2 * AMR narrowband decoder
3 * Copyright (c) 2006-2007 Robert Swain
4 * Copyright (c) 2009 Colin McQuillan
6 * This file is part of Libav.
8 * Libav 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 * Libav 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 Libav; 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 "celp_filters.h"
51 #include "acelp_filters.h"
52 #include "acelp_vectors.h"
53 #include "acelp_pitch_delay.h"
58 #include "amrnbdata.h"
60 #define AMR_BLOCK_SIZE 160 ///< samples per frame
61 #define AMR_SAMPLE_BOUND 32768.0 ///< threshold for synthesis overflow
64 * Scale from constructed speech to [-1,1]
66 * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but
67 * upscales by two (section 6.2.2).
69 * Fundamentally, this scale is determined by energy_mean through
70 * the fixed vector contribution to the excitation vector.
72 #define AMR_SAMPLE_SCALE (2.0 / 32768.0)
74 /** Prediction factor for 12.2kbit/s mode */
75 #define PRED_FAC_MODE_12k2 0.65
77 #define LSF_R_FAC (8000.0 / 32768.0) ///< LSF residual tables to Hertz
78 #define MIN_LSF_SPACING (50.0488 / 8000.0) ///< Ensures stability of LPC filter
79 #define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode
81 /** Initial energy in dB. Also used for bad frames (unimplemented). */
82 #define MIN_ENERGY -14.0
84 /** Maximum sharpening factor
86 * The specification says 0.8, which should be 13107, but the reference C code
87 * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in bitexact G.729.)
89 #define SHARP_MAX 0.79449462890625
91 /** Number of impulse response coefficients used for tilt factor */
92 #define AMR_TILT_RESPONSE 22
93 /** Tilt factor = 1st reflection coefficient * gamma_t */
94 #define AMR_TILT_GAMMA_T 0.8
95 /** Adaptive gain control factor used in post-filter */
96 #define AMR_AGC_ALPHA 0.9
98 typedef struct AMRContext {
99 AVFrame avframe; ///< AVFrame for decoded samples
100 AMRNBFrame frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc)
101 uint8_t bad_frame_indicator; ///< bad frame ? 1 : 0
102 enum Mode cur_frame_mode;
104 int16_t prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe
105 double lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame
106 double prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame
108 float lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing
109 float lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector
111 float lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes
113 uint8_t pitch_lag_int; ///< integer part of pitch lag from current subframe
115 float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history
116 float *excitation; ///< pointer to the current excitation vector in excitation_buf
118 float pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector
119 float fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames)
121 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
122 float pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes
123 float fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes
125 float beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX]
126 uint8_t diff_count; ///< the number of subframes for which diff has been above 0.65
127 uint8_t hang_count; ///< the number of subframes since a hangover period started
129 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset"
130 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
131 uint8_t ir_filter_onset; ///< flag for impulse response filter strength
133 float postfilter_mem[10]; ///< previous intermediate values in the formant filter
134 float tilt_mem; ///< previous input to tilt compensation filter
135 float postfilter_agc; ///< previous factor used for adaptive gain control
136 float high_pass_mem[2]; ///< previous intermediate values in the high-pass filter
138 float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples
142 /** Double version of ff_weighted_vector_sumf() */
143 static void weighted_vector_sumd(double *out, const double *in_a,
144 const double *in_b, double weight_coeff_a,
145 double weight_coeff_b, int length)
149 for (i = 0; i < length; i++)
150 out[i] = weight_coeff_a * in_a[i]
151 + weight_coeff_b * in_b[i];
154 static av_cold int amrnb_decode_init(AVCodecContext *avctx)
156 AMRContext *p = avctx->priv_data;
159 if (avctx->channels > 1) {
160 av_log_missing_feature(avctx, "multi-channel AMR", 0);
161 return AVERROR_PATCHWELCOME;
165 avctx->channel_layout = AV_CH_LAYOUT_MONO;
166 avctx->sample_rate = 8000;
167 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
169 // p->excitation always points to the same position in p->excitation_buf
170 p->excitation = &p->excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1];
172 for (i = 0; i < LP_FILTER_ORDER; i++) {
173 p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15);
174 p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15);
177 for (i = 0; i < 4; i++)
178 p->prediction_error[i] = MIN_ENERGY;
180 avcodec_get_frame_defaults(&p->avframe);
181 avctx->coded_frame = &p->avframe;
188 * Unpack an RFC4867 speech frame into the AMR frame mode and parameters.
