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/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"
44 AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream
45 enum Mode fr_cur_mode; ///< mode index of current frame
46 uint8_t fr_quality; ///< frame quality index (FQI)
47 float isf_cur[LP_ORDER]; ///< working ISF vector from current frame
48 float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame
49 float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame
50 double isp[4][LP_ORDER]; ///< ISP vectors from current frame
51 double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame
53 float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector
55 uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe
56 uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe
58 float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history
59 float *excitation; ///< points to current excitation in excitation_buf[]
61 float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe
62 float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe
64 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
65 float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes
66 float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes
68 float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe
70 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset"
71 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
72 float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold
74 float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz
75 float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling
76 float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz
78 float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters
79 float demph_mem[1]; ///< previous value in the de-emphasis filter
80 float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter
81 float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter
83 AVLFG prng; ///< random number generator for white noise excitation
84 uint8_t first_frame; ///< flag active during decoding of the first frame
87 static av_cold int amrwb_decode_init(AVCodecContext *avctx)
89 AMRWBContext *ctx = avctx->priv_data;
92 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
94 av_lfg_init(&ctx->prng, 1);
96 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1];
99 for (i = 0; i < LP_ORDER; i++)
100 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));
102 for (i = 0; i < 4; i++)
103 ctx->prediction_error[i] = MIN_ENERGY;
109 * Decode the frame header in the "MIME/storage" format. This format
110 * is simpler and does not carry the auxiliary information of the frame
112 * @param[in] ctx The Context
113 * @param[in] buf Pointer to the input buffer
115 * @return The decoded header length in bytes
117 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)
120 init_get_bits(&gb, buf, 8);
122 /* Decode frame header (1st octet) */
123 skip_bits(&gb, 1); // padding bit
124 ctx->fr_cur_mode = get_bits(&gb, 4);
125 ctx->fr_quality = get_bits1(&gb);
126 skip_bits(&gb, 2); // padding bits
132 * Decodes quantized ISF vectors using 36-bit indexes (6K60 mode only)
134 * @param[in] ind Array of 5 indexes
135 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
138 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)
142 for (i = 0; i < 9; i++)
143 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
145 for (i = 0; i < 7; i++)
146 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
148 for (i = 0; i < 5; i++)
149 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));
151 for (i = 0; i < 4; i++)
152 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));
154 for (i = 0; i < 7; i++)
155 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));
159 * Decodes quantized ISF vectors using 46-bit indexes (except 6K60 mode)
161 * @param[in] ind Array of 7 indexes
162 * @param[out] isf_q Buffer for isf_q[LP_ORDER]
165 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)
169 for (i = 0; i < 9; i++)
170 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));
172 for (i = 0; i < 7; i++)
173 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));
175 for (i = 0; i < 3; i++)
176 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));
178 for (i = 0; i < 3; i++)
179 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));
181 for (i = 0; i < 3; i++)
182 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));
184 for (i = 0; i < 3; i++)
185 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));
