FFmpeg: libavcodec/qcelpdec.c Source File

FFmpeg
qcelpdec.c
Go to the documentation of this file.
1 /*
2  * QCELP decoder
3  * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet
4  *
5  * This file is part of FFmpeg.
6  *
7  * FFmpeg is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU Lesser General Public
9  * License as published by the Free Software Foundation; either
10  * version 2.1 of the License, or (at your option) any later version.
11  *
12  * FFmpeg is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15  * Lesser General Public License for more details.
16  *
17  * You should have received a copy of the GNU Lesser General Public
18  * License along with FFmpeg; if not, write to the Free Software
19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20  */
21 
22 /**
23  * @file
24  * QCELP decoder
25  * @author Reynaldo H. Verdejo Pinochet
26  * @remark FFmpeg merging spearheaded by Kenan Gillet
27  * @remark Development mentored by Benjamin Larson
28  */
29 
30 #include <stddef.h>
31 
32 #include "libavutil/avassert.h"
33 #include "libavutil/channel_layout.h"
34 #include "libavutil/float_dsp.h"
35 #include "avcodec.h"
36 #include "internal.h"
37 #include "get_bits.h"
38 #include "qcelpdata.h"
39 #include "celp_filters.h"
40 #include "acelp_filters.h"
41 #include "acelp_vectors.h"
42 #include "lsp.h"
43 
44  typedef enum {
45   I_F_Q = -1, /**< insufficient frame quality */
46   SILENCE,
47   RATE_OCTAVE,
48   RATE_QUARTER,
49   RATE_HALF,
50   RATE_FULL
51 } qcelp_packet_rate;
52 
53  typedef struct {
54   GetBitContext gb;
55   qcelp_packet_rate bitrate;
56   QCELPFrame frame; /**< unpacked data frame */
57 
58   uint8_t erasure_count;
59   uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */
60   float prev_lspf[10];
61   float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */
62   float pitch_synthesis_filter_mem[303];
63   float pitch_pre_filter_mem[303];
64   float rnd_fir_filter_mem[180];
65   float formant_mem[170];
66   float last_codebook_gain;
67   int prev_g1[2];
68   int prev_bitrate;
69   float pitch_gain[4];
70   uint8_t pitch_lag[4];
71   uint16_t first16bits;
72   uint8_t warned_buf_mismatch_bitrate;
73 
74  /* postfilter */
75   float postfilter_synth_mem[10];
76   float postfilter_agc_mem;
77   float postfilter_tilt_mem;
78 } QCELPContext;
79 
80 /**
81  * Initialize the speech codec according to the specification.
82  *
83  * TIA/EIA/IS-733 2.4.9
84  */
85  static av_cold int qcelp_decode_init(AVCodecContext *avctx)
86 {
87  QCELPContext *q = avctx->priv_data;
88  int i;
89 
90  avctx->channels = 1;
91  avctx->channel_layout = AV_CH_LAYOUT_MONO;
92  avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
93 
94  for (i = 0; i < 10; i++)
95  q->prev_lspf[i] = (i + 1) / 11.0;
96 
97  return 0;
98 }
99 
100 /**
101  * Decode the 10 quantized LSP frequencies from the LSPV/LSP
102  * transmission codes of any bitrate and check for badly received packets.
103  *
104  * @param q the context
105  * @param lspf line spectral pair frequencies
106  *
107  * @return 0 on success, -1 if the packet is badly received
108  *
109  * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
110  */
111  static int decode_lspf(QCELPContext *q, float *lspf)
112 {
113  int i;
114  float tmp_lspf, smooth, erasure_coeff;
115  const float *predictors;
116 
117  if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) {
118  predictors = q->prev_bitrate != RATE_OCTAVE &&
119  q->prev_bitrate != I_F_Q ? q->prev_lspf
120  : q->predictor_lspf;
121 
122  if (q->bitrate == RATE_OCTAVE) {
123  q->octave_count++;
124 
125  for (i = 0; i < 10; i++) {
126  q->predictor_lspf[i] =
127  lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
128  : -QCELP_LSP_SPREAD_FACTOR) +
129  predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR +
130  (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11);
131  }
132  smooth = q->octave_count < 10 ? .875 : 0.1;
133  } else {
134  erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
135 
136  av_assert2(q->bitrate == I_F_Q);
137 
138  if (q->erasure_count > 1)
139  erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7;
140 
141  for (i = 0; i < 10; i++) {
142  q->predictor_lspf[i] =
143  lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 +
144  erasure_coeff * predictors[i];
145  }
146  smooth = 0.125;
147  }
148 
149  // Check the stability of the LSP frequencies.