190 * The order of speech bits is specified by 3GPP TS 26.101.
192 * @param p the context
193 * @param buf pointer to the input buffer
194 * @param buf_size size of the input buffer
196 * @return the frame mode
198 static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf,
203 // Decode the first octet.
204 mode = buf[0] >> 3 & 0x0F; // frame type
205 p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit
207 if (mode >= N_MODES || buf_size < frame_sizes_nb[mode] + 1) {
212 ff_amr_bit_reorder((uint16_t *) &p->frame, sizeof(AMRNBFrame), buf + 1,
213 amr_unpacking_bitmaps_per_mode[mode]);
219 /// @name AMR pitch LPC coefficient decoding functions
223 * Interpolate the LSF vector (used for fixed gain smoothing).
224 * The interpolation is done over all four subframes even in MODE_12k2.
226 * @param[in,out] lsf_q LSFs in [0,1] for each subframe
227 * @param[in] lsf_new New LSFs in [0,1] for subframe 4
229 static void interpolate_lsf(float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
233 for (i = 0; i < 4; i++)
234 ff_weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new,
235 0.25 * (3 - i), 0.25 * (i + 1),
240 * Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector.
242 * @param p the context
243 * @param lsp output LSP vector
244 * @param lsf_no_r LSF vector without the residual vector added
245 * @param lsf_quantizer pointers to LSF dictionary tables
246 * @param quantizer_offset offset in tables
247 * @param sign for the 3 dictionary table
248 * @param update store data for computing the next frame's LSFs
250 static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER],
251 const float lsf_no_r[LP_FILTER_ORDER],
252 const int16_t *lsf_quantizer[5],
253 const int quantizer_offset,
254 const int sign, const int update)
256 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
257 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
260 for (i = 0; i < LP_FILTER_ORDER >> 1; i++)
261 memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset],
270 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
272 for (i = 0; i < LP_FILTER_ORDER; i++)
273 lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0);
275 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
278 interpolate_lsf(p->lsf_q, lsf_q);
280 ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER);
284 * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors.
286 * @param p pointer to the AMRContext
288 static void lsf2lsp_5(AMRContext *p)
290 const uint16_t *lsf_param = p->frame.lsf;
291 float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector
292 const int16_t *lsf_quantizer[5];
295 lsf_quantizer[0] = lsf_5_1[lsf_param[0]];
296 lsf_quantizer[1] = lsf_5_2[lsf_param[1]];
297 lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1];
298 lsf_quantizer[3] = lsf_5_4[lsf_param[3]];
299 lsf_quantizer[4] = lsf_5_5[lsf_param[4]];
301 for (i = 0; i < LP_FILTER_ORDER; i++)
302 lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i];
304 lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0);
305 lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1);
307 // interpolate LSP vectors at subframes 1 and 3
308 weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER);
309 weighted_vector_sumd(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER);
313 * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector.