187 for (i = 0; i < 4; i++)
188 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));
192 * Apply mean and past ISF values using the prediction factor
193 * Updates past ISF vector
195 * @param[in,out] isf_q Current quantized ISF
196 * @param[in,out] isf_past Past quantized ISF
199 static void isf_add_mean_and_past(float *isf_q, float *isf_past)
204 for (i = 0; i < LP_ORDER; i++) {
206 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));
207 isf_q[i] += PRED_FACTOR * isf_past[i];
213 * Interpolate the fourth ISP vector from current and past frames
214 * to obtain a ISP vector for each subframe
216 * @param[in,out] isp_q ISPs for each subframe
217 * @param[in] isp4_past Past ISP for subframe 4
219 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)
223 for (k = 0; k < 3; k++) {
224 float c = isfp_inter[k];
225 for (i = 0; i < LP_ORDER; i++)
226 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];
231 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes)
232 * Calculate integer lag and fractional lag always using 1/4 resolution
233 * In 1st and 3rd subframes the index is relative to last subframe integer lag
235 * @param[out] lag_int Decoded integer pitch lag
236 * @param[out] lag_frac Decoded fractional pitch lag
237 * @param[in] pitch_index Adaptive codebook pitch index
238 * @param[in,out] base_lag_int Base integer lag used in relative subframes
239 * @param[in] subframe Current subframe index (0 to 3)
241 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,
242 uint8_t *base_lag_int, int subframe)
244 if (subframe == 0 || subframe == 2) {
245 if (pitch_index < 376) {
246 *lag_int = (pitch_index + 137) >> 2;
247 *lag_frac = pitch_index - (*lag_int << 2) + 136;
248 } else if (pitch_index < 440) {
249 *lag_int = (pitch_index + 257 - 376) >> 1;
250 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;
251 /* the actual resolution is 1/2 but expressed as 1/4 */
253 *lag_int = pitch_index - 280;
256 /* minimum lag for next subframe */
257 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
258 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
259 // XXX: the spec states clearly that *base_lag_int should be
260 // the nearest integer to *lag_int (minus 8), but the ref code
261 // actually always uses its floor, I'm following the latter
263 *lag_int = (pitch_index + 1) >> 2;
264 *lag_frac = pitch_index - (*lag_int << 2);
265 *lag_int += *base_lag_int;
270 * Decode a adaptive codebook index into pitch lag for 8k85 and 6k60 modes
271 * Description is analogous to decode_pitch_lag_high, but in 6k60 relative
272 * index is used for all subframes except the first
274 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,
275 uint8_t *base_lag_int, int subframe, enum Mode mode)
277 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {
278 if (pitch_index < 116) {
279 *lag_int = (pitch_index + 69) >> 1;
280 *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;
282 *lag_int = pitch_index - 24;
285 // XXX: same problem as before
286 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),
287 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);
289 *lag_int = (pitch_index + 1) >> 1;
290 *lag_frac = (pitch_index - (*lag_int << 1)) << 1;
291 *lag_int += *base_lag_int;
296 * Find the pitch vector by interpolating the past excitation at the
297 * pitch delay, which is obtained in this function
299 * @param[in,out] ctx The context
300 * @param[in] amr_subframe Current subframe data
301 * @param[in] subframe Current subframe index (0 to 3)
303 static void decode_pitch_vector(AMRWBContext *ctx,
304 const AMRWBSubFrame *amr_subframe,
307 int pitch_lag_int, pitch_lag_frac;
309 float *exc = ctx->excitation;
310 enum Mode mode = ctx->fr_cur_mode;
312 if (mode <= MODE_8k85) {
313 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
314 &ctx->base_pitch_lag, subframe, mode);
316 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,
317 &ctx->base_pitch_lag, subframe);
319 ctx->pitch_lag_int = pitch_lag_int;
320 pitch_lag_int += pitch_lag_frac > 0;
322 /* Calculate the pitch vector by interpolating the past excitation at the
323 pitch lag using a hamming windowed sinc function */
324 ff_acelp_interpolatef(exc, exc + 1 - pitch_lag_int,
326 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),
327 LP_ORDER, AMRWB_SFR_SIZE + 1);
329 /* Check which pitch signal path should be used
330 * 6k60 and 8k85 modes have the ltp flag set to 0 */
331 if (amr_subframe->ltp) {
