2 * AMR narrowband decoder
3 * Copyright (c) 2006-2007 Robert Swain
4 * Copyright (c) 2009 Colin McQuillan
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 * AMR narrowband decoder
28 * This decoder uses floats for simplicity and so is not bit-exact. One
29 * difference is that differences in phase can accumulate. The test sequences
30 * in 3GPP TS 26.074 can still be useful.
32 * - Comparing this file's output to the output of the ref decoder gives a
33 * PSNR of 30 to 80. Plotting the output samples shows a difference in
34 * phase in some areas.
36 * - Comparing both decoders against their input, this decoder gives a similar
37 * PSNR. If the test sequence homing frames are removed (this decoder does
38 * not detect them), the PSNR is at least as good as the reference on 140
46 #include "libavutil/channel_layout.h"
49 #include "libavutil/common.h"
50 #include "libavutil/avassert.h"
51 #include "celp_math.h"
52 #include "celp_filters.h"
53 #include "acelp_filters.h"
54 #include "acelp_vectors.h"
55 #include "acelp_pitch_delay.h"
59 #include "amrnbdata.h"
61 #define AMR_BLOCK_SIZE 160 ///< samples per frame
62 #define AMR_SAMPLE_BOUND 32768.0 ///< threshold for synthesis overflow
65 * Scale from constructed speech to [-1,1]
67 * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but
68 * upscales by two (section 6.2.2).
70 * Fundamentally, this scale is determined by energy_mean through
71 * the fixed vector contribution to the excitation vector.
73 #define AMR_SAMPLE_SCALE (2.0 / 32768.0)
75 /** Prediction factor for 12.2kbit/s mode */
76 #define PRED_FAC_MODE_12k2 0.65
78 #define LSF_R_FAC (8000.0 / 32768.0) ///< LSF residual tables to Hertz
79 #define MIN_LSF_SPACING (50.0488 / 8000.0) ///< Ensures stability of LPC filter
80 #define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode
82 /** Initial energy in dB. Also used for bad frames (unimplemented). */
83 #define MIN_ENERGY -14.0
85 /** Maximum sharpening factor
87 * The specification says 0.8, which should be 13107, but the reference C code
88 * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in g729dec.c.)
90 #define SHARP_MAX 0.79449462890625
92 /** Number of impulse response coefficients used for tilt factor */
93 #define AMR_TILT_RESPONSE 22
94 /** Tilt factor = 1st reflection coefficient * gamma_t */
95 #define AMR_TILT_GAMMA_T 0.8
96 /** Adaptive gain control factor used in post-filter */
97 #define AMR_AGC_ALPHA 0.9
99 typedef struct AMRContext {
100 AVFrame avframe; ///< AVFrame for decoded samples
101 AMRNBFrame frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc)
102 uint8_t bad_frame_indicator; ///< bad frame ? 1 : 0
103 enum Mode cur_frame_mode;
105 int16_t prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe
106 double lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame
107 double prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame
109 float lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing
110 float lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector
112 float lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes
114 uint8_t pitch_lag_int; ///< integer part of pitch lag from current subframe
116 float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history
117 float *excitation; ///< pointer to the current excitation vector in excitation_buf
119 float pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector
120 float fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames)
122 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
123 float pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes
124 float fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes
126 float beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX]
127 uint8_t diff_count; ///< the number of subframes for which diff has been above 0.65
128 uint8_t hang_count; ///< the number of subframes since a hangover period started
130 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset"
131 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
132 uint8_t ir_filter_onset; ///< flag for impulse response filter strength
134 float postfilter_mem[10]; ///< previous intermediate values in the formant filter
135 float tilt_mem; ///< previous input to tilt compensation filter
136 float postfilter_agc; ///< previous factor used for adaptive gain control
137 float high_pass_mem[2]; ///< previous intermediate values in the high-pass filter
139 float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples
141 ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs
142 ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs
143 CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs
144 CELPMContext celpm_ctx; ///< context for fixed point math operations
148 /** Double version of ff_weighted_vector_sumf() */
149 static void weighted_vector_sumd(double *out, const double *in_a,
150 const double *in_b, double weight_coeff_a,
151 double weight_coeff_b, int length)
155 for (i = 0; i < length; i++)
156 out[i] = weight_coeff_a * in_a[i]
157 + weight_coeff_b * in_b[i];
160 static av_cold int amrnb_decode_init(AVCodecContext *avctx)
162 AMRContext *p = avctx->priv_data;
165 if (avctx->channels > 1) {
166 av_log_missing_feature(avctx, "multi-channel AMR", 0);
167 return AVERROR_PATCHWELCOME;
171 avctx->channel_layout = AV_CH_LAYOUT_MONO;
172 if (!avctx->sample_rate)
173 avctx->sample_rate = 8000;
174 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
176 // p->excitation always points to the same position in p->excitation_buf
177 p->excitation = &p->excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1];
179 for (i = 0; i < LP_FILTER_ORDER; i++) {
180 p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15);
181 p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15);
184 for (i = 0; i < 4; i++)
185 p->prediction_error[i] = MIN_ENERGY;
187 avcodec_get_frame_defaults(&p->avframe);
188 avctx->coded_frame = &p->avframe;
190 ff_acelp_filter_init(&p->acelpf_ctx);
191 ff_acelp_vectors_init(&p->acelpv_ctx);
192 ff_celp_filter_init(&p->celpf_ctx);
193 ff_celp_math_init(&p->celpm_ctx);
200 * Unpack an RFC4867 speech frame into the AMR frame mode and parameters.