315 * @param p pointer to the AMRContext
317 static void lsf2lsp_3(AMRContext *p)
319 const uint16_t *lsf_param = p->frame.lsf;
320 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
321 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
322 const int16_t *lsf_quantizer;
325 lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]];
326 memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r));
328 lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)];
329 memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r));
331 lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]];
332 memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r));
334 // calculate mean-removed LSF vector and add mean
335 for (i = 0; i < LP_FILTER_ORDER; i++)
336 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);
338 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
340 // store data for computing the next frame's LSFs
341 interpolate_lsf(p->lsf_q, lsf_q);
342 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
344 ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
346 // interpolate LSP vectors at subframes 1, 2 and 3
347 for (i = 1; i <= 3; i++)
348 for(j = 0; j < LP_FILTER_ORDER; j++)
349 p->lsp[i-1][j] = p->prev_lsp_sub4[j] +
350 (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i;
356 /// @name AMR pitch vector decoding functions
360 * Like ff_decode_pitch_lag(), but with 1/6 resolution
362 static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index,
363 const int prev_lag_int, const int subframe)
365 if (subframe == 0 || subframe == 2) {
366 if (pitch_index < 463) {
367 *lag_int = (pitch_index + 107) * 10923 >> 16;
368 *lag_frac = pitch_index - *lag_int * 6 + 105;
370 *lag_int = pitch_index - 368;
374 *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1;
375 *lag_frac = pitch_index - *lag_int * 6 - 3;
376 *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2,
377 PITCH_DELAY_MAX - 9);
381 static void decode_pitch_vector(AMRContext *p,
382 const AMRNBSubframe *amr_subframe,
385 int pitch_lag_int, pitch_lag_frac;
386 enum Mode mode = p->cur_frame_mode;
388 if (p->cur_frame_mode == MODE_12k2) {
389 decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac,
390 amr_subframe->p_lag, p->pitch_lag_int,
393 ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac,
395 p->pitch_lag_int, subframe,
396 mode != MODE_4k75 && mode != MODE_5k15,
397 mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6));
399 p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t
401 pitch_lag_frac <<= (p->cur_frame_mode != MODE_12k2);
403 pitch_lag_int += pitch_lag_frac > 0;
405 /* Calculate the pitch vector by interpolating the past excitation at the
406 pitch lag using a b60 hamming windowed sinc function. */
407 ff_acelp_interpolatef(p->excitation, p->excitation + 1 - pitch_lag_int,
409 pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0),
410 10, AMR_SUBFRAME_SIZE);
412 memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float));
418 /// @name AMR algebraic code book (fixed) vector decoding functions
422 * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame.
424 static void decode_10bit_pulse(int code, int pulse_position[8],
425 int i1, int i2, int i3)
427 // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of
428 // the 3 pulses and the upper 7 bits being coded in base 5
429 const uint8_t *positions = base_five_table[code >> 3];
430 pulse_position[i1] = (positions[2] << 1) + ( code & 1);
431 pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1);
432 pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1);
436 * Decode the algebraic codebook index to pulse positions and signs and
437 * construct the algebraic codebook vector for MODE_10k2.
439 * @param fixed_index positions of the eight pulses
440 * @param fixed_sparse pointer to the algebraic codebook vector
442 static void decode_8_pulses_31bits(const int16_t *fixed_index,
443 AMRFixed *fixed_sparse)
445 int pulse_position[8];
448 decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1);
449 decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5);
451 // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of
452 // the 2 pulses and the upper 5 bits being coded in base 5
453 temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5;
454 pulse_position[3] = temp % 5;
455 pulse_position[7] = temp / 5;
456 if (pulse_position[7] & 1)
457 pulse_position[3] = 4 - pulse_position[3];
458 pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1);
459 pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1);
462 for (i = 0; i < 4; i++) {
463 const int pos1 = (pulse_position[i] << 2) + i;
464 const int pos2 = (pulse_position[i + 4] << 2) + i;
465 const float sign = fixed_index[i] ? -1.0 : 1.0;
466 fixed_sparse->x[i ] = pos1;
467 fixed_sparse->x[i + 4] = pos2;
468 fixed_sparse->y[i ] = sign;
469 fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign;
474 * Decode the algebraic codebook index to pulse positions and signs,
475 * then construct the algebraic codebook vector.
