AMRWBContext
771 * @param[in] lpc Pointer to the LPC coefficients
772 * @param[out] excitation Buffer for synthesis final excitation
773 * @param[in] fixed_gain Fixed codebook gain for synthesis
774 * @param[in] fixed_vector Algebraic codebook vector
775 * @param[in,out] samples Pointer to the output samples and memory
777 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
778 float fixed_gain, const float *fixed_vector,
781 ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
782 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
784 /* emphasize pitch vector contribution in low bitrate modes */
785 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
787 float energy = ctx->celpm_ctx.dot_productf(excitation, excitation,
790 // XXX: Weird part in both ref code and spec. A unknown parameter
791 // {beta} seems to be identical to the current pitch gain
792 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
794 for (i = 0; i < AMRWB_SFR_SIZE; i++)
795 excitation[i] += pitch_factor * ctx->pitch_vector[i];
797 ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
798 energy, AMRWB_SFR_SIZE);
801 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
802 AMRWB_SFR_SIZE, LP_ORDER);
806 * Apply to synthesis a de-emphasis filter of the form:
807 * H(z) = 1 / (1 - m * z^-1)
809 * @param[out] out Output buffer
810 * @param[in] in Input samples array with in[-1]
811 * @param[in] m Filter coefficient
812 * @param[in,out] mem State from last filtering
814 static void de_emphasis(float *out, float *in, float m, float mem[1])
818 out[0] = in[0] + m * mem[0];
820 for (i = 1; i < AMRWB_SFR_SIZE; i++)
821 out[i] = in[i] + out[i - 1] * m;
823 mem[0] = out[AMRWB_SFR_SIZE - 1];
827 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using
828 * a FIR interpolation filter. Uses past data from before *in address.
830 * @param[out] out Buffer for interpolated signal
831 * @param[in] in Current signal data (length 0.8*o_size)
832 * @param[in] o_size Output signal length
833 * @param[in] ctx The context
835 static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx)
837 const float *in0 = in - UPS_FIR_SIZE + 1;
839 int int_part = 0, frac_part;
842 for (j = 0; j < o_size / 5; j++) {
843 out[i] = in[int_part];
847 for (k = 1; k < 5; k++) {
848 out[i] = ctx->dot_productf(in0 + int_part,
849 upsample_fir[4 - frac_part],
859 * Calculate the high-band gain based on encoded index (23k85 mode) or
860 * on the low-band speech signal and the Voice Activity Detection flag.
862 * @param[in] ctx The context
863 * @param[in] synth LB speech synthesis at 12.8k
864 * @param[in] hb_idx Gain index for mode 23k85 only
865 * @param[in] vad VAD flag for the frame
867 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
868 uint16_t hb_idx, uint8_t vad)
873 if (ctx->fr_cur_mode == MODE_23k85)
874 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
876 tilt = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
877 ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE);
879 /* return gain bounded by [0.1, 1.0] */
880 return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
884 * Generate the high-band excitation with the same energy from the lower
885 * one and scaled by the given gain.
887 * @param[in] ctx The context
888 * @param[out] hb_exc Buffer for the excitation
889 * @param[in] synth_exc Low-band excitation used for synthesis
890 * @param[in] hb_gain Wanted excitation gain
892 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
893 const float *synth_exc, float hb_gain)
896 float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc,
899 /* Generate a white-noise excitation */
900 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
901 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
903 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc,
904 energy * hb_gain * hb_gain,
909 * Calculate the auto-correlation for the ISF difference vector.
911 static float auto_correlation(float *diff_isf, float mean, int lag)
916 for (i = 7; i < LP_ORDER - 2; i++) {
917 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
924 * Extrapolate a ISF vector to the 16kHz range (20th order LP)
925 * used at mode 6k60 LP filter for the high frequency band.