150  lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
151  for (i = 1; i < 10; i++)
152  lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR);
153 
154  lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR);
155  for (i = 9; i > 0; i--)
156  lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR);
157 
158  // Low-pass filter the LSP frequencies.
159  ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10);
160  } else {
161  q->octave_count = 0;
162 
163  tmp_lspf = 0.0;
164  for (i = 0; i < 5; i++) {
165  lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
166  lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
167  }
168 
169  // Check for badly received packets.
170  if (q->bitrate == RATE_QUARTER) {
171  if (lspf[9] <= .70 || lspf[9] >= .97)
172  return -1;
173  for (i = 3; i < 10; i++)
174  if (fabs(lspf[i] - lspf[i - 2]) < .08)
175  return -1;
176  } else {
177  if (lspf[9] <= .66 || lspf[9] >= .985)
178  return -1;
179  for (i = 4; i < 10; i++)
180  if (fabs(lspf[i] - lspf[i - 4]) < .0931)
181  return -1;
182  }
183  }
184  return 0;
185 }
186 
187 /**
188  * Convert codebook transmission codes to GAIN and INDEX.
189  *
190  * @param q the context
191  * @param gain array holding the decoded gain
192  *
193  * TIA/EIA/IS-733 2.4.6.2
194  */
195  static void decode_gain_and_index(QCELPContext *q, float *gain)
196 {
197  int i, subframes_count, g1[16];
198  float slope;
199 
200  if (q->bitrate >= RATE_QUARTER) {
201  switch (q->bitrate) {
202  case RATE_FULL: subframes_count = 16; break;
203  case RATE_HALF: subframes_count = 4; break;
204  default: subframes_count = 5;
205  }
206  for (i = 0; i < subframes_count; i++) {
207  g1[i] = 4 * q->frame.cbgain[i];
208  if (q->bitrate == RATE_FULL && !((i + 1) & 3)) {
209  g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32);
210  }
211 
212  gain[i] = qcelp_g12ga[g1[i]];
213 
214  if (q->frame.cbsign[i]) {
215  gain[i] = -gain[i];
216  q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127;
217  }
218  }
219 
220  q->prev_g1[0] = g1[i - 2];
221  q->prev_g1[1] = g1[i - 1];
222  q->last_codebook_gain = qcelp_g12ga[g1[i - 1]];
223 
224  if (q->bitrate == RATE_QUARTER) {
225  // Provide smoothing of the unvoiced excitation energy.
226  gain[7] = gain[4];
227  gain[6] = 0.4 * gain[3] + 0.6 * gain[4];
228  gain[5] = gain[3];
229  gain[4] = 0.8 * gain[2] + 0.2 * gain[3];
230  gain[3] = 0.2 * gain[1] + 0.8 * gain[2];
231  gain[2] = gain[1];
232  gain[1] = 0.6 * gain[0] + 0.4 * gain[1];
233  }
234  } else if (q->bitrate != SILENCE) {
235  if (q->bitrate == RATE_OCTAVE) {
236  g1[0] = 2 * q->frame.cbgain[0] +
237  av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
238  subframes_count = 8;
239  } else {
240  av_assert2(q->bitrate == I_F_Q);
241 
242  g1[0] = q->prev_g1[1];
243  switch (q->erasure_count) {
244  case 1 : break;
245  case 2 : g1[0] -= 1; break;
246  case 3 : g1[0] -= 2; break;
247  default: g1[0] -= 6;
248  }
249  if (g1[0] < 0)
250  g1[0] = 0;
251  subframes_count = 4;
252  }
253  // This interpolation is done to produce smoother background noise.
254  slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
255  for (i = 1; i <= subframes_count; i++)
256  gain[i - 1] = q->last_codebook_gain + slope * i;
257 
258  q->last_codebook_gain = gain[i - 2];
259  q->prev_g1[0] = q->prev_g1[1];
260  q->prev_g1[1] = g1[0];
261  }
262 }
263 
264 /**
265  * If the received packet is Rate 1/4 a further sanity check is made of the
266  * codebook gain.