332 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));
334 for (i = 0; i < AMRWB_SFR_SIZE; i++)
335 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +
337 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));
341 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */
342 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))
344 /** Get the bit at specified position */
345 #define BIT_POS(x, p) (((x) >> (p)) & 1)
348 * The next six functions decode_[i]p_track decode exactly i pulses
349 * positions and amplitudes (-1 or 1) in a subframe track using
350 * an encoded pulse indexing (TS 26.190 section 5.8.2)
352 * The results are given in out[], in which a negative number means
353 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) )
355 * @param[out] out Output buffer (writes i elements)
356 * @param[in] code Pulse index (no. of bits varies, see below)
357 * @param[in] m (log2) Number of potential positions
358 * @param[in] off Offset for decoded positions
360 static inline void decode_1p_track(int *out, int code, int m, int off)
362 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits
364 out[0] = BIT_POS(code, m) ? -pos : pos;
367 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits
369 int pos0 = BIT_STR(code, m, m) + off;
370 int pos1 = BIT_STR(code, 0, m) + off;
372 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;
373 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;
374 out[1] = pos0 > pos1 ? -out[1] : out[1];
377 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits
379 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);
381 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
382 m - 1, off + half_2p);
383 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);
386 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits
388 int half_4p, subhalf_2p;
389 int b_offset = 1 << (m - 1);
391 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */
392 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */
393 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses
394 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);
396 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),
397 m - 2, off + half_4p + subhalf_2p);
398 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),
399 m - 1, off + half_4p);
401 case 1: /* 1 pulse in A, 3 pulses in B */
402 decode_1p_track(out, BIT_STR(code, 3*m - 2, m),
404 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),
405 m - 1, off + b_offset);
407 case 2: /* 2 pulses in each half */
408 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),
410 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),
411 m - 1, off + b_offset);
413 case 3: /* 3 pulses in A, 1 pulse in B */
414 decode_3p_track(out, BIT_STR(code, m, 3*m - 2),
416 decode_1p_track(out + 3, BIT_STR(code, 0, m),
417 m - 1, off + b_offset);
422 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits
424 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);
426 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),
427 m - 1, off + half_3p);
429 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);
432 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits
434 int b_offset = 1 << (m - 1);
435 /* which half has more pulses in cases 0 to 2 */
436 int half_more = BIT_POS(code, 6*m - 5) << (m - 1);
437 int half_other = b_offset - half_more;
439 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */
440 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */
441 decode_1p_track(out, BIT_STR(code, 0, m),
442 m - 1, off + half_more);
443 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
444 m - 1, off + half_more);
446 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */
447 decode_1p_track(out, BIT_STR(code, 0, m),
448 m - 1, off + half_other);
449 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),
450 m - 1, off + half_more);
452 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */
453 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),
454 m - 1, off + half_other);
455 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),
456 m - 1, off + half_more);
458 case 3: /* 3 pulses in A, 3 pulses in B */
459 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),
461 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),
462 m - 1, off + b_offset);
468 * Decode the algebraic codebook index to pulse positions and signs,
469 * then construct the algebraic codebook vector
471 * @param[out] fixed_vector Buffer for the fixed codebook excitation