202 * The order of speech bits is specified by 3GPP TS 26.101.
204 * @param p the context
205 * @param buf pointer to the input buffer
206 * @param buf_size size of the input buffer
208 * @return the frame mode
210 static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf,
215 // Decode the first octet.
216 mode = buf[0] >> 3 & 0x0F; // frame type
217 p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit
219 if (mode >= N_MODES || buf_size < frame_sizes_nb[mode] + 1) {
224 ff_amr_bit_reorder((uint16_t *) &p->frame, sizeof(AMRNBFrame), buf + 1,
225 amr_unpacking_bitmaps_per_mode[mode]);
231 /// @name AMR pitch LPC coefficient decoding functions
235 * Interpolate the LSF vector (used for fixed gain smoothing).
236 * The interpolation is done over all four subframes even in MODE_12k2.
238 * @param[in] ctx The Context
239 * @param[in,out] lsf_q LSFs in [0,1] for each subframe
240 * @param[in] lsf_new New LSFs in [0,1] for subframe 4
242 static void interpolate_lsf(ACELPVContext *ctx, float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
246 for (i = 0; i < 4; i++)
247 ctx->weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new,
248 0.25 * (3 - i), 0.25 * (i + 1),
253 * Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector.
255 * @param p the context
256 * @param lsp output LSP vector
257 * @param lsf_no_r LSF vector without the residual vector added
258 * @param lsf_quantizer pointers to LSF dictionary tables
259 * @param quantizer_offset offset in tables
260 * @param sign for the 3 dictionary table
261 * @param update store data for computing the next frame's LSFs
263 static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER],
264 const float lsf_no_r[LP_FILTER_ORDER],
265 const int16_t *lsf_quantizer[5],
266 const int quantizer_offset,
267 const int sign, const int update)
269 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
270 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
273 for (i = 0; i < LP_FILTER_ORDER >> 1; i++)
274 memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset],
283 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
285 for (i = 0; i < LP_FILTER_ORDER; i++)
286 lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0);
288 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
291 interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
293 ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER);
297 * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors.
299 * @param p pointer to the AMRContext
301 static void lsf2lsp_5(AMRContext *p)
303 const uint16_t *lsf_param = p->frame.lsf;
304 float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector
305 const int16_t *lsf_quantizer[5];
308 lsf_quantizer[0] = lsf_5_1[lsf_param[0]];
309 lsf_quantizer[1] = lsf_5_2[lsf_param[1]];
310 lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1];
311 lsf_quantizer[3] = lsf_5_4[lsf_param[3]];
312 lsf_quantizer[4] = lsf_5_5[lsf_param[4]];
314 for (i = 0; i < LP_FILTER_ORDER; i++)
315 lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i];
317 lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0);
318 lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1);
320 // interpolate LSP vectors at subframes 1 and 3
321 weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER);
322 weighted_vector_sumd(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER);
326 * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector.
