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/common.h"
28 #include "libavutil/lfg.h"
32 #include "celp_math.h"
33 #include "celp_filters.h"
34 #include "acelp_filters.h"
35 #include "acelp_vectors.h"
36 #include "acelp_pitch_delay.h"
38 #define AMR_USE_16BIT_TABLES
41 #include "amrwbdata.h"
42 #include "mips/amrwbdec_mips.h"
45 AVFrame avframe; ///< AVFrame for decoded samples
46 AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream
47 enum Mode fr_cur_mode; ///< mode index of current frame
48 uint8_t fr_quality; ///< frame quality index (FQI)
49 float isf_cur[LP_ORDER]; ///< working ISF vector from current frame
50 float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame
51 float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame
52 double isp[4][LP_ORDER]; ///< ISP vectors from current frame
53 double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame
55 float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector
57 uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe
58 uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe
60 float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history
61 float *excitation; ///< points to current excitation in excitation_buf[]
63 float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe
64 float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe
66 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
67 float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes
68 float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes
70 float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe
72 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset"
73 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
74 float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold
76 float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz
77 float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling
78 float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz
80 float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters
81 float demph_mem[1]; ///< previous value in the de-emphasis filter
82 float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter
83 float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter
85 AVLFG prng; ///< random number generator for white noise excitation
86 uint8_t first_frame; ///< flag active during decoding of the first frame
87 ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs
88 ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs
89 CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs
90 CELPMContext celpm_ctx; ///< context for fixed point math operations
94 static av_cold int amrwb_decode_init(AVCodecContext *avctx)
96 AMRWBContext *ctx = avctx->priv_data;
99 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
101 av_lfg_init(&ctx->prng, 1);
103 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1];
104 ctx->first_frame = 1;
106 for (i = 0; i < LP_ORDER; i++)
107 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
109 for (i = 0; i < 4; i++)
110 ctx->prediction_error[i] = MIN_ENERGY;
112 avcodec_get_frame_defaults(&ctx->avframe);
113 avctx->coded_frame = &ctx->avframe;
115 ff_acelp_filter_init(&ctx->acelpf_ctx);
116 ff_acelp_vectors_init(&ctx->acelpv_ctx);
117 ff_celp_filter_init(&ctx->celpf_ctx);
118 ff_celp_math_init(&ctx->celpm_ctx);
124 * Decode the frame header in the "MIME/storage" format. This format
125 * is simpler and does not carry the auxiliary frame information.
127 * @param[in] ctx The Context
128 * @param[in] buf Pointer to the input buffer
130 * @return The decoded header length in bytes
132 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
134 /* Decode frame header (1st octet) */
135 ctx->fr_cur_mode = buf[0] >> 3 & 0x0F;
136 ctx->fr_quality = (buf[0] & 0x4) == 0x4;
142 * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only).
144 * @param[in] ind Array of 5 indexes
145 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
148 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
152 for (i = 0; i < 9; i++)
153 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
155 for (i = 0; i < 7; i++)
156 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
158 for (i = 0; i < 5; i++)
159 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
161 for (i = 0; i < 4; i++)
162 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
164 for (i = 0; i < 7; i++)
165 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
169 * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode).
171 * @param[in] ind Array of 7 indexes
172 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
175 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
179 for (i = 0; i < 9; i++)
180 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
182 for (i = 0; i < 7; i++)
183 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
185 for (i = 0; i < 3; i++)
186 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
188 for (i = 0; i < 3; i++)
189 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
191 for (i = 0; i < 3; i++)
192 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
194 for (i = 0; i < 3; i++)
195 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
197 for (i = 0; i < 4; i++)
198 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
202 * Apply mean and past ISF values using the prediction factor.
203 * Updates past ISF vector.
205 * @param[in,out] isf_q Current quantized ISF
206 * @param[in,out] isf_past Past quantized ISF
209 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
214 for (i = 0; i < LP_ORDER; i++) {
216 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
217 isf_q[i] += PRED_FACTOR * isf_past[i];
223 * Interpolate the fourth ISP vector from current and past frames
224 * to obtain an ISP vector for each subframe.