472 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)
473 * @param[in] pulse_lo LSBs part of the pulse index array
474 * @param[in] mode Mode of the current frame
476 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,
477 const uint16_t *pulse_lo, const enum Mode mode)
479 /* sig_pos stores for each track the decoded pulse position indexes
480 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */
482 int spacing = (mode == MODE_6k60) ? 2 : 4;
487 for (i = 0; i < 2; i++)
488 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);
491 for (i = 0; i < 4; i++)
492 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);
495 for (i = 0; i < 4; i++)
496 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
499 for (i = 0; i < 2; i++)
500 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
501 for (i = 2; i < 4; i++)
502 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);
505 for (i = 0; i < 4; i++)
506 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);
509 for (i = 0; i < 4; i++)
510 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
511 ((int) pulse_hi[i] << 14), 4, 1);
514 for (i = 0; i < 2; i++)
515 decode_5p_track(sig_pos[i], (int) pulse_lo[i] +
516 ((int) pulse_hi[i] << 10), 4, 1);
517 for (i = 2; i < 4; i++)
518 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +
519 ((int) pulse_hi[i] << 14), 4, 1);
523 for (i = 0; i < 4; i++)
524 decode_6p_track(sig_pos[i], (int) pulse_lo[i] +
525 ((int) pulse_hi[i] << 11), 4, 1);
529 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);
531 for (i = 0; i < 4; i++)
532 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {
533 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;
535 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;
540 * Decode pitch gain and fixed gain correction factor
542 * @param[in] vq_gain Vector-quantized index for gains
543 * @param[in] mode Mode of the current frame
544 * @param[out] fixed_gain_factor Decoded fixed gain correction factor
545 * @param[out] pitch_gain Decoded pitch gain
547 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,
548 float *fixed_gain_factor, float *pitch_gain)
550 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :
551 qua_gain_7b[vq_gain]);
553 *pitch_gain = gains[0] * (1.0f / (1 << 14));
554 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));
558 * Apply pitch sharpening filters to the fixed codebook vector
560 * @param[in] ctx The context
561 * @param[in,out] fixed_vector Fixed codebook excitation
563 // XXX: Spec states this procedure should be applied when the pitch
564 // lag is less than 64, but this checking seems absent in reference and AMR-NB
565 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)
570 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)
571 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;
573 /* Periodicity enhancement part */
574 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)
575 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;
579 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced)
581 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors
582 * @param[in] p_gain, f_gain Pitch and fixed gains
584 // XXX: There is something wrong with the precision here! The magnitudes
585 // of the energies are not correct. Please check the reference code carefully
586 static float voice_factor(float *p_vector, float p_gain,
587 float *f_vector, float f_gain)
589 double p_ener = (double) ff_dot_productf(p_vector, p_vector,
590 AMRWB_SFR_SIZE) * p_gain * p_gain;
591 double f_ener = (double) ff_dot_productf(f_vector, f_vector,
592 AMRWB_SFR_SIZE) * f_gain * f_gain;
594 return (p_ener - f_ener) / (p_ener + f_ener);
598 * Reduce fixed vector sparseness by smoothing with one of three IR filters
599 * Also known as "adaptive phase dispersion"
601 * @param[in] ctx The context
602 * @param[in,out] fixed_vector Unfiltered fixed vector
603 * @param[out] buf Space for modified vector if necessary
605 * @return The potentially overwritten filtered fixed vector address
607 static float *anti_sparseness(AMRWBContext *ctx,
608 float *fixed_vector, float *buf)
612 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes
615 if (ctx->pitch_gain[0] < 0.6) {
616 ir_filter_nr = 0; // strong filtering
617 } else if (ctx->pitch_gain[0] < 0.9) {
618 ir_filter_nr = 1; // medium filtering
620 ir_filter_nr = 2; // no filtering
623 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {
624 if (ir_filter_nr < 2)
629 for (i = 0; i < 6; i++)
630 if (ctx->pitch_gain[i] < 0.6)