328 * @param p pointer to the AMRContext
330 static void lsf2lsp_3(AMRContext *p)
332 const uint16_t *lsf_param = p->frame.lsf;
333 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
334 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
335 const int16_t *lsf_quantizer;
338 lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]];
339 memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r));
341 lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)];
342 memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r));
344 lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]];
345 memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r));
347 // calculate mean-removed LSF vector and add mean
348 for (i = 0; i < LP_FILTER_ORDER; i++)
349 lsf_q[i] = (lsf_r[i] + p->prev_lsf_r[i] * pred_fac[i]) * (LSF_R_FAC / 8000.0) + lsf_3_mean[i] * (1.0 / 8000.0);
351 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
353 // store data for computing the next frame's LSFs
354 interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
355 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
357 ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
359 // interpolate LSP vectors at subframes 1, 2 and 3
360 for (i = 1; i <= 3; i++)
361 for(j = 0; j < LP_FILTER_ORDER; j++)
362 p->lsp[i-1][j] = p->prev_lsp_sub4[j] +
363 (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i;
369 /// @name AMR pitch vector decoding functions
373 * Like ff_decode_pitch_lag(), but with 1/6 resolution
375 static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index,
376 const int prev_lag_int, const int subframe)
378 if (subframe == 0 || subframe == 2) {
379 if (pitch_index < 463) {
380 *lag_int = (pitch_index + 107) * 10923 >> 16;
381 *lag_frac = pitch_index - *lag_int * 6 + 105;
383 *lag_int = pitch_index - 368;
387 *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1;
388 *lag_frac = pitch_index - *lag_int * 6 - 3;
389 *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2,
390 PITCH_DELAY_MAX - 9);
394 static void decode_pitch_vector(AMRContext *p,
395 const AMRNBSubframe *amr_subframe,
398 int pitch_lag_int, pitch_lag_frac;
399 enum Mode mode = p->cur_frame_mode;
401 if (p->cur_frame_mode == MODE_12k2) {
402 decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac,
403 amr_subframe->p_lag, p->pitch_lag_int,
406 ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac,
408 p->pitch_lag_int, subframe,
409 mode != MODE_4k75 && mode != MODE_5k15,
410 mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6));
412 p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t
414 pitch_lag_frac <<= (p->cur_frame_mode != MODE_12k2);
416 pitch_lag_int += pitch_lag_frac > 0;
418 /* Calculate the pitch vector by interpolating the past excitation at the
419 pitch lag using a b60 hamming windowed sinc function. */
420 p->acelpf_ctx.acelp_interpolatef(p->excitation,
421 p->excitation + 1 - pitch_lag_int,
423 pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0),
424 10, AMR_SUBFRAME_SIZE);
426 memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float));
432 /// @name AMR algebraic code book (fixed) vector decoding functions
436 * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame.
438 static void decode_10bit_pulse(int code, int pulse_position[8],
439 int i1, int i2, int i3)
441 // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of
442 // the 3 pulses and the upper 7 bits being coded in base 5
443 const uint8_t *positions = base_five_table[code >> 3];
444 pulse_position[i1] = (positions[2] << 1) + ( code & 1);
445 pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1);
446 pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1);
450 * Decode the algebraic codebook index to pulse positions and signs and
451 * construct the algebraic codebook vector for MODE_10k2.
453 * @param fixed_index positions of the eight pulses
454 * @param fixed_sparse pointer to the algebraic codebook vector
456 static void decode_8_pulses_31bits(const int16_t *fixed_index,
457 AMRFixed *fixed_sparse)
459 int pulse_position[8];
462 decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1);
463 decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5);
465 // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of
466 // the 2 pulses and the upper 5 bits being coded in base 5
467 temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5;
468 pulse_position[3] = temp % 5;
469 pulse_position[7] = temp / 5;
470 if (pulse_position[7] & 1)
471 pulse_position[3] = 4 - pulse_position[3];
472 pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1);
473 pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1);
476 for (i = 0; i < 4; i++) {
477 const int pos1 = (pulse_position[i] << 2) + i;
478 const int pos2 = (pulse_position[i + 4] << 2) + i;
479 const float sign = fixed_index[i] ? -1.0 : 1.0;
480 fixed_sparse->x[i ] = pos1;
481 fixed_sparse->x[i + 4] = pos2;
482 fixed_sparse->y[i ] = sign;
483 fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign;
488 * Decode the algebraic codebook index to pulse positions and signs,
489 * then construct the algebraic codebook vector.