226 * @param[in,out] isp_q ISPs for each subframe
227 * @param[in] isp4_past Past ISP for subframe 4
229 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
233 for (k = 0; k < 3; k++) {
234 float c = isfp_inter[k];
235 for (i = 0; i < LP_ORDER; i++)
236 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
241 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes).
242 * Calculate integer lag and fractional lag always using 1/4 resolution.
243 * In 1st and 3rd subframes the index is relative to last subframe integer lag.
245 * @param[out] lag_int Decoded integer pitch lag
246 * @param[out] lag_frac Decoded fractional pitch lag
247 * @param[in] pitch_index Adaptive codebook pitch index
248 * @param[in,out] base_lag_int Base integer lag used in relative subframes
249 * @param[in] subframe Current subframe index (0 to 3)
251 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
252 uint8_t *base_lag_int, int subframe)
254 if (subframe == 0 || subframe == 2) {
255 if (pitch_index < 376) {
256 *lag_int = (pitch_index + 137) >> 2;
257 *lag_frac = pitch_index - (*lag_int << 2) + 136;
258 } else if (pitch_index < 440) {
259 *lag_int = (pitch_index + 257 - 376) >> 1;
260 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;
261 /* the actual resolution is 1/2 but expressed as 1/4 */
263 *lag_int = pitch_index - 280;
266 /* minimum lag for next subframe */
267 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
268 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
269 // XXX: the spec states clearly that *base_lag_int should be
270 // the nearest integer to *lag_int (minus 8), but the ref code
271 // actually always uses its floor, I'm following the latter
273 *lag_int = (pitch_index + 1) >> 2;
274 *lag_frac = pitch_index - (*lag_int << 2);
275 *lag_int += *base_lag_int;
280 * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes.
281 * The description is analogous to decode_pitch_lag_high, but in 6k60 the
282 * relative index is used for all subframes except the first.
284 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
285 uint8_t *base_lag_int, int subframe, enum Mode mode)
287 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
288 if (pitch_index < 116) {
289 *lag_int = (pitch_index + 69) >> 1;
290 *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;
292 *lag_int = pitch_index - 24;
295 // XXX: same problem as before
296 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
297 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
299 *lag_int = (pitch_index + 1) >> 1;
300 *lag_frac = (pitch_index - (*lag_int << 1)) << 1;
301 *lag_int += *base_lag_int;
306 * Find the pitch vector by interpolating the past excitation at the
307 * pitch delay, which is obtained in this function.
309 * @param[in,out] ctx The context
310 * @param[in] amr_subframe Current subframe data
311 * @param[in] subframe Current subframe index (0 to 3)
313 static void decode_pitch_vector(AMRWBContext *ctx,
314 const AMRWBSubFrame *amr_subframe,
317 int pitch_lag_int, pitch_lag_frac;
319 float *exc = ctx->excitation;
320 enum Mode mode = ctx->fr_cur_mode;
322 if (mode <= MODE_8k85) {
323 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
324 &ctx->base_pitch_lag, subframe, mode);
326 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
327 &ctx->base_pitch_lag, subframe);
329 ctx->pitch_lag_int = pitch_lag_int;
330 pitch_lag_int += pitch_lag_frac > 0;
332 /* Calculate the pitch vector by interpolating the past excitation at the
333 pitch lag using a hamming windowed sinc function */
334 ctx->acelpf_ctx.acelp_interpolatef(exc,
335 exc + 1 - pitch_lag_int,
337 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
338 LP_ORDER, AMRWB_SFR_SIZE + 1);
340 /* Check which pitch signal path should be used
341 * 6k60 and 8k85 modes have the ltp flag set to 0 */
342 if (amr_subframe->ltp) {
343 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
345 for (i = 0; i < AMRWB_SFR_SIZE; i++)
346 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
348 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
352 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */
353 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))
355 /** Get the bit at specified position */
356 #define BIT_POS(x, p) (((x) >> (p)) & 1)
359 * The next six functions decode_[i]p_track decode exactly i pulses
360 * positions and amplitudes (-1 or 1) in a subframe track using
361 * an encoded pulse indexing (TS 26.190 section 5.8.2).
363 * The results are given in out[], in which a negative number means
364 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ).