267  *
268  * @param cbgain the unpacked cbgain array
269  * @return -1 if the sanity check fails, 0 otherwise
270  *
271  * TIA/EIA/IS-733 2.4.8.7.3
272  */
273  static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
274 {
275  int i, diff, prev_diff = 0;
276 
277  for (i = 1; i < 5; i++) {
278  diff = cbgain[i] - cbgain[i-1];
279  if (FFABS(diff) > 10)
280  return -1;
281  else if (FFABS(diff - prev_diff) > 12)
282  return -1;
283  prev_diff = diff;
284  }
285  return 0;
286 }
287 
288 /**
289  * Compute the scaled codebook vector Cdn From INDEX and GAIN
290  * for all rates.
291  *
292  * The specification lacks some information here.
293  *
294  * TIA/EIA/IS-733 has an omission on the codebook index determination
295  * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
296  * you have to subtract the decoded index parameter from the given scaled
297  * codebook vector index 'n' to get the desired circular codebook index, but
298  * it does not mention that you have to clamp 'n' to [0-9] in order to get
299  * RI-compliant results.
300  *
301  * The reason for this mistake seems to be the fact they forgot to mention you
302  * have to do these calculations per codebook subframe and adjust given
303  * equation values accordingly.
304  *
305  * @param q the context
306  * @param gain array holding the 4 pitch subframe gain values
307  * @param cdn_vector array for the generated scaled codebook vector
308  */
309  static void compute_svector(QCELPContext *q, const float *gain,
310  float *cdn_vector)
311 {
312  int i, j, k;
313  uint16_t cbseed, cindex;
314  float *rnd, tmp_gain, fir_filter_value;
315 
316  switch (q->bitrate) {
317  case RATE_FULL:
318  for (i = 0; i < 16; i++) {
319  tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
320  cindex = -q->frame.cindex[i];
321  for (j = 0; j < 10; j++)
322  *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
323  }
324  break;
325  case RATE_HALF:
326  for (i = 0; i < 4; i++) {
327  tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
328  cindex = -q->frame.cindex[i];
329  for (j = 0; j < 40; j++)
330  *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
331  }
332  break;
333  case RATE_QUARTER:
334  cbseed = (0x0003 & q->frame.lspv[4]) << 14 |
335  (0x003F & q->frame.lspv[3]) << 8 |
336  (0x0060 & q->frame.lspv[2]) << 1 |
337  (0x0007 & q->frame.lspv[1]) << 3 |
338  (0x0038 & q->frame.lspv[0]) >> 3;
339  rnd = q->rnd_fir_filter_mem + 20;
340  for (i = 0; i < 8; i++) {
341  tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
342  for (k = 0; k < 20; k++) {
343  cbseed = 521 * cbseed + 259;
344  *rnd = (int16_t) cbseed;
345 
346  // FIR filter
347  fir_filter_value = 0.0;
348  for (j = 0; j < 10; j++)
349  fir_filter_value += qcelp_rnd_fir_coefs[j] *
350  (rnd[-j] + rnd[-20+j]);
351 
352  fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
353  *cdn_vector++ = tmp_gain * fir_filter_value;
354  rnd++;
355  }
356  }
357  memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160,
358  20 * sizeof(float));
359  break;
360  case RATE_OCTAVE:
361  cbseed = q->first16bits;
362  for (i = 0; i < 8; i++) {
363  tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
364  for (j = 0; j < 20; j++) {
365  cbseed = 521 * cbseed + 259;
366  *cdn_vector++ = tmp_gain * (int16_t) cbseed;
367  }
368  }
369  break;
370  case I_F_Q:
371  cbseed = -44; // random codebook index
372  for (i = 0; i < 4; i++) {
373  tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
374  for (j = 0; j < 40; j++)
375  *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
376  }
377  break;
378  case SILENCE:
379  memset(cdn_vector, 0, 160 * sizeof(float));
380  break;
381  }
382 }
383 
384 /**
385  * Apply generic gain control.