477 * nb of pulses | bits encoding pulses
478 * For MODE_4k75 or MODE_5k15, 2 | 1-3, 4-6, 7
479 * MODE_5k9, 2 | 1, 2-4, 5-6, 7-9
480 * MODE_6k7, 3 | 1-3, 4, 5-7, 8, 9-11
481 * MODE_7k4 or MODE_7k95, 4 | 1-3, 4-6, 7-9, 10, 11-13
483 * @param fixed_sparse pointer to the algebraic codebook vector
484 * @param pulses algebraic codebook indexes
485 * @param mode mode of the current frame
486 * @param subframe current subframe number
488 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses,
489 const enum Mode mode, const int subframe)
491 assert(MODE_4k75 <= mode && mode <= MODE_12k2);
493 if (mode == MODE_12k2) {
494 ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3);
495 } else if (mode == MODE_10k2) {
496 decode_8_pulses_31bits(pulses, fixed_sparse);
498 int *pulse_position = fixed_sparse->x;
500 const int fixed_index = pulses[0];
502 if (mode <= MODE_5k15) {
503 pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1);
504 pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset];
505 pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1];
507 } else if (mode == MODE_5k9) {
508 pulse_subset = ((fixed_index & 1) << 1) + 1;
509 pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset;
510 pulse_subset = (fixed_index >> 4) & 3;
511 pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0);
512 fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2;
513 } else if (mode == MODE_6k7) {
514 pulse_position[0] = (fixed_index & 7) * 5;
515 pulse_subset = (fixed_index >> 2) & 2;
516 pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1;
517 pulse_subset = (fixed_index >> 6) & 2;
518 pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2;
520 } else { // mode <= MODE_7k95
521 pulse_position[0] = gray_decode[ fixed_index & 7];
522 pulse_position[1] = gray_decode[(fixed_index >> 3) & 7] + 1;
523 pulse_position[2] = gray_decode[(fixed_index >> 6) & 7] + 2;
524 pulse_subset = (fixed_index >> 9) & 1;
525 pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3;
528 for (i = 0; i < fixed_sparse->n; i++)
529 fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0;
534 * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2)
536 * @param p the context
537 * @param subframe unpacked amr subframe
538 * @param mode mode of the current frame
539 * @param fixed_sparse sparse respresentation of the fixed vector
541 static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode,
542 AMRFixed *fixed_sparse)
544 // The spec suggests the current pitch gain is always used, but in other
545 // modes the pitch and codebook gains are joinly quantized (sec 5.8.2)
546 // so the codebook gain cannot depend on the quantized pitch gain.
547 if (mode == MODE_12k2)
548 p->beta = FFMIN(p->pitch_gain[4], 1.0);
550 fixed_sparse->pitch_lag = p->pitch_lag_int;
551 fixed_sparse->pitch_fac = p->beta;
553 // Save pitch sharpening factor for the next subframe
554 // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from
555 // the fact that the gains for two subframes are jointly quantized.
556 if (mode != MODE_4k75 || subframe & 1)
557 p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX);
562 /// @name AMR gain decoding functions
566 * fixed gain smoothing
567 * Note that where the spec specifies the "spectrum in the q domain"
568 * in section 6.1.4, in fact frequencies should be used.
570 * @param p the context
571 * @param lsf LSFs for the current subframe, in the range [0,1]
572 * @param lsf_avg averaged LSFs
573 * @param mode mode of the current frame
575 * @return fixed gain smoothed
577 static float fixed_gain_smooth(AMRContext *p , const float *lsf,
578 const float *lsf_avg, const enum Mode mode)
583 for (i = 0; i < LP_FILTER_ORDER; i++)
584 diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i];
586 // If diff is large for ten subframes, disable smoothing for a 40-subframe
592 if (p->diff_count > 10) {
594 p->diff_count--; // don't let diff_count overflow
597 if (p->hang_count < 40) {
599 } else if (mode < MODE_7k4 || mode == MODE_10k2) {
600 const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0);
601 const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] +
602 p->fixed_gain[2] + p->fixed_gain[3] +
603 p->fixed_gain[4]) * 0.2;
604 return smoothing_factor * p->fixed_gain[4] +
605 (1.0 - smoothing_factor) * fixed_gain_mean;
607 return p->fixed_gain[4];
611 * Decode pitch gain and fixed gain factor (part of section 6.1.3).
613 * @param p the context
614 * @param amr_subframe unpacked amr subframe
615 * @param mode mode of the current frame
616 * @param subframe current subframe number
617 * @param fixed_gain_factor decoded gain correction factor
619 static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe,
620 const enum Mode mode, const int subframe,
621 float *fixed_gain_factor)
623 if (mode == MODE_12k2 || mode == MODE_7k95) {
624 p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ]
626 *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain]
629 const uint16_t *gains;
631 if (mode >= MODE_6k7) {
632 gains = gains_high[amr_subframe->p_gain];
633 } else if (mode >= MODE_5k15) {
634 gains = gains_low [amr_subframe->p_gain];
636 // gain index is only coded in subframes 0,2 for MODE_4k75
637 gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)];
640 p->pitch_gain[4] = gains[0] * (1.0 / 16384.0);
641 *fixed_gain_factor = gains[1] * (1.0 / 4096.0);
648 /// @name AMR preprocessing functions
652 * Circularly convolve a sparse fixed vector with a phase dispersion impulse
653 * response filter (D.6.2 of G.729 and 6.1.5 of AMR).
