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/channel_layout.h"
33 #include "libavutil/float_dsp.h"
34 #include "avcodec.h"
35 #include "internal.h"
36 #include "get_bits.h"
37 #include "qcelpdata.h"
38 #include "celp_filters.h"
39 #include "acelp_filters.h"
40 #include "acelp_vectors.h"
41 #include "lsp.h"
42 
43 #undef NDEBUG
44 #include <assert.h>
45 
46  typedef enum {
47   I_F_Q = -1, /**< insufficient frame quality */
48   SILENCE,
49   RATE_OCTAVE,
50   RATE_QUARTER,
51   RATE_HALF,
52   RATE_FULL
53 } qcelp_packet_rate;
54 
55  typedef struct {
56   GetBitContext gb;
57   qcelp_packet_rate bitrate;
58   QCELPFrame frame; /**< unpacked data frame */
59 
60   uint8_t erasure_count;
61   uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */
62   float prev_lspf[10];
63   float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */
64   float pitch_synthesis_filter_mem[303];
65   float pitch_pre_filter_mem[303];
66   float rnd_fir_filter_mem[180];
67   float formant_mem[170];
68   float last_codebook_gain;
69   int prev_g1[2];
70   int prev_bitrate;
71   float pitch_gain[4];
72   uint8_t pitch_lag[4];
73   uint16_t first16bits;
74   uint8_t warned_buf_mismatch_bitrate;
75 
76  /* postfilter */
77   float postfilter_synth_mem[10];
78   float postfilter_agc_mem;
79   float postfilter_tilt_mem;
80 } QCELPContext;
81 
82 /**
83  * Initialize the speech codec according to the specification.
84  *
85  * TIA/EIA/IS-733 2.4.9
86  */
87  static av_cold int qcelp_decode_init(AVCodecContext *avctx)
88 {
89  QCELPContext *q = avctx->priv_data;
90  int i;
91 
92  avctx->channels = 1;
93  avctx->channel_layout = AV_CH_LAYOUT_MONO;
94  avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
95 
96  for (i = 0; i < 10; i++)
97  q->prev_lspf[i] = (i + 1) / 11.;
98 
99  return 0;
100 }
101 
102 /**
103  * Decode the 10 quantized LSP frequencies from the LSPV/LSP
104  * transmission codes of any bitrate and check for badly received packets.
105  *
106  * @param q the context
107  * @param lspf line spectral pair frequencies
108  *
109  * @return 0 on success, -1 if the packet is badly received
110  *
111  * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
112  */
113  static int decode_lspf(QCELPContext *q, float *lspf)
114 {
115  int i;
116  float tmp_lspf, smooth, erasure_coeff;
117  const float *predictors;
118 
119  if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) {
120  predictors = q->prev_bitrate != RATE_OCTAVE &&
121  q->prev_bitrate != I_F_Q ? q->prev_lspf
122  : q->predictor_lspf;
123 
124  if (q->bitrate == RATE_OCTAVE) {
125  q->octave_count++;
126 
127  for (i = 0; i < 10; i++) {
128  q->predictor_lspf[i] =
129  lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
130  : -QCELP_LSP_SPREAD_FACTOR) +
131  predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR +
132  (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11);
133  }
134  smooth = q->octave_count < 10 ? .875 : 0.1;
135  } else {
136  erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
137 
138  assert(q->bitrate == I_F_Q);
139 
140  if (q->erasure_count > 1)
141  erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7;
142 
143  for (i = 0; i < 10; i++) {
144  q->predictor_lspf[i] =
145  lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 +
146  erasure_coeff * predictors[i];
147  }
148  smooth = 0.125;
149  }
150 
151  // Check the stability of the LSP frequencies.
152  lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
153  for (i = 1; i < 10; i++)
154  lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR);
155 
156  lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR);
157  for (i = 9; i > 0; i--)
158  lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR);
159 
160  // Low-pass filter the LSP frequencies.
161  ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10);
162  } else {
163  q->octave_count = 0;
164 
165  tmp_lspf = 0.;
166  for (i = 0; i < 5; i++) {
167  lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
168  lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
169  }
170 
171  // Check for badly received packets.
172  if (q->bitrate == RATE_QUARTER) {
173  if (lspf[9] <= .70 || lspf[9] >= .97)
174  return -1;
175  for (i = 3; i < 10; i++)
176  if (fabs(lspf[i] - lspf[i - 2]) < .08)
177  return -1;
178  } else {
179  if (lspf[9] <= .66 || lspf[9] >= .985)
180  return -1;
181  for (i = 4; i < 10; i++)
182  if (fabs(lspf[i] - lspf[i - 4]) < .0931)
183  return -1;
184  }
185  }
186  return 0;
187 }
188 
189 /**
190  * Convert codebook transmission codes to GAIN and INDEX.