927 * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER
930 static void extrapolate_isf(float isf[LP_ORDER_16k])
932 float diff_isf[LP_ORDER - 2], diff_mean;
935 int i, j, i_max_corr;
937 isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
939 /* Calculate the difference vector */
940 for (i = 0; i < LP_ORDER - 2; i++)
941 diff_isf[i] = isf[i + 1] - isf[i];
944 for (i = 2; i < LP_ORDER - 2; i++)
945 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
947 /* Find which is the maximum autocorrelation */
949 for (i = 0; i < 3; i++) {
950 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
952 if (corr_lag[i] > corr_lag[i_max_corr])
957 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
958 isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
959 - isf[i - 2 - i_max_corr];
961 /* Calculate an estimate for ISF(18) and scale ISF based on the error */
962 est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
963 scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
964 (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
966 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
967 diff_isf[j] = scale * (isf[i] - isf[i - 1]);
969 /* Stability insurance */
970 for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++)
971 if (diff_isf[i] + diff_isf[i - 1] < 5.0) {
972 if (diff_isf[i] > diff_isf[i - 1]) {
973 diff_isf[i - 1] = 5.0 - diff_isf[i];
975 diff_isf[i] = 5.0 - diff_isf[i - 1];
978 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)
979 isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15));
981 /* Scale the ISF vector for 16000 Hz */
982 for (i = 0; i < LP_ORDER_16k - 1; i++)
987 * Spectral expand the LP coefficients using the equation:
988 * y[i] = x[i] * (gamma ** i)
990 * @param[out] out Output buffer (may use input array)
991 * @param[in] lpc LP coefficients array
992 * @param[in] gamma Weighting factor
993 * @param[in] size LP array size
995 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
1000 for (i = 0; i < size; i++) {
1001 out[i] = lpc[i] * fac;
1007 * Conduct 20th order linear predictive coding synthesis for the high
1008 * frequency band excitation at 16kHz.
1010 * @param[in] ctx The context
1011 * @param[in] subframe Current subframe index (0 to 3)
1012 * @param[in,out] samples Pointer to the output speech samples
1013 * @param[in] exc Generated white-noise scaled excitation
1014 * @param[in] isf Current frame isf vector
1015 * @param[in] isf_past Past frame final isf vector
1017 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
1018 const float *exc, const float *isf, const float *isf_past)
1020 float hb_lpc[LP_ORDER_16k];
1021 enum Mode mode = ctx->fr_cur_mode;
1023 if (mode == MODE_6k60) {
1024 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1025 double e_isp[LP_ORDER_16k];
1027 ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1028 1.0 - isfp_inter[subframe], LP_ORDER);
1030 extrapolate_isf(e_isf);
1032 e_isf[LP_ORDER_16k - 1] *= 2.0;
1033 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1034 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1036 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1038 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1041 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1042 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1046 * Apply a 15th order filter to high-band samples.
1047 * The filter characteristic depends on the given coefficients.
1049 * @param[out] out Buffer for filtered output
1050 * @param[in] fir_coef Filter coefficients
1051 * @param[in,out] mem State from last filtering (updated)
1052 * @param[in] in Input speech data (high-band)
1054 * @remark It is safe to pass the same array in in and out parameters
1057 #ifndef hb_fir_filter
1058 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1059 float mem[HB_FIR_SIZE], const float *in)
1062 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1064 memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1065 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1067 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1069 for (j = 0; j <= HB_FIR_SIZE; j++)
1070 out[i] += data[i + j] * fir_coef[j];
1073 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1075 #endif /* hb_fir_filter */
1078 * Update context state before the next subframe.