366 * @param[out] out Output buffer (writes i elements)
367 * @param[in] code Pulse index (no. of bits varies, see below)
368 * @param[in] m (log2) Number of potential positions
369 * @param[in] off Offset for decoded positions
371 static inline void decode_1p_track(int *out, int code, int m, int off)
373 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits
375 out[0] = BIT_POS(code, m) ? -pos : pos;
378 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits
380 int pos0 = BIT_STR(code, m, m) + off;
381 int pos1 = BIT_STR(code, 0, m) + off;
383 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
384 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
385 out[1] = pos0 > pos1 ? -out[1] : out[1];
388 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits
390 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
392 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
393 m - 1, off + half_2p);
394 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
397 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits
399 int half_4p, subhalf_2p;
400 int b_offset = 1 << (m - 1);
402 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
403 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
404 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
405 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
407 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
408 m - 2, off + half_4p + subhalf_2p);
409 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
410 m - 1, off + half_4p);
412 case 1: /* 1 pulse in A, 3 pulses in B */
413 decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
415 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
416 m - 1, off + b_offset);
418 case 2: /* 2 pulses in each half */
419 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
421 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
422 m - 1, off + b_offset);
424 case 3: /* 3 pulses in A, 1 pulse in B */
425 decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
427 decode_1p_track(out + 3, BIT_STR(code, 0, m),
428 m - 1, off + b_offset);
433 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits
435 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
437 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
438 m - 1, off + half_3p);
440 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
443 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits
445 int b_offset = 1 << (m - 1);
446 /* which half has more pulses in cases 0 to 2 */
447 int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
448 int half_other = b_offset - half_more;
450 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
451 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
452 decode_1p_track(out, BIT_STR(code, 0, m),
453 m - 1, off + half_more);
454 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
455 m - 1, off + half_more);
457 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
458 decode_1p_track(out, BIT_STR(code, 0, m),
459 m - 1, off + half_other);
460 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
461 m - 1, off + half_more);
463 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
464 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
465 m - 1, off + half_other);
466 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
467 m - 1, off + half_more);
469 case 3: /* 3 pulses in A, 3 pulses in B */
470 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
472 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
473 m - 1, off + b_offset);
479 * Decode the algebraic codebook index to pulse positions and signs,
480 * then construct the algebraic codebook vector.
482 * @param[out] fixed_vector Buffer for the fixed codebook excitation
483 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)
484 * @param[in] pulse_lo LSBs part of the pulse index array
485 * @param[in] mode Mode of the current frame
487 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
488 const uint16_t *pulse_lo, const enum Mode mode)
490 /* sig_pos stores for each track the decoded pulse position indexes
491 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
493 int spacing = (mode == MODE_6k60) ? 2 : 4;
498 for (i = 0; i < 2; i++)
499 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
502 for (i = 0; i < 4; i++)
503 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
506 for (i = 0; i < 4; i++)
507 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
510 for (i = 0; i < 2; i++)
511 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
512 for (i = 2; i < 4; i++)
513 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
516 for (i = 0; i < 4; i++)
517 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
520 for (i = 0; i < 4; i++)
521 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
522 ((int) pulse_hi[i] << 14), 4, 1);
525 for (i = 0; i < 2; i++)
526 decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
527 ((int) pulse_hi[i] << 10), 4, 1);
528 for (i = 2; i < 4; i++)
529 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
530 ((int) pulse_hi[i] << 14), 4, 1);
534 for (i = 0; i < 4; i++)
535 decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
536 ((int) pulse_hi[i] << 11), 4, 1);
540 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
542 for (i = 0; i < 4; i++)
543 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
544 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
546 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
551 * Decode pitch gain and fixed gain correction factor.
553 * @param[in] vq_gain Vector-quantized index for gains
554 * @param[in] mode Mode of the current frame
555 * @param[out] fixed_gain_factor Decoded fixed gain correction factor
556 * @param[out] pitch_gain Decoded pitch gain
558 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
559 float *fixed_gain_factor, float *pitch_gain)
561 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
562 qua_gain_7b[vq_gain]);
564 *pitch_gain = gains[0] * (1.0f / (1 << 14));
565 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
569 * Apply pitch sharpening filters to the fixed codebook vector.
571 * @param[in] ctx The context
572 * @param[in,out] fixed_vector Fixed codebook excitation
574 // XXX: Spec states this procedure should be applied when the pitch
575 // lag is less than 64, but this checking seems absent in reference and AMR-NB
576 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
581 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
582 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
584 /* Periodicity enhancement part */
585 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
586 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
590 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced).
