FFmpeg
 All Data Structures Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
amrnbdec.c
Go to the documentation of this file.
1 /*
2  * AMR narrowband decoder
3  * Copyright (c) 2006-2007 Robert Swain
4  * Copyright (c) 2009 Colin McQuillan
5  *
6  * This file is part of FFmpeg.
7  *
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.
12  *
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.
17  *
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
21  */
22 
23 
24 /**
25  * @file
26  * AMR narrowband decoder
27  *
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.
31  *
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.
35  *
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
39  * out of 169 tests.
40  */
41 
42 
43 #include <string.h>
44 #include <math.h>
45 
47 #include "avcodec.h"
48 #include "dsputil.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"
56 #include "lsp.h"
57 #include "amr.h"
58 #include "internal.h"
59 
60 #include "amrnbdata.h"
61 
62 #define AMR_BLOCK_SIZE 160 ///< samples per frame
63 #define AMR_SAMPLE_BOUND 32768.0 ///< threshold for synthesis overflow
64 
65 /**
66  * Scale from constructed speech to [-1,1]
67  *
68  * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but
69  * upscales by two (section 6.2.2).
70  *
71  * Fundamentally, this scale is determined by energy_mean through
72  * the fixed vector contribution to the excitation vector.
73  */
74 #define AMR_SAMPLE_SCALE (2.0 / 32768.0)
75 
76 /** Prediction factor for 12.2kbit/s mode */
77 #define PRED_FAC_MODE_12k2 0.65
78 
79 #define LSF_R_FAC (8000.0 / 32768.0) ///< LSF residual tables to Hertz
80 #define MIN_LSF_SPACING (50.0488 / 8000.0) ///< Ensures stability of LPC filter
81 #define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode
82 
83 /** Initial energy in dB. Also used for bad frames (unimplemented). */
84 #define MIN_ENERGY -14.0
85 
86 /** Maximum sharpening factor
87  *
88  * The specification says 0.8, which should be 13107, but the reference C code
89  * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in g729dec.c.)
90  */
91 #define SHARP_MAX 0.79449462890625
92 
93 /** Number of impulse response coefficients used for tilt factor */
94 #define AMR_TILT_RESPONSE 22
95 /** Tilt factor = 1st reflection coefficient * gamma_t */
96 #define AMR_TILT_GAMMA_T 0.8
97 /** Adaptive gain control factor used in post-filter */
98 #define AMR_AGC_ALPHA 0.9
99 
100 typedef struct AMRContext {
101  AVFrame avframe; ///< AVFrame for decoded samples
102  AMRNBFrame frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc)
103  uint8_t bad_frame_indicator; ///< bad frame ? 1 : 0
105 
106  int16_t prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe
107  double lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame
108  double prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame
109 
110  float lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing
111  float lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector
112 
113  float lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes
114 
115  uint8_t pitch_lag_int; ///< integer part of pitch lag from current subframe
116 
117  float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history
118  float *excitation; ///< pointer to the current excitation vector in excitation_buf
119 
120  float pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector
121  float fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames)
122 
123  float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes
124  float pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes
125  float fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes
126 
127  float beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX]
128  uint8_t diff_count; ///< the number of subframes for which diff has been above 0.65
129  uint8_t hang_count; ///< the number of subframes since a hangover period started
130 
131  float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset"
132  uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none
133  uint8_t ir_filter_onset; ///< flag for impulse response filter strength
134 
135  float postfilter_mem[10]; ///< previous intermediate values in the formant filter
136  float tilt_mem; ///< previous input to tilt compensation filter
137  float postfilter_agc; ///< previous factor used for adaptive gain control
138  float high_pass_mem[2]; ///< previous intermediate values in the high-pass filter
139 
140  float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples
141 
142  ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs
143  ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs
144  CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs
145  CELPMContext celpm_ctx; ///< context for fixed point math operations
146 
147 } AMRContext;
148 
149 /** Double version of ff_weighted_vector_sumf() */
150 static void weighted_vector_sumd(double *out, const double *in_a,
151  const double *in_b, double weight_coeff_a,
152  double weight_coeff_b, int length)
153 {
154  int i;
155 
156  for (i = 0; i < length; i++)
157  out[i] = weight_coeff_a * in_a[i]
158  + weight_coeff_b * in_b[i];
159 }
160 
162 {
163  AMRContext *p = avctx->priv_data;
164  int i;
165 
166  if (avctx->channels > 1) {
167  av_log_missing_feature(avctx, "multi-channel AMR", 0);
168  return AVERROR_PATCHWELCOME;
169  }
170 
171  avctx->channels = 1;
173  if (!avctx->sample_rate)
174  avctx->sample_rate = 8000;
175  avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
176 
177  // p->excitation always points to the same position in p->excitation_buf
179 
180  for (i = 0; i < LP_FILTER_ORDER; i++) {
181  p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15);
182  p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15);
183  }
184 
185  for (i = 0; i < 4; i++)
187 
189  avctx->coded_frame = &p->avframe;
190 
195 
196  return 0;
197 }
198 
199 
200 /**
201  * Unpack an RFC4867 speech frame into the AMR frame mode and parameters.