491 * nb of pulses | bits encoding pulses
492 * For MODE_4k75 or MODE_5k15, 2 | 1-3, 4-6, 7
493 * MODE_5k9, 2 | 1, 2-4, 5-6, 7-9
494 * MODE_6k7, 3 | 1-3, 4, 5-7, 8, 9-11
495 * MODE_7k4 or MODE_7k95, 4 | 1-3, 4-6, 7-9, 10, 11-13
497 * @param fixed_sparse pointer to the algebraic codebook vector
498 * @param pulses algebraic codebook indexes
499 * @param mode mode of the current frame
500 * @param subframe current subframe number
502 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses,
503 const enum Mode mode, const int subframe)
505 av_assert1(MODE_4k75 <= (signed)mode && mode <= MODE_12k2);
507 if (mode == MODE_12k2) {
508 ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3);
509 } else if (mode == MODE_10k2) {
510 decode_8_pulses_31bits(pulses, fixed_sparse);
512 int *pulse_position = fixed_sparse->x;
514 const int fixed_index = pulses[0];
516 if (mode <= MODE_5k15) {
517 pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1);
518 pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset];
519 pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1];
521 } else if (mode == MODE_5k9) {
522 pulse_subset = ((fixed_index & 1) << 1) + 1;
523 pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset;
524 pulse_subset = (fixed_index >> 4) & 3;
525 pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0);
526 fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2;
527 } else if (mode == MODE_6k7) {
528 pulse_position[0] = (fixed_index & 7) * 5;
529 pulse_subset = (fixed_index >> 2) & 2;
530 pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1;
531 pulse_subset = (fixed_index >> 6) & 2;
532 pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2;
534 } else { // mode <= MODE_7k95
535 pulse_position[0] = gray_decode[ fixed_index & 7];
536 pulse_position[1] = gray_decode[(fixed_index >> 3) & 7] + 1;
537 pulse_position[2] = gray_decode[(fixed_index >> 6) & 7] + 2;
538 pulse_subset = (fixed_index >> 9) & 1;
539 pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3;
542 for (i = 0; i < fixed_sparse->n; i++)
543 fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0;
548 * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2)
550 * @param p the context
551 * @param subframe unpacked amr subframe
552 * @param mode mode of the current frame
553 * @param fixed_sparse sparse respresentation of the fixed vector
555 static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode,
556 AMRFixed *fixed_sparse)
558 // The spec suggests the current pitch gain is always used, but in other
559 // modes the pitch and codebook gains are joinly quantized (sec 5.8.2)
560 // so the codebook gain cannot depend on the quantized pitch gain.
561 if (mode == MODE_12k2)
562 p->beta = FFMIN(p->pitch_gain[4], 1.0);
564 fixed_sparse->pitch_lag = p->pitch_lag_int;
565 fixed_sparse->pitch_fac = p->beta;
567 // Save pitch sharpening factor for the next subframe
568 // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from
569 // the fact that the gains for two subframes are jointly quantized.
570 if (mode != MODE_4k75 || subframe & 1)
571 p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX);
576 /// @name AMR gain decoding functions
580 * fixed gain smoothing
581 * Note that where the spec specifies the "spectrum in the q domain"
582 * in section 6.1.4, in fact frequencies should be used.
584 * @param p the context
585 * @param lsf LSFs for the current subframe, in the range [0,1]
586 * @param lsf_avg averaged LSFs
587 * @param mode mode of the current frame
589 * @return fixed gain smoothed
591 static float fixed_gain_smooth(AMRContext *p , const float *lsf,
592 const float *lsf_avg, const enum Mode mode)
597 for (i = 0; i < LP_FILTER_ORDER; i++)
598 diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i];
600 // If diff is large for ten subframes, disable smoothing for a 40-subframe
606 if (p->diff_count > 10) {
608 p->diff_count--; // don't let diff_count overflow
611 if (p->hang_count < 40) {
613 } else if (mode < MODE_7k4 || mode == MODE_10k2) {
614 const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0);
615 const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] +
616 p->fixed_gain[2] + p->fixed_gain[3] +
617 p->fixed_gain[4]) * 0.2;
618 return smoothing_factor * p->fixed_gain[4] +
619 (1.0 - smoothing_factor) * fixed_gain_mean;
621 return p->fixed_gain[4];
625 * Decode pitch gain and fixed gain factor (part of section 6.1.3).