592 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors
593 * @param[in] p_gain, f_gain Pitch and fixed gains
594 * @param[in] ctx The context
596 // XXX: There is something wrong with the precision here! The magnitudes
597 // of the energies are not correct. Please check the reference code carefully
598 static float voice_factor(float *p_vector, float p_gain,
599 float *f_vector, float f_gain,
602 double p_ener = (double) ctx->dot_productf(p_vector, p_vector,
603 AMRWB_SFR_SIZE) * p_gain * p_gain;
604 double f_ener = (double) ctx->dot_productf(f_vector, f_vector,
605 AMRWB_SFR_SIZE) * f_gain * f_gain;
607 return (p_ener - f_ener) / (p_ener + f_ener);
611 * Reduce fixed vector sparseness by smoothing with one of three IR filters,
612 * also known as "adaptive phase dispersion".
614 * @param[in] ctx The context
615 * @param[in,out] fixed_vector Unfiltered fixed vector
616 * @param[out] buf Space for modified vector if necessary
618 * @return The potentially overwritten filtered fixed vector address
620 static float *anti_sparseness(AMRWBContext *ctx,
621 float *fixed_vector, float *buf)
625 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
628 if (ctx->pitch_gain[0] < 0.6) {
629 ir_filter_nr = 0; // strong filtering
630 } else if (ctx->pitch_gain[0] < 0.9) {
631 ir_filter_nr = 1; // medium filtering
633 ir_filter_nr = 2; // no filtering
636 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
637 if (ir_filter_nr < 2)
642 for (i = 0; i < 6; i++)
643 if (ctx->pitch_gain[i] < 0.6)
649 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
653 /* update ir filter strength history */
654 ctx->prev_ir_filter_nr = ir_filter_nr;
656 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
658 if (ir_filter_nr < 2) {
660 const float *coef = ir_filters_lookup[ir_filter_nr];
662 /* Circular convolution code in the reference
663 * decoder was modified to avoid using one
664 * extra array. The filtered vector is given by:
666 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
669 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
670 for (i = 0; i < AMRWB_SFR_SIZE; i++)
672 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
681 * Calculate a stability factor {teta} based on distance between
682 * current and past isf. A value of 1 shows maximum signal stability.
684 static float stability_factor(const float *isf, const float *isf_past)
689 for (i = 0; i < LP_ORDER - 1; i++)
690 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
692 // XXX: This part is not so clear from the reference code
693 // the result is more accurate changing the "/ 256" to "* 512"
694 return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
698 * Apply a non-linear fixed gain smoothing in order to reduce
699 * fluctuation in the energy of excitation.
701 * @param[in] fixed_gain Unsmoothed fixed gain
702 * @param[in,out] prev_tr_gain Previous threshold gain (updated)
703 * @param[in] voice_fac Frame voicing factor
704 * @param[in] stab_fac Frame stability factor
706 * @return The smoothed gain
708 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
709 float voice_fac, float stab_fac)
711 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
714 // XXX: the following fixed-point constants used to in(de)crement
715 // gain by 1.5dB were taken from the reference code, maybe it could
717 if (fixed_gain < *prev_tr_gain) {
718 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
719 (6226 * (1.0f / (1 << 15)))); // +1.5 dB
721 g0 = FFMAX(*prev_tr_gain, fixed_gain *
722 (27536 * (1.0f / (1 << 15)))); // -1.5 dB
724 *prev_tr_gain = g0; // update next frame threshold
726 return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
730 * Filter the fixed_vector to emphasize the higher frequencies.
732 * @param[in,out] fixed_vector Fixed codebook vector
733 * @param[in] voice_fac Frame voicing factor
735 static void pitch_enhancer(float *fixed_vector, float voice_fac)
738 float cpe = 0.125 * (1 + voice_fac);
739 float last = fixed_vector[0]; // holds c(i - 1)
741 fixed_vector[0] -= cpe * fixed_vector[1];
743 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
744 float cur = fixed_vector[i];
746 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
750 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
754 * Conduct 16th order linear predictive coding synthesis from excitation.