202  *
203  * The order of speech bits is specified by 3GPP TS 26.101.
204  *
205  * @param p the context
206  * @param buf pointer to the input buffer
207  * @param buf_size size of the input buffer
208  *
209  * @return the frame mode
210  */
211 static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf,
212  int buf_size)
213 {
214  enum Mode mode;
215 
216  // Decode the first octet.
217  mode = buf[0] >> 3 & 0x0F; // frame type
218  p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit
219 
220  if (mode >= N_MODES || buf_size < frame_sizes_nb[mode] + 1) {
221  return NO_DATA;
222  }
223 
224  if (mode < MODE_DTX)
225  ff_amr_bit_reorder((uint16_t *) &p->frame, sizeof(AMRNBFrame), buf + 1,
227 
228  return mode;
229 }
230 
231 
232 /// @name AMR pitch LPC coefficient decoding functions
233 /// @{
234 
235 /**
236  * Interpolate the LSF vector (used for fixed gain smoothing).
237  * The interpolation is done over all four subframes even in MODE_12k2.
238  *
239  * @param[in] ctx The Context
240  * @param[in,out] lsf_q LSFs in [0,1] for each subframe
241  * @param[in] lsf_new New LSFs in [0,1] for subframe 4
242  */
243 static void interpolate_lsf(ACELPVContext *ctx, float lsf_q[4][LP_FILTER_ORDER], float *lsf_new)
244 {
245  int i;
246 
247  for (i = 0; i < 4; i++)
248  ctx->weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new,
249  0.25 * (3 - i), 0.25 * (i + 1),
250  LP_FILTER_ORDER);
251 }
252 
253 /**
254  * Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector.
255  *
256  * @param p the context
257  * @param lsp output LSP vector
258  * @param lsf_no_r LSF vector without the residual vector added
259  * @param lsf_quantizer pointers to LSF dictionary tables
260  * @param quantizer_offset offset in tables
261  * @param sign for the 3 dictionary table
262  * @param update store data for computing the next frame's LSFs
263  */
264 static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER],
265  const float lsf_no_r[LP_FILTER_ORDER],
266  const int16_t *lsf_quantizer[5],
267  const int quantizer_offset,
268  const int sign, const int update)
269 {
270  int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
271  float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
272  int i;
273 
274  for (i = 0; i < LP_FILTER_ORDER >> 1; i++)
275  memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset],
276  2 * sizeof(*lsf_r));
277 
278  if (sign) {
279  lsf_r[4] *= -1;
280  lsf_r[5] *= -1;
281  }
282 
283  if (update)
284  memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
285 
286  for (i = 0; i < LP_FILTER_ORDER; i++)
287  lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0);
288 
289  ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
290 
291  if (update)
292  interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
293 
294  ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER);
295 }
296 
297 /**
298  * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors.
299  *
300  * @param p pointer to the AMRContext
301  */
302 static void lsf2lsp_5(AMRContext *p)
303 {
304  const uint16_t *lsf_param = p->frame.lsf;
305  float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector
306  const int16_t *lsf_quantizer[5];
307  int i;
308 
309  lsf_quantizer[0] = lsf_5_1[lsf_param[0]];
310  lsf_quantizer[1] = lsf_5_2[lsf_param[1]];
311  lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1];
312  lsf_quantizer[3] = lsf_5_4[lsf_param[3]];
313  lsf_quantizer[4] = lsf_5_5[lsf_param[4]];
314 
315  for (i = 0; i < LP_FILTER_ORDER; i++)
316  lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i];
317 
318  lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0);
319  lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1);
320 
321  // interpolate LSP vectors at subframes 1 and 3
322  weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER);
323  weighted_vector_sumd(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER);
324 }
325 
326 /**
327  * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector.