1080 static void update_sub_state(AMRWBContext *ctx)
1082 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1083 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1085 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1086 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1088 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1089 LP_ORDER * sizeof(float));
1090 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1091 UPS_MEM_SIZE * sizeof(float));
1092 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1093 LP_ORDER_16k * sizeof(float));
1096 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1097 int *got_frame_ptr, AVPacket *avpkt)
1099 AMRWBContext *ctx = avctx->priv_data;
1100 AMRWBFrame *cf = &ctx->frame;
1101 const uint8_t *buf = avpkt->data;
1102 int buf_size = avpkt->size;
1103 int expected_fr_size, header_size;
1105 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1106 float fixed_gain_factor; // fixed gain correction factor (gamma)
1107 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1108 float synth_fixed_gain; // the fixed gain that synthesis should use
1109 float voice_fac, stab_fac; // parameters used for gain smoothing
1110 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1111 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1112 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1116 /* get output buffer */
1117 ctx->avframe.nb_samples = 4 * AMRWB_SFR_SIZE_16k;
1118 if ((ret = ff_get_buffer(avctx, &ctx->avframe)) < 0) {
1119 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1122 buf_out = (float *)ctx->avframe.data[0];
1124 header_size = decode_mime_header(ctx, buf);
1125 if (ctx->fr_cur_mode > MODE_SID) {
1126 av_log(avctx, AV_LOG_ERROR,
1127 "Invalid mode %d\n", ctx->fr_cur_mode);
1128 return AVERROR_INVALIDDATA;
1130 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1132 if (buf_size < expected_fr_size) {
1133 av_log(avctx, AV_LOG_ERROR,
1134 "Frame too small (%d bytes). Truncated file?\n", buf_size);
1136 return AVERROR_INVALIDDATA;
1139 if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1140 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1142 if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1143 av_log_missing_feature(avctx, "SID mode", 1);
1144 return AVERROR_PATCHWELCOME;
1147 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1148 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1150 /* Decode the quantized ISF vector */
1151 if (ctx->fr_cur_mode == MODE_6k60) {
1152 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur);
1154 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur);
1157 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past);
1158 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1);
1160 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1162 ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1163 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1165 /* Generate a ISP vector for each subframe */
1166 if (ctx->first_frame) {
1167 ctx->first_frame = 0;
1168 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1170 interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1172 for (sub = 0; sub < 4; sub++)
1173 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1175 for (sub = 0; sub < 4; sub++) {
1176 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1177 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1179 /* Decode adaptive codebook (pitch vector) */
1180 decode_pitch_vector(ctx, cur_subframe, sub);
1181 /* Decode innovative codebook (fixed vector) */
1182 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1183 cur_subframe->pul_il, ctx->fr_cur_mode);
1185 pitch_sharpening(ctx, ctx->fixed_vector);
1187 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1188 &fixed_gain_factor, &ctx->pitch_gain[0]);
1190 ctx->fixed_gain[0] =
1191 ff_amr_set_fixed_gain(fixed_gain_factor,
1192 ctx->celpm_ctx.dot_productf(ctx->fixed_vector,
1196 ctx->prediction_error,
1197 ENERGY_MEAN, energy_pred_fac);
1199 /* Calculate voice factor and store tilt for next subframe */
1200 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1201 ctx->fixed_vector, ctx->fixed_gain[0],
1203 ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1205 /* Construct current excitation */
1206 for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1207 ctx->excitation[i] *= ctx->pitch_gain[0];
1208 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1209 ctx->excitation[i] = truncf(ctx->excitation[i]);
1212 /* Post-processing of excitation elements */
1213 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1214 voice_fac, stab_fac);
1216 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1219 pitch_enhancer(synth_fixed_vector, voice_fac);
1221 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1222 synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1224 /* Synthesis speech post-processing */
1225 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],
1226 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1228 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
1229 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,
1230 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1232 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1233 AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx);
1235 /* High frequency band (6.4 - 7.0 kHz) generation part */
1236 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(hb_samples,
1237 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,
1238 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1240 hb_gain = find_hb_gain(ctx, hb_samples,
1241 cur_subframe->hb_gain, cf->vad);
1243 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1245 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1246 hb_exc, ctx->isf_cur, ctx->isf_past_final);
1248 /* High-band post-processing filters */
1249 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1250 &ctx->samples_hb[LP_ORDER_16k]);
1252 if (ctx->fr_cur_mode == MODE_23k85)
1253 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1256 /* Add the low and high frequency bands */
1257 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1258 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1260 /* Update buffers and history */
1261 update_sub_state(ctx);
1264 /* update state for next frame */
1265 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1266 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1269 *(AVFrame *)data = ctx->avframe;
1271 return expected_fr_size;
1274 AVCodec ff_amrwb_decoder = {
1276 .type = AVMEDIA_TYPE_AUDIO,
1277 .id = AV_CODEC_ID_AMR_WB,
1278 .priv_data_size = sizeof(AMRWBContext),
1279 .init = amrwb_decode_init,
1280 .decode = amrwb_decode_frame,
1281 .capabilities = CODEC_CAP_DR1,
1282 .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"),
1283 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1284 AV_SAMPLE_FMT_NONE },