756 * @param[in] ctx Pointer to the AMRWBContext
757 * @param[in] lpc Pointer to the LPC coefficients
758 * @param[out] excitation Buffer for synthesis final excitation
759 * @param[in] fixed_gain Fixed codebook gain for synthesis
760 * @param[in] fixed_vector Algebraic codebook vector
761 * @param[in,out] samples Pointer to the output samples and memory
763 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
764 float fixed_gain, const float *fixed_vector,
767 ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
768 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
770 /* emphasize pitch vector contribution in low bitrate modes */
771 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
773 float energy = ctx->celpm_ctx.dot_productf(excitation, excitation,
776 // XXX: Weird part in both ref code and spec. A unknown parameter
777 // {beta} seems to be identical to the current pitch gain
778 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
780 for (i = 0; i < AMRWB_SFR_SIZE; i++)
781 excitation[i] += pitch_factor * ctx->pitch_vector[i];
783 ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
784 energy, AMRWB_SFR_SIZE);
787 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
788 AMRWB_SFR_SIZE, LP_ORDER);
792 * Apply to synthesis a de-emphasis filter of the form:
793 * H(z) = 1 / (1 - m * z^-1)
795 * @param[out] out Output buffer
796 * @param[in] in Input samples array with in[-1]
797 * @param[in] m Filter coefficient
798 * @param[in,out] mem State from last filtering
800 static void de_emphasis(float *out, float *in, float m, float mem[1])
804 out[0] = in[0] + m * mem[0];
806 for (i = 1; i < AMRWB_SFR_SIZE; i++)
807 out[i] = in[i] + out[i - 1] * m;
809 mem[0] = out[AMRWB_SFR_SIZE - 1];
813 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using
814 * a FIR interpolation filter. Uses past data from before *in address.
816 * @param[out] out Buffer for interpolated signal
817 * @param[in] in Current signal data (length 0.8*o_size)
818 * @param[in] o_size Output signal length
819 * @param[in] ctx The context
821 static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx)
823 const float *in0 = in - UPS_FIR_SIZE + 1;
825 int int_part = 0, frac_part;
828 for (j = 0; j < o_size / 5; j++) {
829 out[i] = in[int_part];
833 for (k = 1; k < 5; k++) {
834 out[i] = ctx->dot_productf(in0 + int_part,
835 upsample_fir[4 - frac_part],
845 * Calculate the high-band gain based on encoded index (23k85 mode) or
846 * on the low-band speech signal and the Voice Activity Detection flag.
848 * @param[in] ctx The context
849 * @param[in] synth LB speech synthesis at 12.8k
850 * @param[in] hb_idx Gain index for mode 23k85 only
851 * @param[in] vad VAD flag for the frame
853 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
854 uint16_t hb_idx, uint8_t vad)
859 if (ctx->fr_cur_mode == MODE_23k85)
860 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
862 tilt = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
863 ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE);
865 /* return gain bounded by [0.1, 1.0] */
866 return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
870 * Generate the high-band excitation with the same energy from the lower
871 * one and scaled by the given gain.
873 * @param[in] ctx The context
874 * @param[out] hb_exc Buffer for the excitation
875 * @param[in] synth_exc Low-band excitation used for synthesis
876 * @param[in] hb_gain Wanted excitation gain
878 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
879 const float *synth_exc, float hb_gain)
882 float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc, AMRWB_SFR_SIZE);
884 /* Generate a white-noise excitation */
885 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
886 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
888 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc,
889 energy * hb_gain * hb_gain,
894 * Calculate the auto-correlation for the ISF difference vector.
896 static float auto_correlation(float *diff_isf, float mean, int lag)
901 for (i = 7; i < LP_ORDER - 2; i++) {
902 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
909 * Extrapolate a ISF vector to the 16kHz range (20th order LP)
910 * used at mode 6k60 LP filter for the high frequency band.