386  *
387  * @param v_out output vector
388  * @param v_in gain-controlled vector
389  * @param v_ref vector to control gain of
390  *
391  * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6
392  */
393  static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in)
394 {
395  int i;
396 
397  for (i = 0; i < 160; i += 40) {
398  float res = avpriv_scalarproduct_float_c(v_ref + i, v_ref + i, 40);
399  ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, res, 40);
400  }
401 }
402 
403 /**
404  * Apply filter in pitch-subframe steps.
405  *
406  * @param memory buffer for the previous state of the filter
407  * - must be able to contain 303 elements
408  * - the 143 first elements are from the previous state
409  * - the next 160 are for output
410  * @param v_in input filter vector
411  * @param gain per-subframe gain array, each element is between 0.0 and 2.0
412  * @param lag per-subframe lag array, each element is
413  * - between 16 and 143 if its corresponding pfrac is 0,
414  * - between 16 and 139 otherwise
415  * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
416  * otherwise
417  *
418  * @return filter output vector
419  */
420  static const float *do_pitchfilter(float memory[303], const float v_in[160],
421  const float gain[4], const uint8_t *lag,
422  const uint8_t pfrac[4])
423 {
424  int i, j;
425  float *v_lag, *v_out;
426  const float *v_len;
427 
428  v_out = memory + 143; // Output vector starts at memory[143].
429 
430  for (i = 0; i < 4; i++) {
431  if (gain[i]) {
432  v_lag = memory + 143 + 40 * i - lag[i];
433  for (v_len = v_in + 40; v_in < v_len; v_in++) {
434  if (pfrac[i]) { // If it is a fractional lag...
435  for (j = 0, *v_out = 0.0; j < 4; j++)
436  *v_out += qcelp_hammsinc_table[j] * (v_lag[j - 4] + v_lag[3 - j]);
437  } else
438  *v_out = *v_lag;
439 
440  *v_out = *v_in + gain[i] * *v_out;
441 
442  v_lag++;
443  v_out++;
444  }
445  } else {
446  memcpy(v_out, v_in, 40 * sizeof(float));
447  v_in += 40;
448  v_out += 40;
449  }
450  }
451 
452  memmove(memory, memory + 160, 143 * sizeof(float));
453  return memory + 143;
454 }
455 
456 /**
457  * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
458  * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
459  *
460  * @param q the context
461  * @param cdn_vector the scaled codebook vector
462  */
463  static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
464 {
465  int i;
466  const float *v_synthesis_filtered, *v_pre_filtered;
467 
468  if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE ||
469  (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
470 
471  if (q->bitrate >= RATE_HALF) {
472  // Compute gain & lag for the whole frame.
473  for (i = 0; i < 4; i++) {
474  q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
475 
476  q->pitch_lag[i] = q->frame.plag[i] + 16;
477  }
478  } else {
479  float max_pitch_gain;
480 
481  if (q->bitrate == I_F_Q) {
482  if (q->erasure_count < 3)
483  max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
484  else
485  max_pitch_gain = 0.0;
486  } else {
487  av_assert2(q->bitrate == SILENCE);
488  max_pitch_gain = 1.0;
489  }
490  for (i = 0; i < 4; i++)
491  q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
492 
493  memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
494  }
495 
496  // pitch synthesis filter
497  v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
498  cdn_vector, q->pitch_gain,
499  q->pitch_lag, q->frame.pfrac);
500 
501  // pitch prefilter update
502  for (i = 0; i < 4; i++)
503  q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
504 
505  v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
506  v_synthesis_filtered,
507  q->pitch_gain, q->pitch_lag,
508  q->frame.pfrac);
509 
510  apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
511  } else {
512  memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17, 143 * sizeof(float));
513  memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
514  memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
515  memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
516  }
517 }
518 
519 /**
520  * Reconstruct LPC coefficients from the line spectral pair frequencies
521  * and perform bandwidth expansion.