636 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)
640 /* update ir filter strength history */
641 ctx->prev_ir_filter_nr = ir_filter_nr;
643 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);
645 if (ir_filter_nr < 2) {
647 const float *coef = ir_filters_lookup[ir_filter_nr];
649 /* Circular convolution code in the reference
650 * decoder was modified to avoid using one
651 * extra array. The filtered vector is given by:
653 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }
656 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);
657 for (i = 0; i < AMRWB_SFR_SIZE; i++)
659 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],
668 * Calculate a stability factor {teta} based on distance between
669 * current and past isf. A value of 1 shows maximum signal stability
671 static float stability_factor(const float *isf, const float *isf_past)
676 for (i = 0; i < LP_ORDER - 1; i++)
677 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);
679 // XXX: This part is not so clear from the reference code
680 // the result is more accurate changing the "/ 256" to "* 512"
681 return FFMAX(0.0, 1.25 - acc * 0.8 * 512);
685 * Apply a non-linear fixed gain smoothing in order to reduce
686 * fluctuation in the energy of excitation
688 * @param[in] fixed_gain Unsmoothed fixed gain
689 * @param[in,out] prev_tr_gain Previous threshold gain (updated)
690 * @param[in] voice_fac Frame voicing factor
691 * @param[in] stab_fac Frame stability factor
693 * @return The smoothed gain
695 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,
696 float voice_fac, float stab_fac)
698 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;
701 // XXX: the following fixed-point constants used to in(de)crement
702 // gain by 1.5dB were taken from the reference code, maybe it could
704 if (fixed_gain < *prev_tr_gain) {
705 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *
706 (6226 * (1.0f / (1 << 15)))); // +1.5 dB
708 g0 = FFMAX(*prev_tr_gain, fixed_gain *
709 (27536 * (1.0f / (1 << 15)))); // -1.5 dB
711 *prev_tr_gain = g0; // update next frame threshold
713 return sm_fac * g0 + (1 - sm_fac) * fixed_gain;
717 * Filter the fixed_vector to emphasize the higher frequencies
719 * @param[in,out] fixed_vector Fixed codebook vector
720 * @param[in] voice_fac Frame voicing factor
722 static void pitch_enhancer(float *fixed_vector, float voice_fac)
725 float cpe = 0.125 * (1 + voice_fac);
726 float last = fixed_vector[0]; // holds c(i - 1)
728 fixed_vector[0] -= cpe * fixed_vector[1];
730 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {
731 float cur = fixed_vector[i];
733 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);
737 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;
741 * Conduct 16th order linear predictive coding synthesis from excitation
743 * @param[in] ctx Pointer to the AMRWBContext
744 * @param[in] lpc Pointer to the LPC coefficients
745 * @param[out] excitation Buffer for synthesis final excitation
746 * @param[in] fixed_gain Fixed codebook gain for synthesis
747 * @param[in] fixed_vector Algebraic codebook vector
748 * @param[in,out] samples Pointer to the output samples and memory
750 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,
751 float fixed_gain, const float *fixed_vector,
754 ff_weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,
755 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);
757 /* emphasize pitch vector contribution in low bitrate modes */
758 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {
760 float energy = ff_dot_productf(excitation, excitation,
763 // XXX: Weird part in both ref code and spec. A unknown parameter
764 // {beta} seems to be identical to the current pitch gain
765 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];
767 for (i = 0; i < AMRWB_SFR_SIZE; i++)
768 excitation[i] += pitch_factor * ctx->pitch_vector[i];
770 ff_scale_vector_to_given_sum_of_squares(excitation, excitation,
771 energy, AMRWB_SFR_SIZE);
774 ff_celp_lp_synthesis_filterf(samples, lpc, excitation,
775 AMRWB_SFR_SIZE, LP_ORDER);
779 * Apply to synthesis a de-emphasis filter of the form:
780 * H(z) = 1 / (1 - m * z^-1)
782 * @param[out] out Output buffer
783 * @param[in] in Input samples array with in[-1]
784 * @param[in] m Filter coefficient
785 * @param[in,out] mem State from last filtering
787 static void de_emphasis(float *out, float *in, float m, float mem[1])
791 out[0] = in[0] + m * mem[0];
793 for (i = 1; i < AMRWB_SFR_SIZE; i++)
794 out[i] = in[i] + out[i - 1] * m;