912 * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER
915 static void extrapolate_isf(float isf[LP_ORDER_16k])
917 float diff_isf[LP_ORDER - 2], diff_mean;
918 float *diff_hi = diff_isf - LP_ORDER + 1; // diff array for extrapolated indexes
923 isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
925 /* Calculate the difference vector */
926 for (i = 0; i < LP_ORDER - 2; i++)
927 diff_isf[i] = isf[i + 1] - isf[i];
930 for (i = 2; i < LP_ORDER - 2; i++)
931 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
933 /* Find which is the maximum autocorrelation */
935 for (i = 0; i < 3; i++) {
936 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
938 if (corr_lag[i] > corr_lag[i_max_corr])
943 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
944 isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
945 - isf[i - 2 - i_max_corr];
947 /* Calculate an estimate for ISF(18) and scale ISF based on the error */
948 est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;
949 scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /
950 (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);
952 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
953 diff_hi[i] = scale * (isf[i] - isf[i - 1]);
955 /* Stability insurance */
956 for (i = LP_ORDER; i < LP_ORDER_16k - 1; i++)
957 if (diff_hi[i] + diff_hi[i - 1] < 5.0) {
958 if (diff_hi[i] > diff_hi[i - 1]) {
959 diff_hi[i - 1] = 5.0 - diff_hi[i];
961 diff_hi[i] = 5.0 - diff_hi[i - 1];
964 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
965 isf[i] = isf[i - 1] + diff_hi[i] * (1.0f / (1 << 15));
967 /* Scale the ISF vector for 16000 Hz */
968 for (i = 0; i < LP_ORDER_16k - 1; i++)
973 * Spectral expand the LP coefficients using the equation:
974 * y[i] = x[i] * (gamma ** i)
976 * @param[out] out Output buffer (may use input array)
977 * @param[in] lpc LP coefficients array
978 * @param[in] gamma Weighting factor
979 * @param[in] size LP array size
981 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
986 for (i = 0; i < size; i++) {
987 out[i] = lpc[i] * fac;
993 * Conduct 20th order linear predictive coding synthesis for the high
994 * frequency band excitation at 16kHz.
996 * @param[in] ctx The context
997 * @param[in] subframe Current subframe index (0 to 3)
998 * @param[in,out] samples Pointer to the output speech samples
999 * @param[in] exc Generated white-noise scaled excitation
1000 * @param[in] isf Current frame isf vector
1001 * @param[in] isf_past Past frame final isf vector
1003 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
1004 const float *exc, const float *isf, const float *isf_past)
1006 float hb_lpc[LP_ORDER_16k];
1007 enum Mode mode = ctx->fr_cur_mode;
1009 if (mode == MODE_6k60) {
1010 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
1011 double e_isp[LP_ORDER_16k];
1013 ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1014 1.0 - isfp_inter[subframe], LP_ORDER);
1016 extrapolate_isf(e_isf);
1018 e_isf[LP_ORDER_16k - 1] *= 2.0;
1019 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1020 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1022 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1024 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1027 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1028 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1032 * Apply a 15th order filter to high-band samples.
1033 * The filter characteristic depends on the given coefficients.
1035 * @param[out] out Buffer for filtered output
1036 * @param[in] fir_coef Filter coefficients
1037 * @param[in,out] mem State from last filtering (updated)
1038 * @param[in] in Input speech data (high-band)
1040 * @remark It is safe to pass the same array in in and out parameters
1043 #ifndef hb_fir_filter
1044 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1045 float mem[HB_FIR_SIZE], const float *in)
1048 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1050 memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1051 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1053 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1055 for (j = 0; j <= HB_FIR_SIZE; j++)
1056 out[i] += data[i + j] * fir_coef[j];
1059 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1061 #endif /* hb_fir_filter */
1064 * Update context state before the next subframe.