627 * @param p the context
628 * @param amr_subframe unpacked amr subframe
629 * @param mode mode of the current frame
630 * @param subframe current subframe number
631 * @param fixed_gain_factor decoded gain correction factor
633 static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe,
634 const enum Mode mode, const int subframe,
635 float *fixed_gain_factor)
637 if (mode == MODE_12k2 || mode == MODE_7k95) {
638 p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ]
640 *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain]
643 const uint16_t *gains;
645 if (mode >= MODE_6k7) {
646 gains = gains_high[amr_subframe->p_gain];
647 } else if (mode >= MODE_5k15) {
648 gains = gains_low [amr_subframe->p_gain];
650 // gain index is only coded in subframes 0,2 for MODE_4k75
651 gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)];
654 p->pitch_gain[4] = gains[0] * (1.0 / 16384.0);
655 *fixed_gain_factor = gains[1] * (1.0 / 4096.0);
662 /// @name AMR preprocessing functions
666 * Circularly convolve a sparse fixed vector with a phase dispersion impulse
667 * response filter (D.6.2 of G.729 and 6.1.5 of AMR).
669 * @param out vector with filter applied
670 * @param in source vector
671 * @param filter phase filter coefficients
673 * out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] }
675 static void apply_ir_filter(float *out, const AMRFixed *in,
678 float filter1[AMR_SUBFRAME_SIZE], ///< filters at pitch lag*1 and *2
679 filter2[AMR_SUBFRAME_SIZE];
680 int lag = in->pitch_lag;
681 float fac = in->pitch_fac;
684 if (lag < AMR_SUBFRAME_SIZE) {
685 ff_celp_circ_addf(filter1, filter, filter, lag, fac,
688 if (lag < AMR_SUBFRAME_SIZE >> 1)
689 ff_celp_circ_addf(filter2, filter, filter1, lag, fac,
693 memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE);
694 for (i = 0; i < in->n; i++) {
697 const float *filterp;
699 if (x >= AMR_SUBFRAME_SIZE - lag) {
701 } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) {
706 ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE);
711 * Reduce fixed vector sparseness by smoothing with one of three IR filters.
712 * Also know as "adaptive phase dispersion".
714 * This implements 3GPP TS 26.090 section 6.1(5).
716 * @param p the context
717 * @param fixed_sparse algebraic codebook vector
718 * @param fixed_vector unfiltered fixed vector
719 * @param fixed_gain smoothed gain
720 * @param out space for modified vector if necessary
722 static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse,
723 const float *fixed_vector,
724 float fixed_gain, float *out)
728 if (p->pitch_gain[4] < 0.6) {
729 ir_filter_nr = 0; // strong filtering
730 } else if (p->pitch_gain[4] < 0.9) {
731 ir_filter_nr = 1; // medium filtering
733 ir_filter_nr = 2; // no filtering
736 if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) {
737 p->ir_filter_onset = 2;
738 } else if (p->ir_filter_onset)
739 p->ir_filter_onset--;
741 if (!p->ir_filter_onset) {
744 for (i = 0; i < 5; i++)
745 if (p->pitch_gain[i] < 0.6)
750 if (ir_filter_nr > p->prev_ir_filter_nr + 1)
752 } else if (ir_filter_nr < 2)
755 // Disable filtering for very low level of fixed_gain.
756 // Note this step is not specified in the technical description but is in
757 // the reference source in the function Ph_disp.
758 if (fixed_gain < 5.0)
761 if (p->cur_frame_mode != MODE_7k4 && p->cur_frame_mode < MODE_10k2
762 && ir_filter_nr < 2) {
763 apply_ir_filter(out, fixed_sparse,
764 (p->cur_frame_mode == MODE_7k95 ?
765 ir_filters_lookup_MODE_7k95 :
766 ir_filters_lookup)[ir_filter_nr]);
770 // update ir filter strength history
771 p->prev_ir_filter_nr = ir_filter_nr;
772 p->prev_sparse_fixed_gain = fixed_gain;
780 /// @name AMR synthesis functions
784 * Conduct 10th order linear predictive coding synthesis.
786 * @param p pointer to the AMRContext
787 * @param lpc pointer to the LPC coefficients
788 * @param fixed_gain fixed codebook gain for synthesis
789 * @param fixed_vector algebraic codebook vector
790 * @param samples pointer to the output speech samples
791 * @param overflow 16-bit overflow flag
793 static int synthesis(AMRContext *p, float *lpc,
794 float fixed_gain, const float *fixed_vector,
795 float *samples, uint8_t overflow)
798 float excitation[AMR_SUBFRAME_SIZE];
800 // if an overflow has been detected, the pitch vector is scaled down by a
803 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
804 p->pitch_vector[i] *= 0.25;
806 p->acelpv_ctx.weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector,
807 p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE);
809 // emphasize pitch vector contribution
810 if (p->pitch_gain[4] > 0.5 && !overflow) {
811 float energy = p->celpm_ctx.dot_productf(excitation, excitation,
815 (p->cur_frame_mode == MODE_12k2 ?