328  *
329  * @param p pointer to the AMRContext
330  */
331 static void lsf2lsp_3(AMRContext *p)
332 {
333  const uint16_t *lsf_param = p->frame.lsf;
334  int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector
335  float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector
336  const int16_t *lsf_quantizer;
337  int i, j;
338 
339  lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]];
340  memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r));
341 
342  lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)];
343  memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r));
344 
345  lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]];
346  memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r));
347 
348  // calculate mean-removed LSF vector and add mean
349  for (i = 0; i < LP_FILTER_ORDER; i++)
350  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 
352  ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER);
353 
354  // store data for computing the next frame's LSFs
355  interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q);
356  memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r));
357 
358  ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER);
359 
360  // interpolate LSP vectors at subframes 1, 2 and 3
361  for (i = 1; i <= 3; i++)
362  for(j = 0; j < LP_FILTER_ORDER; j++)
363  p->lsp[i-1][j] = p->prev_lsp_sub4[j] +
364  (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i;
365 }
366 
367 /// @}
368 
369 
370 /// @name AMR pitch vector decoding functions
371 /// @{
372 
373 /**
374  * Like ff_decode_pitch_lag(), but with 1/6 resolution
375  */
376 static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index,
377  const int prev_lag_int, const int subframe)
378 {
379  if (subframe == 0 || subframe == 2) {
380  if (pitch_index < 463) {
381  *lag_int = (pitch_index + 107) * 10923 >> 16;
382  *lag_frac = pitch_index - *lag_int * 6 + 105;
383  } else {
384  *lag_int = pitch_index - 368;
385  *lag_frac = 0;
386  }
387  } else {
388  *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1;
389  *lag_frac = pitch_index - *lag_int * 6 - 3;
390  *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2,
391  PITCH_DELAY_MAX - 9);
392  }
393 }
394 
396  const AMRNBSubframe *amr_subframe,
397  const int subframe)
398 {
399  int pitch_lag_int, pitch_lag_frac;
400  enum Mode mode = p->cur_frame_mode;
401 
402  if (p->cur_frame_mode == MODE_12k2) {
403  decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac,
404  amr_subframe->p_lag, p->pitch_lag_int,
405  subframe);
406  } else
407  ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac,
408  amr_subframe->p_lag,
409  p->pitch_lag_int, subframe,
410  mode != MODE_4k75 && mode != MODE_5k15,
411  mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6));
412 
413  p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t
414 
415  pitch_lag_frac <<= (p->cur_frame_mode != MODE_12k2);
416 
417  pitch_lag_int += pitch_lag_frac > 0;
418 
419  /* Calculate the pitch vector by interpolating the past excitation at the
420  pitch lag using a b60 hamming windowed sinc function. */
422  p->excitation + 1 - pitch_lag_int,
423  ff_b60_sinc, 6,
424  pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0),
425  10, AMR_SUBFRAME_SIZE);
426 
427  memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float));
428 }
429 
430 /// @}
431 
432 
433 /// @name AMR algebraic code book (fixed) vector decoding functions
434 /// @{
435 
436 /**
437  * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame.
438  */
439 static void decode_10bit_pulse(int code, int pulse_position[8],
440  int i1, int i2, int i3)
441 {
442  // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of
443  // the 3 pulses and the upper 7 bits being coded in base 5
444  const uint8_t *positions = base_five_table[code >> 3];
445  pulse_position[i1] = (positions[2] << 1) + ( code & 1);
446  pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1);
447  pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1);
448 }
449 
450 /**
451  * Decode the algebraic codebook index to pulse positions and signs and
452  * construct the algebraic codebook vector for MODE_10k2.