796 mem[0] = out[AMRWB_SFR_SIZE - 1];
800 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using
801 * a FIR interpolation filter. Uses past data from before *in address
803 * @param[out] out Buffer for interpolated signal
804 * @param[in] in Current signal data (length 0.8*o_size)
805 * @param[in] o_size Output signal length
807 static void upsample_5_4(float *out, const float *in, int o_size)
809 const float *in0 = in - UPS_FIR_SIZE + 1;
811 int int_part = 0, frac_part;
814 for (j = 0; j < o_size / 5; j++) {
815 out[i] = in[int_part];
819 for (k = 1; k < 5; k++) {
820 out[i] = ff_dot_productf(in0 + int_part, upsample_fir[4 - frac_part],
830 * Calculate the high-band gain based on encoded index (23k85 mode) or
831 * on the low-band speech signal and the Voice Activity Detection flag
833 * @param[in] ctx The context
834 * @param[in] synth LB speech synthesis at 12.8k
835 * @param[in] hb_idx Gain index for mode 23k85 only
836 * @param[in] vad VAD flag for the frame
838 static float find_hb_gain(AMRWBContext *ctx, const float *synth,
839 uint16_t hb_idx, uint8_t vad)
844 if (ctx->fr_cur_mode == MODE_23k85)
845 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));
847 tilt = ff_dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /
848 ff_dot_productf(synth, synth, AMRWB_SFR_SIZE);
850 /* return gain bounded by [0.1, 1.0] */
851 return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);
855 * Generate the high-band excitation with the same energy from the lower
856 * one and scaled by the given gain
858 * @param[in] ctx The context
859 * @param[out] hb_exc Buffer for the excitation
860 * @param[in] synth_exc Low-band excitation used for synthesis
861 * @param[in] hb_gain Wanted excitation gain
863 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,
864 const float *synth_exc, float hb_gain)
867 float energy = ff_dot_productf(synth_exc, synth_exc, AMRWB_SFR_SIZE);
869 /* Generate a white-noise excitation */
870 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
871 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);
873 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc,
874 energy * hb_gain * hb_gain,
879 * Calculate the auto-correlation for the ISF difference vector
881 static float auto_correlation(float *diff_isf, float mean, int lag)
886 for (i = 7; i < LP_ORDER - 2; i++) {
887 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);
894 * Extrapolate a ISF vector to the 16kHz range (20th order LP)
895 * used at mode 6k60 LP filter for the high frequency band
897 * @param[out] out Buffer for extrapolated isf
898 * @param[in] isf Input isf vector
900 static void extrapolate_isf(float out[LP_ORDER_16k], float isf[LP_ORDER])
902 float diff_isf[LP_ORDER - 2], diff_mean;
903 float *diff_hi = diff_isf - LP_ORDER + 1; // diff array for extrapolated indexes
908 memcpy(out, isf, (LP_ORDER - 1) * sizeof(float));
909 out[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];
911 /* Calculate the difference vector */
912 for (i = 0; i < LP_ORDER - 2; i++)
913 diff_isf[i] = isf[i + 1] - isf[i];
916 for (i = 2; i < LP_ORDER - 2; i++)
917 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));
919 /* Find which is the maximum autocorrelation */
921 for (i = 0; i < 3; i++) {
922 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);
924 if (corr_lag[i] > corr_lag[i_max_corr])
929 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
930 out[i] = isf[i - 1] + isf[i - 1 - i_max_corr]
931 - isf[i - 2 - i_max_corr];
933 /* Calculate an estimate for ISF(18) and scale ISF based on the error */
934 est = 7965 + (out[2] - out[3] - out[4]) / 6.0;
935 scale = 0.5 * (FFMIN(est, 7600) - out[LP_ORDER - 2]) /
936 (out[LP_ORDER_16k - 2] - out[LP_ORDER - 2]);
938 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
939 diff_hi[i] = scale * (out[i] - out[i - 1]);
941 /* Stability insurance */
942 for (i = LP_ORDER; i < LP_ORDER_16k - 1; i++)
943 if (diff_hi[i] + diff_hi[i - 1] < 5.0) {
944 if (diff_hi[i] > diff_hi[i - 1]) {
945 diff_hi[i - 1] = 5.0 - diff_hi[i];
947 diff_hi[i] = 5.0 - diff_hi[i - 1];
950 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)
951 out[i] = out[i - 1] + diff_hi[i] * (1.0f / (1 << 15));
953 /* Scale the ISF vector for 16000 Hz */
954 for (i = 0; i < LP_ORDER_16k - 1; i++)
959 * Spectral expand the LP coefficients using the equation:
960 * y[i] = x[i] * (gamma ** i)
962 * @param[out] out Output buffer (may use input array)
963 * @param[in] lpc LP coefficients array
964 * @param[in] gamma Weighting factor