1066 static void update_sub_state(AMRWBContext *ctx)
1068 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1069 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1071 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1072 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1074 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1075 LP_ORDER * sizeof(float));
1076 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1077 UPS_MEM_SIZE * sizeof(float));
1078 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1079 LP_ORDER_16k * sizeof(float));
1082 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,
1083 int *got_frame_ptr, AVPacket *avpkt)
1085 AMRWBContext *ctx = avctx->priv_data;
1086 AMRWBFrame *cf = &ctx->frame;
1087 const uint8_t *buf = avpkt->data;
1088 int buf_size = avpkt->size;
1089 int expected_fr_size, header_size;
1091 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1092 float fixed_gain_factor; // fixed gain correction factor (gamma)
1093 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1094 float synth_fixed_gain; // the fixed gain that synthesis should use
1095 float voice_fac, stab_fac; // parameters used for gain smoothing
1096 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1097 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1098 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1102 /* get output buffer */
1103 ctx->avframe.nb_samples = 4 * AMRWB_SFR_SIZE_16k;
1104 if ((ret = avctx->get_buffer(avctx, &ctx->avframe)) < 0) {
1105 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1108 buf_out = (float *)ctx->avframe.data[0];
1110 header_size = decode_mime_header(ctx, buf);
1111 if (ctx->fr_cur_mode > MODE_SID) {
1112 av_log(avctx, AV_LOG_ERROR,
1113 "Invalid mode %d\n", ctx->fr_cur_mode);
1114 return AVERROR_INVALIDDATA;
1116 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1118 if (buf_size < expected_fr_size) {
1119 av_log(avctx, AV_LOG_ERROR,
1120 "Frame too small (%d bytes). Truncated file?\n", buf_size);
1122 return AVERROR_INVALIDDATA;
1125 if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1126 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1128 if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */
1129 av_log_missing_feature(avctx, "SID mode", 1);
1133 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1134 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1136 /* Decode the quantized ISF vector */
1137 if (ctx->fr_cur_mode == MODE_6k60) {
1138 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur);
1140 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur);
1143 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past);
1144 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1);
1146 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1148 ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1149 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1151 /* Generate a ISP vector for each subframe */
1152 if (ctx->first_frame) {
1153 ctx->first_frame = 0;
1154 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1156 interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1158 for (sub = 0; sub < 4; sub++)
1159 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1161 for (sub = 0; sub < 4; sub++) {
1162 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1163 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1165 /* Decode adaptive codebook (pitch vector) */
1166 decode_pitch_vector(ctx, cur_subframe, sub);
1167 /* Decode innovative codebook (fixed vector) */
1168 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1169 cur_subframe->pul_il, ctx->fr_cur_mode);
1171 pitch_sharpening(ctx, ctx->fixed_vector);
1173 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1174 &fixed_gain_factor, &ctx->pitch_gain[0]);
1176 ctx->fixed_gain[0] =
1177 ff_amr_set_fixed_gain(fixed_gain_factor,
1178 ctx->celpm_ctx.dot_productf(ctx->fixed_vector, ctx->fixed_vector,
1179 AMRWB_SFR_SIZE) / AMRWB_SFR_SIZE,
1180 ctx->prediction_error,
1181 ENERGY_MEAN, energy_pred_fac);
1183 /* Calculate voice factor and store tilt for next subframe */
1184 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1185 ctx->fixed_vector, ctx->fixed_gain[0],
1187 ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1189 /* Construct current excitation */
1190 for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1191 ctx->excitation[i] *= ctx->pitch_gain[0];
1192 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1193 ctx->excitation[i] = truncf(ctx->excitation[i]);
1196 /* Post-processing of excitation elements */
1197 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1198 voice_fac, stab_fac);
1200 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1203 pitch_enhancer(synth_fixed_vector, voice_fac);
1205 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1206 synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1208 /* Synthesis speech post-processing */
1209 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],
1210 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1212 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
1213 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,
1214 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1216 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1217 AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx);
1219 /* High frequency band (6.4 - 7.0 kHz) generation part */
1220 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(hb_samples,
1221 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,
1222 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1224 hb_gain = find_hb_gain(ctx, hb_samples,
1225 cur_subframe->hb_gain, cf->vad);
1227 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1229 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1230 hb_exc, ctx->isf_cur, ctx->isf_past_final);
1232 /* High-band post-processing filters */
1233 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1234 &ctx->samples_hb[LP_ORDER_16k]);
1236 if (ctx->fr_cur_mode == MODE_23k85)
1237 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1240 /* Add the low and high frequency bands */
1241 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1242 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1244 /* Update buffers and history */
1245 update_sub_state(ctx);
1248 /* update state for next frame */
1249 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1250 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1253 *(AVFrame *)data = ctx->avframe;
1255 return expected_fr_size;
1258 AVCodec ff_amrwb_decoder = {
1260 .type = AVMEDIA_TYPE_AUDIO,
1261 .id = AV_CODEC_ID_AMR_WB,
1262 .priv_data_size = sizeof(AMRWBContext),
1263 .init = amrwb_decode_init,
1264 .decode = amrwb_decode_frame,
1265 .capabilities = CODEC_CAP_DR1,
1266 .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"),
1267 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1268 AV_SAMPLE_FMT_NONE },