453  *
454  * @param fixed_index positions of the eight pulses
455  * @param fixed_sparse pointer to the algebraic codebook vector
456  */
457 static void decode_8_pulses_31bits(const int16_t *fixed_index,
458  AMRFixed *fixed_sparse)
459 {
460  int pulse_position[8];
461  int i, temp;
462 
463  decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1);
464  decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5);
465 
466  // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of
467  // the 2 pulses and the upper 5 bits being coded in base 5
468  temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5;
469  pulse_position[3] = temp % 5;
470  pulse_position[7] = temp / 5;
471  if (pulse_position[7] & 1)
472  pulse_position[3] = 4 - pulse_position[3];
473  pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1);
474  pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1);
475 
476  fixed_sparse->n = 8;
477  for (i = 0; i < 4; i++) {
478  const int pos1 = (pulse_position[i] << 2) + i;
479  const int pos2 = (pulse_position[i + 4] << 2) + i;
480  const float sign = fixed_index[i] ? -1.0 : 1.0;
481  fixed_sparse->x[i ] = pos1;
482  fixed_sparse->x[i + 4] = pos2;
483  fixed_sparse->y[i ] = sign;
484  fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign;
485  }
486 }
487 
488 /**
489  * Decode the algebraic codebook index to pulse positions and signs,
490  * then construct the algebraic codebook vector.
491  *
492  * nb of pulses | bits encoding pulses
493  * For MODE_4k75 or MODE_5k15, 2 | 1-3, 4-6, 7
494  * MODE_5k9, 2 | 1, 2-4, 5-6, 7-9
495  * MODE_6k7, 3 | 1-3, 4, 5-7, 8, 9-11
496  * MODE_7k4 or MODE_7k95, 4 | 1-3, 4-6, 7-9, 10, 11-13
497  *
498  * @param fixed_sparse pointer to the algebraic codebook vector
499  * @param pulses algebraic codebook indexes
500  * @param mode mode of the current frame
501  * @param subframe current subframe number
502  */
503 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses,
504  const enum Mode mode, const int subframe)
505 {
506  av_assert1(MODE_4k75 <= (signed)mode && mode <= MODE_12k2);
507 
508  if (mode == MODE_12k2) {
509  ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3);
510  } else if (mode == MODE_10k2) {
511  decode_8_pulses_31bits(pulses, fixed_sparse);
512  } else {
513  int *pulse_position = fixed_sparse->x;
514  int i, pulse_subset;
515  const int fixed_index = pulses[0];
516 
517  if (mode <= MODE_5k15) {
518  pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1);
519  pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset];
520  pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1];
521  fixed_sparse->n = 2;
522  } else if (mode == MODE_5k9) {
523  pulse_subset = ((fixed_index & 1) << 1) + 1;
524  pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset;
525  pulse_subset = (fixed_index >> 4) & 3;
526  pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0);
527  fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2;
528  } else if (mode == MODE_6k7) {
529  pulse_position[0] = (fixed_index & 7) * 5;
530  pulse_subset = (fixed_index >> 2) & 2;
531  pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1;
532  pulse_subset = (fixed_index >> 6) & 2;
533  pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2;
534  fixed_sparse->n = 3;
535  } else { // mode <= MODE_7k95
536  pulse_position[0] = gray_decode[ fixed_index & 7];
537  pulse_position[1] = gray_decode[(fixed_index >> 3) & 7] + 1;
538  pulse_position[2] = gray_decode[(fixed_index >> 6) & 7] + 2;
539  pulse_subset = (fixed_index >> 9) & 1;
540  pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3;
541  fixed_sparse->n = 4;
542  }
543  for (i = 0; i < fixed_sparse->n; i++)
544  fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0;
545  }
546 }
547 
548 /**
549  * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2)
550  *
551  * @param p the context
552  * @param subframe unpacked amr subframe
553  * @param mode mode of the current frame
554  * @param fixed_sparse sparse respresentation of the fixed vector
555  */
556 static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode,
557  AMRFixed *fixed_sparse)
558 {
559  // The spec suggests the current pitch gain is always used, but in other
560  // modes the pitch and codebook gains are joinly quantized (sec 5.8.2)
561  // so the codebook gain cannot depend on the quantized pitch gain.
562  if (mode == MODE_12k2)
563  p->beta = FFMIN(p->pitch_gain[4], 1.0);
564 
565  fixed_sparse->pitch_lag = p->pitch_lag_int;
566  fixed_sparse->pitch_fac = p->beta;
567 
568  // Save pitch sharpening factor for the next subframe
569  // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from
570  // the fact that the gains for two subframes are jointly quantized.