522  *
523  * @param lspf line spectral pair frequencies
524  * @param lpc linear predictive coding coefficients
525  *
526  * @note: bandwidth_expansion_coeff could be precalculated into a table
527  * but it seems to be slower on x86
528  *
529  * TIA/EIA/IS-733 2.4.3.3.5
530  */
531  static void lspf2lpc(const float *lspf, float *lpc)
532 {
533  double lsp[10];
534  double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
535  int i;
536 
537  for (i = 0; i < 10; i++)
538  lsp[i] = cos(M_PI * lspf[i]);
539 
540  ff_acelp_lspd2lpc(lsp, lpc, 5);
541 
542  for (i = 0; i < 10; i++) {
543  lpc[i] *= bandwidth_expansion_coeff;
544  bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
545  }
546 }
547 
548 /**
549  * Interpolate LSP frequencies and compute LPC coefficients
550  * for a given bitrate & pitch subframe.
551  *
552  * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
553  *
554  * @param q the context
555  * @param curr_lspf LSP frequencies vector of the current frame
556  * @param lpc float vector for the resulting LPC
557  * @param subframe_num frame number in decoded stream
558  */
559  static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
560  float *lpc, const int subframe_num)
561 {
562  float interpolated_lspf[10];
563  float weight;
564 
565  if (q->bitrate >= RATE_QUARTER)
566  weight = 0.25 * (subframe_num + 1);
567  else if (q->bitrate == RATE_OCTAVE && !subframe_num)
568  weight = 0.625;
569  else
570  weight = 1.0;
571 
572  if (weight != 1.0) {
573  ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
574  weight, 1.0 - weight, 10);
575  lspf2lpc(interpolated_lspf, lpc);
576  } else if (q->bitrate >= RATE_QUARTER ||
577  (q->bitrate == I_F_Q && !subframe_num))
578  lspf2lpc(curr_lspf, lpc);
579  else if (q->bitrate == SILENCE && !subframe_num)
580  lspf2lpc(q->prev_lspf, lpc);
581 }
582 
583  static qcelp_packet_rate buf_size2bitrate(const int buf_size)
584 {
585  switch (buf_size) {
586  case 35: return RATE_FULL;
587  case 17: return RATE_HALF;
588  case 8: return RATE_QUARTER;
589  case 4: return RATE_OCTAVE;
590  case 1: return SILENCE;
591  }
592 
593  return I_F_Q;
594 }
595 
596 /**
597  * Determine the bitrate from the frame size and/or the first byte of the frame.
598  *
599  * @param avctx the AV codec context
600  * @param buf_size length of the buffer
601  * @param buf the bufffer
602  *
603  * @return the bitrate on success,
604  * I_F_Q if the bitrate cannot be satisfactorily determined
605  *
606  * TIA/EIA/IS-733 2.4.8.7.1
607  */
608  static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx,
609  const int buf_size,
610  const uint8_t **buf)
611 {
612  qcelp_packet_rate bitrate;
613 
614  if ((bitrate = buf_size2bitrate(buf_size)) >= 0) {
615  if (bitrate > **buf) {
616  QCELPContext *q = avctx->priv_data;
617  if (!q->warned_buf_mismatch_bitrate) {
618  av_log(avctx, AV_LOG_WARNING,
619  "Claimed bitrate and buffer size mismatch.\n");
620  q->warned_buf_mismatch_bitrate = 1;
621  }
622  bitrate = **buf;
623  } else if (bitrate < **buf) {
624  av_log(avctx, AV_LOG_ERROR,
625  "Buffer is too small for the claimed bitrate.\n");
626  return I_F_Q;
627  }
628  (*buf)++;
629  } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) {
630  av_log(avctx, AV_LOG_WARNING,
631  "Bitrate byte is missing, guessing the bitrate from packet size.\n");
632  } else
633  return I_F_Q;
634 
635  if (bitrate == SILENCE) {
636  // FIXME: Remove this warning when tested with samples.