965 * @param[in] size LP array size
967 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)
972 for (i = 0; i < size; i++) {
973 out[i] = lpc[i] * fac;
979 * Conduct 20th order linear predictive coding synthesis for the high
980 * frequency band excitation at 16kHz
982 * @param[in] ctx The context
983 * @param[in] subframe Current subframe index (0 to 3)
984 * @param[in,out] samples Pointer to the output speech samples
985 * @param[in] exc Generated white-noise scaled excitation
986 * @param[in] isf Current frame isf vector
987 * @param[in] isf_past Past frame final isf vector
989 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,
990 const float *exc, const float *isf, const float *isf_past)
992 float hb_lpc[LP_ORDER_16k];
993 enum Mode mode = ctx->fr_cur_mode;
995 if (mode == MODE_6k60) {
996 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation
997 double e_isp[LP_ORDER_16k];
999 ff_weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],
1000 1.0 - isfp_inter[subframe], LP_ORDER);
1002 extrapolate_isf(e_isf, e_isf);
1004 e_isf[LP_ORDER_16k - 1] *= 2.0;
1005 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);
1006 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);
1008 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);
1010 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);
1013 ff_celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,
1014 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);
1018 * Apply to high-band samples a 15th order filter
1019 * The filter characteristic depends on the given coefficients
1021 * @param[out] out Buffer for filtered output
1022 * @param[in] fir_coef Filter coefficients
1023 * @param[in,out] mem State from last filtering (updated)
1024 * @param[in] in Input speech data (high-band)
1026 * @remark It is safe to pass the same array in in and out parameters
1028 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],
1029 float mem[HB_FIR_SIZE], const float *in)
1032 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples
1034 memcpy(data, mem, HB_FIR_SIZE * sizeof(float));
1035 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));
1037 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {
1039 for (j = 0; j <= HB_FIR_SIZE; j++)
1040 out[i] += data[i + j] * fir_coef[j];
1043 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));
1047 * Update context state before the next subframe
1049 static void update_sub_state(AMRWBContext *ctx)
1051 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],
1052 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));
1054 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));
1055 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));
1057 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],
1058 LP_ORDER * sizeof(float));
1059 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],
1060 UPS_MEM_SIZE * sizeof(float));
1061 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],
1062 LP_ORDER_16k * sizeof(float));
1065 static int amrwb_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
1068 AMRWBContext *ctx = avctx->priv_data;
1069 AMRWBFrame *cf = &ctx->frame;
1070 const uint8_t *buf = avpkt->data;
1071 int buf_size = avpkt->size;
1072 int expected_fr_size, header_size;
1073 float *buf_out = data;
1074 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing
1075 float fixed_gain_factor; // fixed gain correction factor (gamma)
1076 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
1077 float synth_fixed_gain; // the fixed gain that synthesis should use
1078 float voice_fac, stab_fac; // parameters used for gain smoothing
1079 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis
1080 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band
1081 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis
1085 header_size = decode_mime_header(ctx, buf);
1086 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;
1088 if (buf_size < expected_fr_size) {
1089 av_log(avctx, AV_LOG_ERROR,
1090 "Frame too small (%d bytes). Truncated file?\n", buf_size);
1095 if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)
1096 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");
1098 if (ctx->fr_cur_mode == MODE_SID) /* Comfort noise frame */
1099 av_log_missing_feature(avctx, "SID mode", 1);
1101 if (ctx->fr_cur_mode >= MODE_SID)
1104 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),
1105 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);
1107 /* Decode the quantized ISF vector */