571  if (mode != MODE_4k75 || subframe & 1)
572  p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX);
573 }
574 /// @}
575 
576 
577 /// @name AMR gain decoding functions
578 /// @{
579 
580 /**
581  * fixed gain smoothing
582  * Note that where the spec specifies the "spectrum in the q domain"
583  * in section 6.1.4, in fact frequencies should be used.
584  *
585  * @param p the context
586  * @param lsf LSFs for the current subframe, in the range [0,1]
587  * @param lsf_avg averaged LSFs
588  * @param mode mode of the current frame
589  *
590  * @return fixed gain smoothed
591  */
592 static float fixed_gain_smooth(AMRContext *p , const float *lsf,
593  const float *lsf_avg, const enum Mode mode)
594 {
595  float diff = 0.0;
596  int i;
597 
598  for (i = 0; i < LP_FILTER_ORDER; i++)
599  diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i];
600 
601  // If diff is large for ten subframes, disable smoothing for a 40-subframe
602  // hangover period.
603  p->diff_count++;
604  if (diff <= 0.65)
605  p->diff_count = 0;
606 
607  if (p->diff_count > 10) {
608  p->hang_count = 0;
609  p->diff_count--; // don't let diff_count overflow
610  }
611 
612  if (p->hang_count < 40) {
613  p->hang_count++;
614  } else if (mode < MODE_7k4 || mode == MODE_10k2) {
615  const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0);
616  const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] +
617  p->fixed_gain[2] + p->fixed_gain[3] +
618  p->fixed_gain[4]) * 0.2;
619  return smoothing_factor * p->fixed_gain[4] +
620  (1.0 - smoothing_factor) * fixed_gain_mean;
621  }
622  return p->fixed_gain[4];
623 }
624 
625 /**
626  * Decode pitch gain and fixed gain factor (part of section 6.1.3).
627  *
628  * @param p the context
629  * @param amr_subframe unpacked amr subframe
630  * @param mode mode of the current frame
631  * @param subframe current subframe number
632  * @param fixed_gain_factor decoded gain correction factor
633  */
634 static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe,
635  const enum Mode mode, const int subframe,
636  float *fixed_gain_factor)
637 {
638  if (mode == MODE_12k2 || mode == MODE_7k95) {
639  p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ]
640  * (1.0 / 16384.0);
641  *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain]
642  * (1.0 / 2048.0);
643  } else {
644  const uint16_t *gains;
645 
646  if (mode >= MODE_6k7) {
647  gains = gains_high[amr_subframe->p_gain];
648  } else if (mode >= MODE_5k15) {
649  gains = gains_low [amr_subframe->p_gain];
650  } else {
651  // gain index is only coded in subframes 0,2 for MODE_4k75
652  gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)];
653  }
654 
655  p->pitch_gain[4] = gains[0] * (1.0 / 16384.0);
656  *fixed_gain_factor = gains[1] * (1.0 / 4096.0);
657  }
658 }
659 
660 /// @}
661 
662 
663 /// @name AMR preprocessing functions
664 /// @{
665 
666 /**
667  * Circularly convolve a sparse fixed vector with a phase dispersion impulse
668  * response filter (D.6.2 of G.729 and 6.1.5 of AMR).
669  *
670  * @param out vector with filter applied
671  * @param in source vector
672  * @param filter phase filter coefficients
673  *
674  * out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] }
675  */
676 static void apply_ir_filter(float *out, const AMRFixed *in,
677  const float *filter)
678 {
679  float filter1[AMR_SUBFRAME_SIZE], ///< filters at pitch lag*1 and *2
680  filter2[AMR_SUBFRAME_SIZE];
681  int lag = in->pitch_lag;
682  float fac = in->pitch_fac;
683  int i;
684 
685  if (lag < AMR_SUBFRAME_SIZE) {
686  ff_celp_circ_addf(filter1, filter, filter, lag, fac,
688 
689  if (lag < AMR_SUBFRAME_SIZE >> 1)
690  ff_celp_circ_addf(filter2, filter, filter1, lag, fac,
692  }
693 
694  memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE);
695  for (i = 0; i < in->n; i++) {
696  int x = in->x[i];
697  float y = in->y[i];
698  const float *filterp;
699 
700  if (x >= AMR_SUBFRAME_SIZE - lag) {
701  filterp = filter;
702  } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) {
703  filterp = filter1;
704  } else
705  filterp = filter2;
706 
707  ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE);
708  }
709 }
710 
711 /**
712  * Reduce fixed vector sparseness by smoothing with one of three IR filters.