655 * @param out vector with filter applied
656 * @param in source vector
657 * @param filter phase filter coefficients
659 * out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] }
661 static void apply_ir_filter(float *out, const AMRFixed *in,
664 float filter1[AMR_SUBFRAME_SIZE], ///< filters at pitch lag*1 and *2
665 filter2[AMR_SUBFRAME_SIZE];
666 int lag = in->pitch_lag;
667 float fac = in->pitch_fac;
670 if (lag < AMR_SUBFRAME_SIZE) {
671 ff_celp_circ_addf(filter1, filter, filter, lag, fac,
674 if (lag < AMR_SUBFRAME_SIZE >> 1)
675 ff_celp_circ_addf(filter2, filter, filter1, lag, fac,
679 memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE);
680 for (i = 0; i < in->n; i++) {
683 const float *filterp;
685 if (x >= AMR_SUBFRAME_SIZE - lag) {
687 } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) {
692 ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE);
697 * Reduce fixed vector sparseness by smoothing with one of three IR filters.
698 * Also know as "adaptive phase dispersion".
700 * This implements 3GPP TS 26.090 section 6.1(5).
702 * @param p the context
703 * @param fixed_sparse algebraic codebook vector
704 * @param fixed_vector unfiltered fixed vector
705 * @param fixed_gain smoothed gain
706 * @param out space for modified vector if necessary
708 static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse,
709 const float *fixed_vector,
710 float fixed_gain, float *out)
714 if (p->pitch_gain[4] < 0.6) {
715 ir_filter_nr = 0; // strong filtering
716 } else if (p->pitch_gain[4] < 0.9) {
717 ir_filter_nr = 1; // medium filtering
719 ir_filter_nr = 2; // no filtering
722 if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) {
723 p->ir_filter_onset = 2;
724 } else if (p->ir_filter_onset)
725 p->ir_filter_onset--;
727 if (!p->ir_filter_onset) {
730 for (i = 0; i < 5; i++)
731 if (p->pitch_gain[i] < 0.6)
736 if (ir_filter_nr > p->prev_ir_filter_nr + 1)
738 } else if (ir_filter_nr < 2)
741 // Disable filtering for very low level of fixed_gain.
742 // Note this step is not specified in the technical description but is in
743 // the reference source in the function Ph_disp.
744 if (fixed_gain < 5.0)
747 if (p->cur_frame_mode != MODE_7k4 && p->cur_frame_mode < MODE_10k2
748 && ir_filter_nr < 2) {
749 apply_ir_filter(out, fixed_sparse,
750 (p->cur_frame_mode == MODE_7k95 ?
751 ir_filters_lookup_MODE_7k95 :
752 ir_filters_lookup)[ir_filter_nr]);
756 // update ir filter strength history
757 p->prev_ir_filter_nr = ir_filter_nr;
758 p->prev_sparse_fixed_gain = fixed_gain;
766 /// @name AMR synthesis functions
770 * Conduct 10th order linear predictive coding synthesis.
772 * @param p pointer to the AMRContext
773 * @param lpc pointer to the LPC coefficients
774 * @param fixed_gain fixed codebook gain for synthesis
775 * @param fixed_vector algebraic codebook vector
776 * @param samples pointer to the output speech samples
777 * @param overflow 16-bit overflow flag
779 static int synthesis(AMRContext *p, float *lpc,
780 float fixed_gain, const float *fixed_vector,
781 float *samples, uint8_t overflow)
784 float excitation[AMR_SUBFRAME_SIZE];
786 // if an overflow has been detected, the pitch vector is scaled down by a
789 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
790 p->pitch_vector[i] *= 0.25;
792 ff_weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector,
793 p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE);
795 // emphasize pitch vector contribution
796 if (p->pitch_gain[4] > 0.5 && !overflow) {
797 float energy = ff_scalarproduct_float_c(excitation, excitation,
801 (p->cur_frame_mode == MODE_12k2 ?