637  avpriv_request_sample(avctx, "Blank frame handling");
638  }
639  return bitrate;
640 }
641 
642  static void warn_insufficient_frame_quality(AVCodecContext *avctx,
643  const char *message)
644 {
645  av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n",
646  avctx->frame_number, message);
647 }
648 
649  static void postfilter(QCELPContext *q, float *samples, float *lpc)
650 {
651  static const float pow_0_775[10] = {
652  0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
653  0.216676, 0.167924, 0.130141, 0.100859, 0.078166
654  }, pow_0_625[10] = {
655  0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
656  0.059605, 0.037253, 0.023283, 0.014552, 0.009095
657  };
658  float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
659  int n;
660 
661  for (n = 0; n < 10; n++) {
662  lpc_s[n] = lpc[n] * pow_0_625[n];
663  lpc_p[n] = lpc[n] * pow_0_775[n];
664  }
665 
666  ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
667  q->formant_mem + 10, 160, 10);
668  memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10);
669  ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
670  memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
671 
672  ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
673 
674  ff_adaptive_gain_control(samples, pole_out + 10,
675  avpriv_scalarproduct_float_c(q->formant_mem + 10,
676  q->formant_mem + 10,
677  160),
678  160, 0.9375, &q->postfilter_agc_mem);
679 }
680 
681  static int qcelp_decode_frame(AVCodecContext *avctx, void *data,
682  int *got_frame_ptr, AVPacket *avpkt)
683 {
684  const uint8_t *buf = avpkt->data;
685  int buf_size = avpkt->size;
686  QCELPContext *q = avctx->priv_data;
687  AVFrame *frame = data;
688  float *outbuffer;
689  int i, ret;
690  float quantized_lspf[10], lpc[10];
691  float gain[16];
692  float *formant_mem;
693 
694  /* get output buffer */
695  frame->nb_samples = 160;
696  if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
697  return ret;
698  outbuffer = (float *)frame->data[0];
699 
700  if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) {
701  warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
702  goto erasure;
703  }
704 
705  if (q->bitrate == RATE_OCTAVE &&
706  (q->first16bits = AV_RB16(buf)) == 0xFFFF) {
707  warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
708  goto erasure;
709  }
710 
711  if (q->bitrate > SILENCE) {
712  const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
713  const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] +
714  qcelp_unpacking_bitmaps_lengths[q->bitrate];
715  uint8_t *unpacked_data = (uint8_t *)&q->frame;
716 
717  init_get_bits(&q->gb, buf, 8 * buf_size);
718 
719  memset(&q->frame, 0, sizeof(QCELPFrame));
720 
721  for (; bitmaps < bitmaps_end; bitmaps++)
722  unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
723 
724  // Check for erasures/blanks on rates 1, 1/4 and 1/8.
725  if (q->frame.reserved) {
726  warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
727  goto erasure;
728  }
729  if (q->bitrate == RATE_QUARTER &&
730  codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) {
731  warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
732  goto erasure;
733  }
734 
735  if (q->bitrate >= RATE_HALF) {
736  for (i = 0; i < 4; i++) {
737  if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) {
738  warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
739  goto erasure;
740  }
741  }
742  }
743  }
744 
745  decode_gain_and_index(q, gain);
746  compute_svector(q, gain, outbuffer);
747 
748  if (decode_lspf(q, quantized_lspf) < 0) {
749  warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
750  goto erasure;
751  }
752 
753  apply_pitch_filters(q, outbuffer);
754 
755  if (q->bitrate == I_F_Q) {
756 erasure:
757  q->bitrate = I_F_Q;
758  q->erasure_count++;
759  decode_gain_and_index(q, gain);
760  compute_svector(q, gain, outbuffer);
761  decode_lspf(q, quantized_lspf);
762  apply_pitch_filters(q, outbuffer);
763  } else
764  q->erasure_count = 0;
765 
766  formant_mem = q->formant_mem + 10;
767  for (i = 0; i < 4; i++) {
768  interpolate_lpc(q, quantized_lspf, lpc, i);
769  ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40, 10);
770  formant_mem += 40;
771  }
772 
773  // postfilter, as per TIA/EIA/IS-733 2.4.8.6
774  postfilter(q, outbuffer, lpc);
775 
776  memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
777 
778  memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
779  q->prev_bitrate = q->bitrate;
780 
781  *got_frame_ptr = 1;
782 
783  return buf_size;
784 }
785 
786  AVCodec ff_qcelp_decoder = {
787  .name = "qcelp",
788  .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
789  .type = AVMEDIA_TYPE_AUDIO,
790  .id = AV_CODEC_ID_QCELP,
791  .init = qcelp_decode_init,
792  .decode = qcelp_decode_frame,
793  .capabilities = CODEC_CAP_DR1,
794  .priv_data_size = sizeof(QCELPContext),
795 };
Generated on Sat Jan 25 2014 19:51:53 for FFmpeg by   doxygen 1.8.2

AltStyle によって変換されたページ (->オリジナル) /