713  * Also know as "adaptive phase dispersion".
714  *
715  * This implements 3GPP TS 26.090 section 6.1(5).
716  *
717  * @param p the context
718  * @param fixed_sparse algebraic codebook vector
719  * @param fixed_vector unfiltered fixed vector
720  * @param fixed_gain smoothed gain
721  * @param out space for modified vector if necessary
722  */
723 static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse,
724  const float *fixed_vector,
725  float fixed_gain, float *out)
726 {
727  int ir_filter_nr;
728 
729  if (p->pitch_gain[4] < 0.6) {
730  ir_filter_nr = 0; // strong filtering
731  } else if (p->pitch_gain[4] < 0.9) {
732  ir_filter_nr = 1; // medium filtering
733  } else
734  ir_filter_nr = 2; // no filtering
735 
736  // detect 'onset'
737  if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) {
738  p->ir_filter_onset = 2;
739  } else if (p->ir_filter_onset)
740  p->ir_filter_onset--;
741 
742  if (!p->ir_filter_onset) {
743  int i, count = 0;
744 
745  for (i = 0; i < 5; i++)
746  if (p->pitch_gain[i] < 0.6)
747  count++;
748  if (count > 2)
749  ir_filter_nr = 0;
750 
751  if (ir_filter_nr > p->prev_ir_filter_nr + 1)
752  ir_filter_nr--;
753  } else if (ir_filter_nr < 2)
754  ir_filter_nr++;
755 
756  // Disable filtering for very low level of fixed_gain.
757  // Note this step is not specified in the technical description but is in
758  // the reference source in the function Ph_disp.
759  if (fixed_gain < 5.0)
760  ir_filter_nr = 2;
761 
763  && ir_filter_nr < 2) {
764  apply_ir_filter(out, fixed_sparse,
765  (p->cur_frame_mode == MODE_7k95 ?
767  ir_filters_lookup)[ir_filter_nr]);
768  fixed_vector = out;
769  }
770 
771  // update ir filter strength history
772  p->prev_ir_filter_nr = ir_filter_nr;
773  p->prev_sparse_fixed_gain = fixed_gain;
774 
775  return fixed_vector;
776 }
777 
778 /// @}
779 
780 
781 /// @name AMR synthesis functions
782 /// @{
783 
784 /**
785  * Conduct 10th order linear predictive coding synthesis.
786  *
787  * @param p pointer to the AMRContext
788  * @param lpc pointer to the LPC coefficients
789  * @param fixed_gain fixed codebook gain for synthesis
790  * @param fixed_vector algebraic codebook vector
791  * @param samples pointer to the output speech samples
792  * @param overflow 16-bit overflow flag
793  */
794 static int synthesis(AMRContext *p, float *lpc,
795  float fixed_gain, const float *fixed_vector,
796  float *samples, uint8_t overflow)
797 {
798  int i;
799  float excitation[AMR_SUBFRAME_SIZE];
800 
801  // if an overflow has been detected, the pitch vector is scaled down by a
802  // factor of 4
803  if (overflow)
804  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
805  p->pitch_vector[i] *= 0.25;
806 
807  p->acelpv_ctx.weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector,
808  p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE);
809 
810  // emphasize pitch vector contribution
811  if (p->pitch_gain[4] > 0.5 && !overflow) {
812  float energy = p->celpm_ctx.dot_productf(excitation, excitation,
813  AMR_SUBFRAME_SIZE);
814  float pitch_factor =
815  p->pitch_gain[4] *
816  (p->cur_frame_mode == MODE_12k2 ?
817  0.25 * FFMIN(p->pitch_gain[4], 1.0) :
818  0.5 * FFMIN(p->pitch_gain[4], SHARP_MAX));
819 
820  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
821  excitation[i] += pitch_factor * p->pitch_vector[i];
822 
823  ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy,
824  AMR_SUBFRAME_SIZE);
825  }
826 
827  p->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,
828  AMR_SUBFRAME_SIZE,
830 
831  // detect overflow
832  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
833  if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) {
834  return 1;
835  }
836 
837  return 0;
838 }
839 
840 /// @}
841 
842 
843 /// @name AMR update functions
844 /// @{
845 
846 /**
847  * Update buffers and history at the end of decoding a subframe.