816 0.25 * FFMIN(p->pitch_gain[4], 1.0) :
817 0.5 * FFMIN(p->pitch_gain[4], SHARP_MAX));
819 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
820 excitation[i] += pitch_factor * p->pitch_vector[i];
822 ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy,
826 p->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
831 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
832 if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) {
842 /// @name AMR update functions
846 * Update buffers and history at the end of decoding a subframe.
848 * @param p pointer to the AMRContext
850 static void update_state(AMRContext *p)
852 memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0]));
854 memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE],
855 (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float));
857 memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float));
858 memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float));
860 memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE],
861 LP_FILTER_ORDER * sizeof(float));
867 /// @name AMR Postprocessing functions
871 * Get the tilt factor of a formant filter from its transfer function
873 * @param p The Context
874 * @param lpc_n LP_FILTER_ORDER coefficients of the numerator
875 * @param lpc_d LP_FILTER_ORDER coefficients of the denominator
877 static float tilt_factor(AMRContext *p, float *lpc_n, float *lpc_d)
879 float rh0, rh1; // autocorrelation at lag 0 and 1
881 // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf
882 float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 };
883 float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response
886 memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER);
887 p->celpf_ctx.celp_lp_synthesis_filterf(hf, lpc_d, hf,
891 rh0 = p->celpm_ctx.dot_productf(hf, hf, AMR_TILT_RESPONSE);
892 rh1 = p->celpm_ctx.dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1);
894 // The spec only specifies this check for 12.2 and 10.2 kbit/s
895 // modes. But in the ref source the tilt is always non-negative.
896 return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0;
900 * Perform adaptive post-filtering to enhance the quality of the speech.
903 * @param p pointer to the AMRContext
904 * @param lpc interpolated LP coefficients for this subframe
905 * @param buf_out output of the filter
907 static void postfilter(AMRContext *p, float *lpc, float *buf_out)
910 float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input
912 float speech_gain = p->celpm_ctx.dot_productf(samples, samples,
915 float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter
916 const float *gamma_n, *gamma_d; // Formant filter factor table
917 float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients
919 if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) {
920 gamma_n = ff_pow_0_7;
921 gamma_d = ff_pow_0_75;
923 gamma_n = ff_pow_0_55;
924 gamma_d = ff_pow_0_7;
927 for (i = 0; i < LP_FILTER_ORDER; i++) {
928 lpc_n[i] = lpc[i] * gamma_n[i];
929 lpc_d[i] = lpc[i] * gamma_d[i];
932 memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER);
933 p->celpf_ctx.celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples,
934 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
935 memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE,
936 sizeof(float) * LP_FILTER_ORDER);
938 p->celpf_ctx.celp_lp_zero_synthesis_filterf(buf_out, lpc_n,
939 pole_out + LP_FILTER_ORDER,
940 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
942 ff_tilt_compensation(&p->tilt_mem, tilt_factor(p, lpc_n, lpc_d), buf_out,
945 ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE,
946 AMR_AGC_ALPHA, &p->postfilter_agc);
951 static int amrnb_decode_frame(AVCodecContext *avctx, void *data,
952 int *got_frame_ptr, AVPacket *avpkt)
955 AMRContext *p = avctx->priv_data; // pointer to private data
956 const uint8_t *buf = avpkt->data;
957 int buf_size = avpkt->size;
958 float *buf_out; // pointer to the output data buffer
959 int i, subframe, ret;
960 float fixed_gain_factor;
961 AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing
962 float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing
963 float synth_fixed_gain; // the fixed gain that synthesis should use
964 const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
966 /* get output buffer */
967 p->avframe.nb_samples = AMR_BLOCK_SIZE;
968 if ((ret = avctx->get_buffer(avctx, &p->avframe)) < 0) {
969 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
972 buf_out = (float *)p->avframe.data[0];
974 p->cur_frame_mode = unpack_bitstream(p, buf, buf_size);
975 if (p->cur_frame_mode == NO_DATA) {
976 av_log(avctx, AV_LOG_ERROR, "Corrupt bitstream\n");
977 return AVERROR_INVALIDDATA;
979 if (p->cur_frame_mode == MODE_DTX) {
980 av_log_missing_feature(avctx, "dtx mode", 0);
981 av_log(avctx, AV_LOG_INFO, "Note: libopencore_amrnb supports dtx\n");
982 return AVERROR_PATCHWELCOME;
985 if (p->cur_frame_mode == MODE_12k2) {
990 for (i = 0; i < 4; i++)
991 ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5);
993 for (subframe = 0; subframe < 4; subframe++) {
994 const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe];
996 decode_pitch_vector(p, amr_subframe, subframe);
998 decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses,
999 p->cur_frame_mode, subframe);
1001 // The fixed gain (section 6.1.3) depends on the fixed vector
1002 // (section 6.1.2), but the fixed vector calculation uses
1003 // pitch sharpening based on the on the pitch gain (section 6.1.3).