802 0.25 * FFMIN(p->pitch_gain[4], 1.0) :
803 0.5 * FFMIN(p->pitch_gain[4], SHARP_MAX));
805 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
806 excitation[i] += pitch_factor * p->pitch_vector[i];
808 ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy,
812 ff_celp_lp_synthesis_filterf(samples, lpc, excitation, AMR_SUBFRAME_SIZE,
816 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
817 if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) {
827 /// @name AMR update functions
831 * Update buffers and history at the end of decoding a subframe.
833 * @param p pointer to the AMRContext
835 static void update_state(AMRContext *p)
837 memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0]));
839 memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE],
840 (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float));
842 memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float));
843 memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float));
845 memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE],
846 LP_FILTER_ORDER * sizeof(float));
852 /// @name AMR Postprocessing functions
856 * Get the tilt factor of a formant filter from its transfer function
858 * @param lpc_n LP_FILTER_ORDER coefficients of the numerator
859 * @param lpc_d LP_FILTER_ORDER coefficients of the denominator
861 static float tilt_factor(float *lpc_n, float *lpc_d)
863 float rh0, rh1; // autocorrelation at lag 0 and 1
865 // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf
866 float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 };
867 float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response
870 memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER);
871 ff_celp_lp_synthesis_filterf(hf, lpc_d, hf, AMR_TILT_RESPONSE,
874 rh0 = ff_scalarproduct_float_c(hf, hf, AMR_TILT_RESPONSE);
875 rh1 = ff_scalarproduct_float_c(hf, hf + 1, AMR_TILT_RESPONSE - 1);
877 // The spec only specifies this check for 12.2 and 10.2 kbit/s
878 // modes. But in the ref source the tilt is always non-negative.
879 return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0;
883 * Perform adaptive post-filtering to enhance the quality of the speech.
886 * @param p pointer to the AMRContext
887 * @param lpc interpolated LP coefficients for this subframe
888 * @param buf_out output of the filter
890 static void postfilter(AMRContext *p, float *lpc, float *buf_out)
893 float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input
895 float speech_gain = ff_scalarproduct_float_c(samples, samples,
898 float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter
899 const float *gamma_n, *gamma_d; // Formant filter factor table
900 float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients
902 if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) {
903 gamma_n = ff_pow_0_7;
904 gamma_d = ff_pow_0_75;
906 gamma_n = ff_pow_0_55;
907 gamma_d = ff_pow_0_7;
910 for (i = 0; i < LP_FILTER_ORDER; i++) {
911 lpc_n[i] = lpc[i] * gamma_n[i];
912 lpc_d[i] = lpc[i] * gamma_d[i];
915 memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER);
916 ff_celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples,
917 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
918 memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE,
919 sizeof(float) * LP_FILTER_ORDER);
921 ff_celp_lp_zero_synthesis_filterf(buf_out, lpc_n,
922 pole_out + LP_FILTER_ORDER,
923 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
925 ff_tilt_compensation(&p->tilt_mem, tilt_factor(lpc_n, lpc_d), buf_out,
928 ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE,
929 AMR_AGC_ALPHA, &p->postfilter_agc);
934 static int amrnb_decode_frame(AVCodecContext *avctx, void *data,
935 int *got_frame_ptr, AVPacket *avpkt)
938 AMRContext *p = avctx->priv_data; // pointer to private data
939 const uint8_t *buf = avpkt->data;
940 int buf_size = avpkt->size;
941 float *buf_out; // pointer to the output data buffer
942 int i, subframe, ret;
943 float fixed_gain_factor;
944 AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing
945 float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing
946 float synth_fixed_gain; // the fixed gain that synthesis should use
947 const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
949 /* get output buffer */
950 p->avframe.nb_samples = AMR_BLOCK_SIZE;
951 if ((ret = ff_get_buffer(avctx, &p->avframe)) < 0) {
952 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
955 buf_out = (float *)p->avframe.data[0];
957 p->cur_frame_mode = unpack_bitstream(p, buf, buf_size);
958 if (p->cur_frame_mode == NO_DATA) {
959 av_log(avctx, AV_LOG_ERROR, "Corrupt bitstream\n");
960 return AVERROR_INVALIDDATA;
962 if (p->cur_frame_mode == MODE_DTX) {
963 av_log_missing_feature(avctx, "dtx mode", 1);
964 return AVERROR_PATCHWELCOME;
967 if (p->cur_frame_mode == MODE_12k2) {
972 for (i = 0; i < 4; i++)
973 ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5);
975 for (subframe = 0; subframe < 4; subframe++) {
976 const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe];
978 decode_pitch_vector(p, amr_subframe, subframe);
980 decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses,
981 p->cur_frame_mode, subframe);
983 // The fixed gain (section 6.1.3) depends on the fixed vector
984 // (section 6.1.2), but the fixed vector calculation uses
985 // pitch sharpening based on the on the pitch gain (section 6.1.3).