848  *
849  * @param p pointer to the AMRContext
850  */
851 static void update_state(AMRContext *p)
852 {
853  memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0]));
854 
855  memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE],
856  (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float));
857 
858  memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float));
859  memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float));
860 
861  memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE],
862  LP_FILTER_ORDER * sizeof(float));
863 }
864 
865 /// @}
866 
867 
868 /// @name AMR Postprocessing functions
869 /// @{
870 
871 /**
872  * Get the tilt factor of a formant filter from its transfer function
873  *
874  * @param p The Context
875  * @param lpc_n LP_FILTER_ORDER coefficients of the numerator
876  * @param lpc_d LP_FILTER_ORDER coefficients of the denominator
877  */
878 static float tilt_factor(AMRContext *p, float *lpc_n, float *lpc_d)
879 {
880  float rh0, rh1; // autocorrelation at lag 0 and 1
881 
882  // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf
883  float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 };
884  float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response
885 
886  hf[0] = 1.0;
887  memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER);
888  p->celpf_ctx.celp_lp_synthesis_filterf(hf, lpc_d, hf,
890  LP_FILTER_ORDER);
891 
892  rh0 = p->celpm_ctx.dot_productf(hf, hf, AMR_TILT_RESPONSE);
893  rh1 = p->celpm_ctx.dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1);
894 
895  // The spec only specifies this check for 12.2 and 10.2 kbit/s
896  // modes. But in the ref source the tilt is always non-negative.
897  return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0;
898 }
899 
900 /**
901  * Perform adaptive post-filtering to enhance the quality of the speech.
902  * See section 6.2.1.
903  *
904  * @param p pointer to the AMRContext
905  * @param lpc interpolated LP coefficients for this subframe
906  * @param buf_out output of the filter
907  */
908 static void postfilter(AMRContext *p, float *lpc, float *buf_out)
909 {
910  int i;
911  float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input
912 
913  float speech_gain = p->celpm_ctx.dot_productf(samples, samples,
915 
916  float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter
917  const float *gamma_n, *gamma_d; // Formant filter factor table
918  float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients
919 
920  if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) {
921  gamma_n = ff_pow_0_7;
922  gamma_d = ff_pow_0_75;
923  } else {
924  gamma_n = ff_pow_0_55;
925  gamma_d = ff_pow_0_7;
926  }
927 
928  for (i = 0; i < LP_FILTER_ORDER; i++) {
929  lpc_n[i] = lpc[i] * gamma_n[i];
930  lpc_d[i] = lpc[i] * gamma_d[i];
931  }
932 
933  memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER);
934  p->celpf_ctx.celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples,
935  AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
936  memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE,
937  sizeof(float) * LP_FILTER_ORDER);
938 
939  p->celpf_ctx.celp_lp_zero_synthesis_filterf(buf_out, lpc_n,
940  pole_out + LP_FILTER_ORDER,
941  AMR_SUBFRAME_SIZE, LP_FILTER_ORDER);
942 
943  ff_tilt_compensation(&p->tilt_mem, tilt_factor(p, lpc_n, lpc_d), buf_out,
945 
946  ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE,
948 }
949 
950 /// @}
951 
952 static int amrnb_decode_frame(AVCodecContext *avctx, void *data,
953  int *got_frame_ptr, AVPacket *avpkt)
954 {
955 
956  AMRContext *p = avctx->priv_data; // pointer to private data
957  const uint8_t *buf = avpkt->data;
958  int buf_size = avpkt->size;
959  float *buf_out; // pointer to the output data buffer
960  int i, subframe, ret;
961  float fixed_gain_factor;
962  AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing
963  float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing
964  float synth_fixed_gain; // the fixed gain that synthesis should use
965  const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use
966 
967  /* get output buffer */
969  if ((ret = ff_get_buffer(avctx, &p->avframe)) < 0) {
970  av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
971  return ret;
972  }
973  buf_out = (float *)p->avframe.data[0];
974 
975  p->cur_frame_mode = unpack_bitstream(p, buf, buf_size);
976  if (p->cur_frame_mode == NO_DATA) {
977  av_log(avctx, AV_LOG_ERROR, "Corrupt bitstream\n");
978  return AVERROR_INVALIDDATA;
979  }
980  if (p->cur_frame_mode == MODE_DTX) {
981  av_log_missing_feature(avctx, "dtx mode", 0);
982  av_log(avctx, AV_LOG_INFO, "Note: libopencore_amrnb supports dtx\n");
983  return AVERROR_PATCHWELCOME;
984  }
985 
986  if (p->cur_frame_mode == MODE_12k2) {
987  lsf2lsp_5(p);
988  } else
989  lsf2lsp_3(p);
990 
991  for (i = 0; i < 4; i++)
992  ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5);
993 
994  for (subframe = 0; subframe < 4; subframe++) {
995  const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe];
996 
997  decode_pitch_vector(p, amr_subframe, subframe);
998 
999  decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses,
1000  p->cur_frame_mode, subframe);
1001 
1002  // The fixed gain (section 6.1.3) depends on the fixed vector
1003  // (section 6.1.2), but the fixed vector calculation uses
1004  // pitch sharpening based on the on the pitch gain (section 6.1.3).