1004 // So the correct order is: pitch gain, pitch sharpening, fixed gain.
1005 decode_gains(p, amr_subframe, p->cur_frame_mode, subframe,
1006 &fixed_gain_factor);
1008 pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse);
1010 if (fixed_sparse.pitch_lag == 0) {
1011 av_log(avctx, AV_LOG_ERROR, "The file is corrupted, pitch_lag = 0 is not allowed\n");
1012 return AVERROR_INVALIDDATA;
1014 ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0,
1018 ff_amr_set_fixed_gain(fixed_gain_factor,
1019 p->celpm_ctx.dot_productf(p->fixed_vector,
1021 AMR_SUBFRAME_SIZE) /
1023 p->prediction_error,
1024 energy_mean[p->cur_frame_mode], energy_pred_fac);
1026 // The excitation feedback is calculated without any processing such
1027 // as fixed gain smoothing. This isn't mentioned in the specification.
1028 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1029 p->excitation[i] *= p->pitch_gain[4];
1030 ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4],
1033 // In the ref decoder, excitation is stored with no fractional bits.
1034 // This step prevents buzz in silent periods. The ref encoder can
1035 // emit long sequences with pitch factor greater than one. This
1036 // creates unwanted feedback if the excitation vector is nonzero.
1037 // (e.g. test sequence T19_795.COD in 3GPP TS 26.074)
1038 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1039 p->excitation[i] = truncf(p->excitation[i]);
1041 // Smooth fixed gain.
1042 // The specification is ambiguous, but in the reference source, the
1043 // smoothed value is NOT fed back into later fixed gain smoothing.
1044 synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe],
1045 p->lsf_avg, p->cur_frame_mode);
1047 synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector,
1048 synth_fixed_gain, spare_vector);
1050 if (synthesis(p, p->lpc[subframe], synth_fixed_gain,
1051 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0))
1052 // overflow detected -> rerun synthesis scaling pitch vector down
1053 // by a factor of 4, skipping pitch vector contribution emphasis
1054 // and adaptive gain control
1055 synthesis(p, p->lpc[subframe], synth_fixed_gain,
1056 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1);
1058 postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE);
1060 // update buffers and history
1061 ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE);
1065 p->acelpf_ctx.acelp_apply_order_2_transfer_function(buf_out,
1066 buf_out, highpass_zeros,
1068 highpass_gain * AMR_SAMPLE_SCALE,
1069 p->high_pass_mem, AMR_BLOCK_SIZE);
1071 /* Update averaged lsf vector (used for fixed gain smoothing).
1073 * Note that lsf_avg should not incorporate the current frame's LSFs
1074 * for fixed_gain_smooth.
1075 * The specification has an incorrect formula: the reference decoder uses
1076 * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */
1077 p->acelpv_ctx.weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3],
1078 0.84, 0.16, LP_FILTER_ORDER);
1081 *(AVFrame *)data = p->avframe;
1083 /* return the amount of bytes consumed if everything was OK */
1084 return frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC
1088 AVCodec ff_amrnb_decoder = {
1090 .type = AVMEDIA_TYPE_AUDIO,
1091 .id = AV_CODEC_ID_AMR_NB,
1092 .priv_data_size = sizeof(AMRContext),
1093 .init = amrnb_decode_init,
1094 .decode = amrnb_decode_frame,
1095 .capabilities = CODEC_CAP_DR1,
1096 .long_name = NULL_IF_CONFIG_SMALL("AMR-NB (Adaptive Multi-Rate NarrowBand)"),
1097 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1098 AV_SAMPLE_FMT_NONE },