986 // So the correct order is: pitch gain, pitch sharpening, fixed gain.
987 decode_gains(p, amr_subframe, p->cur_frame_mode, subframe,
990 pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse);
992 if (fixed_sparse.pitch_lag == 0) {
993 av_log(avctx, AV_LOG_ERROR, "The file is corrupted, pitch_lag = 0 is not allowed\n");
994 return AVERROR_INVALIDDATA;
996 ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0,
1000 ff_amr_set_fixed_gain(fixed_gain_factor,
1001 ff_scalarproduct_float_c(p->fixed_vector,
1003 AMR_SUBFRAME_SIZE) /
1005 p->prediction_error,
1006 energy_mean[p->cur_frame_mode], energy_pred_fac);
1008 // The excitation feedback is calculated without any processing such
1009 // as fixed gain smoothing. This isn't mentioned in the specification.
1010 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1011 p->excitation[i] *= p->pitch_gain[4];
1012 ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4],
1015 // In the ref decoder, excitation is stored with no fractional bits.
1016 // This step prevents buzz in silent periods. The ref encoder can
1017 // emit long sequences with pitch factor greater than one. This
1018 // creates unwanted feedback if the excitation vector is nonzero.
1019 // (e.g. test sequence T19_795.COD in 3GPP TS 26.074)
1020 for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1021 p->excitation[i] = truncf(p->excitation[i]);
1023 // Smooth fixed gain.
1024 // The specification is ambiguous, but in the reference source, the
1025 // smoothed value is NOT fed back into later fixed gain smoothing.
1026 synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe],
1027 p->lsf_avg, p->cur_frame_mode);
1029 synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector,
1030 synth_fixed_gain, spare_vector);
1032 if (synthesis(p, p->lpc[subframe], synth_fixed_gain,
1033 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0))
1034 // overflow detected -> rerun synthesis scaling pitch vector down
1035 // by a factor of 4, skipping pitch vector contribution emphasis
1036 // and adaptive gain control
1037 synthesis(p, p->lpc[subframe], synth_fixed_gain,
1038 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1);
1040 postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE);
1042 // update buffers and history
1043 ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE);
1047 ff_acelp_apply_order_2_transfer_function(buf_out, buf_out, highpass_zeros,
1049 highpass_gain * AMR_SAMPLE_SCALE,
1050 p->high_pass_mem, AMR_BLOCK_SIZE);
1052 /* Update averaged lsf vector (used for fixed gain smoothing).
1054 * Note that lsf_avg should not incorporate the current frame's LSFs
1055 * for fixed_gain_smooth.
1056 * The specification has an incorrect formula: the reference decoder uses
1057 * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */
1058 ff_weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3],
1059 0.84, 0.16, LP_FILTER_ORDER);
1062 *(AVFrame *)data = p->avframe;
1064 /* return the amount of bytes consumed if everything was OK */
1065 return frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC
1069 AVCodec ff_amrnb_decoder = {
1071 .type = AVMEDIA_TYPE_AUDIO,
1072 .id = AV_CODEC_ID_AMR_NB,
1073 .priv_data_size = sizeof(AMRContext),
1074 .init = amrnb_decode_init,
1075 .decode = amrnb_decode_frame,
1076 .capabilities = CODEC_CAP_DR1,
1077 .long_name = NULL_IF_CONFIG_SMALL("AMR-NB (Adaptive Multi-Rate NarrowBand)"),
1078 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1079 AV_SAMPLE_FMT_NONE },