1005  // So the correct order is: pitch gain, pitch sharpening, fixed gain.
1006  decode_gains(p, amr_subframe, p->cur_frame_mode, subframe,
1007  &fixed_gain_factor);
1008 
1009  pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse);
1010 
1011  if (fixed_sparse.pitch_lag == 0) {
1012  av_log(avctx, AV_LOG_ERROR, "The file is corrupted, pitch_lag = 0 is not allowed\n");
1013  return AVERROR_INVALIDDATA;
1014  }
1015  ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0,
1017 
1018  p->fixed_gain[4] =
1019  ff_amr_set_fixed_gain(fixed_gain_factor,
1021  p->fixed_vector,
1024  p->prediction_error,
1026 
1027  // The excitation feedback is calculated without any processing such
1028  // as fixed gain smoothing. This isn't mentioned in the specification.
1029  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1030  p->excitation[i] *= p->pitch_gain[4];
1031  ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4],
1032  AMR_SUBFRAME_SIZE);
1033 
1034  // In the ref decoder, excitation is stored with no fractional bits.
1035  // This step prevents buzz in silent periods. The ref encoder can
1036  // emit long sequences with pitch factor greater than one. This
1037  // creates unwanted feedback if the excitation vector is nonzero.
1038  // (e.g. test sequence T19_795.COD in 3GPP TS 26.074)
1039  for (i = 0; i < AMR_SUBFRAME_SIZE; i++)
1040  p->excitation[i] = truncf(p->excitation[i]);
1041 
1042  // Smooth fixed gain.
1043  // The specification is ambiguous, but in the reference source, the
1044  // smoothed value is NOT fed back into later fixed gain smoothing.
1045  synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe],
1046  p->lsf_avg, p->cur_frame_mode);
1047 
1048  synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector,
1049  synth_fixed_gain, spare_vector);
1050 
1051  if (synthesis(p, p->lpc[subframe], synth_fixed_gain,
1052  synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0))
1053  // overflow detected -> rerun synthesis scaling pitch vector down
1054  // by a factor of 4, skipping pitch vector contribution emphasis
1055  // and adaptive gain control
1056  synthesis(p, p->lpc[subframe], synth_fixed_gain,
1057  synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1);
1058 
1059  postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE);
1060 
1061  // update buffers and history
1062  ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE);
1063  update_state(p);
1064  }
1065 
1067  buf_out, highpass_zeros,
1071 
1072  /* Update averaged lsf vector (used for fixed gain smoothing).
1073  *
1074  * Note that lsf_avg should not incorporate the current frame's LSFs
1075  * for fixed_gain_smooth.
1076  * The specification has an incorrect formula: the reference decoder uses
1077  * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */
1079  0.84, 0.16, LP_FILTER_ORDER);
1080 
1081  *got_frame_ptr = 1;
1082  *(AVFrame *)data = p->avframe;
1083 
1084  /* return the amount of bytes consumed if everything was OK */
1085  return frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC
1086 }
1087 
1088 
1090  .name = "amrnb",
1091  .type = AVMEDIA_TYPE_AUDIO,
1092  .id = AV_CODEC_ID_AMR_NB,
1093  .priv_data_size = sizeof(AMRContext),
1096  .capabilities = CODEC_CAP_DR1,
1097  .long_name = NULL_IF_CONFIG_SMALL("AMR-NB (Adaptive Multi-Rate NarrowBand)"),
1098  .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,
1100 };