1108 if (ctx->fr_cur_mode == MODE_6k60) {
1109 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur);
1111 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur);
1114 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past);
1115 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1);
1117 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);
1119 ctx->isf_cur[LP_ORDER - 1] *= 2.0;
1120 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);
1122 /* Generate a ISP vector for each subframe */
1123 if (ctx->first_frame) {
1124 ctx->first_frame = 0;
1125 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));
1127 interpolate_isp(ctx->isp, ctx->isp_sub4_past);
1129 for (sub = 0; sub < 4; sub++)
1130 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);
1132 for (sub = 0; sub < 4; sub++) {
1133 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];
1134 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;
1136 /* Decode adaptive codebook (pitch vector) */
1137 decode_pitch_vector(ctx, cur_subframe, sub);
1138 /* Decode innovative codebook (fixed vector) */
1139 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,
1140 cur_subframe->pul_il, ctx->fr_cur_mode);
1142 pitch_sharpening(ctx, ctx->fixed_vector);
1144 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,
1145 &fixed_gain_factor, &ctx->pitch_gain[0]);
1147 ctx->fixed_gain[0] =
1148 ff_amr_set_fixed_gain(fixed_gain_factor,
1149 ff_dot_productf(ctx->fixed_vector, ctx->fixed_vector,
1150 AMRWB_SFR_SIZE) / AMRWB_SFR_SIZE,
1151 ctx->prediction_error,
1152 ENERGY_MEAN, energy_pred_fac);
1154 /* Calculate voice factor and store tilt for next subframe */
1155 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],
1156 ctx->fixed_vector, ctx->fixed_gain[0]);
1157 ctx->tilt_coef = voice_fac * 0.25 + 0.25;
1159 /* Construct current excitation */
1160 for (i = 0; i < AMRWB_SFR_SIZE; i++) {
1161 ctx->excitation[i] *= ctx->pitch_gain[0];
1162 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];
1163 ctx->excitation[i] = truncf(ctx->excitation[i]);
1166 /* Post-processing of excitation elements */
1167 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,
1168 voice_fac, stab_fac);
1170 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,
1173 pitch_enhancer(synth_fixed_vector, voice_fac);
1175 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,
1176 synth_fixed_vector, &ctx->samples_az[LP_ORDER]);
1178 /* Synthesis speech post-processing */
1179 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],
1180 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);
1182 ff_acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],
1183 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,
1184 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);
1186 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],
1187 AMRWB_SFR_SIZE_16k);
1189 /* High frequency band (6.4 - 7.0 kHz) generation part */
1190 ff_acelp_apply_order_2_transfer_function(hb_samples,
1191 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,
1192 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);
1194 hb_gain = find_hb_gain(ctx, hb_samples,
1195 cur_subframe->hb_gain, cf->vad);
1197 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);
1199 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],
1200 hb_exc, ctx->isf_cur, ctx->isf_past_final);
1202 /* High-band post-processing filters */
1203 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,
1204 &ctx->samples_hb[LP_ORDER_16k]);
1206 if (ctx->fr_cur_mode == MODE_23k85)
1207 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,
1210 /* Add the low and high frequency bands */
1211 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)
1212 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));
1214 /* Update buffers and history */
1215 update_sub_state(ctx);
1218 /* update state for next frame */
1219 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));
1220 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));
1222 /* report how many samples we got */
1223 *data_size = 4 * AMRWB_SFR_SIZE_16k * sizeof(float);
1225 return expected_fr_size;
1228 AVCodec ff_amrwb_decoder = {
1230 .type = AVMEDIA_TYPE_AUDIO,
1231 .id = CODEC_ID_AMR_WB,
1232 .priv_data_size = sizeof(AMRWBContext),
1233 .init = amrwb_decode_init,
1234 .decode = amrwb_decode_frame,
1235 .long_name = NULL_IF_CONFIG_SMALL("Adaptive Multi-Rate WideBand"),
1236 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_FLT,AV_SAMPLE_FMT_NONE},