191  *
192  * @param q the context
193  * @param gain array holding the decoded gain
194  *
195  * TIA/EIA/IS-733 2.4.6.2
196  */
197  static void decode_gain_and_index(QCELPContext *q, float *gain)
198 {
199  int i, subframes_count, g1[16];
200  float slope;
201 
202  if (q->bitrate >= RATE_QUARTER) {
203  switch (q->bitrate) {
204  case RATE_FULL: subframes_count = 16; break;
205  case RATE_HALF: subframes_count = 4; break;
206  default: subframes_count = 5;
207  }
208  for (i = 0; i < subframes_count; i++) {
209  g1[i] = 4 * q->frame.cbgain[i];
210  if (q->bitrate == RATE_FULL && !((i + 1) & 3)) {
211  g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32);
212  }
213 
214  gain[i] = qcelp_g12ga[g1[i]];
215 
216  if (q->frame.cbsign[i]) {
217  gain[i] = -gain[i];
218  q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127;
219  }
220  }
221 
222  q->prev_g1[0] = g1[i - 2];
223  q->prev_g1[1] = g1[i - 1];
224  q->last_codebook_gain = qcelp_g12ga[g1[i - 1]];
225 
226  if (q->bitrate == RATE_QUARTER) {
227  // Provide smoothing of the unvoiced excitation energy.
228  gain[7] = gain[4];
229  gain[6] = 0.4 * gain[3] + 0.6 * gain[4];
230  gain[5] = gain[3];
231  gain[4] = 0.8 * gain[2] + 0.2 * gain[3];
232  gain[3] = 0.2 * gain[1] + 0.8 * gain[2];
233  gain[2] = gain[1];
234  gain[1] = 0.6 * gain[0] + 0.4 * gain[1];
235  }
236  } else if (q->bitrate != SILENCE) {
237  if (q->bitrate == RATE_OCTAVE) {
238  g1[0] = 2 * q->frame.cbgain[0] +
239  av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
240  subframes_count = 8;
241  } else {
242  assert(q->bitrate == I_F_Q);
243 
244  g1[0] = q->prev_g1[1];
245  switch (q->erasure_count) {
246  case 1 : break;
247  case 2 : g1[0] -= 1; break;
248  case 3 : g1[0] -= 2; break;
249  default: g1[0] -= 6;
250  }
251  if (g1[0] < 0)
252  g1[0] = 0;
253  subframes_count = 4;
254  }
255  // This interpolation is done to produce smoother background noise.
256  slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
257  for (i = 1; i <= subframes_count; i++)
258  gain[i - 1] = q->last_codebook_gain + slope * i;
259 
260  q->last_codebook_gain = gain[i - 2];
261  q->prev_g1[0] = q->prev_g1[1];
262  q->prev_g1[1] = g1[0];
263  }
264 }
265 
266 /**
267  * If the received packet is Rate 1/4 a further sanity check is made of the
268  * codebook gain.
269  *
270  * @param cbgain the unpacked cbgain array
271  * @return -1 if the sanity check fails, 0 otherwise
272  *
273  * TIA/EIA/IS-733 2.4.8.7.3
274  */
275  static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
276 {
277  int i, diff, prev_diff = 0;
278 
279  for (i = 1; i < 5; i++) {
280  diff = cbgain[i] - cbgain[i-1];
281  if (FFABS(diff) > 10)
282  return -1;
283  else if (FFABS(diff - prev_diff) > 12)
284  return -1;
285  prev_diff = diff;
286  }
287  return 0;
288 }
289 
290 /**
291  * Compute the scaled codebook vector Cdn From INDEX and GAIN
292  * for all rates.
293  *
294  * The specification lacks some information here.
295  *
296  * TIA/EIA/IS-733 has an omission on the codebook index determination
297  * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
298  * you have to subtract the decoded index parameter from the given scaled
299  * codebook vector index 'n' to get the desired circular codebook index, but
300  * it does not mention that you have to clamp 'n' to [0-9] in order to get
301  * RI-compliant results.
302  *
303  * The reason for this mistake seems to be the fact they forgot to mention you
304  * have to do these calculations per codebook subframe and adjust given
305  * equation values accordingly.
306  *
307  * @param q the context
308  * @param gain array holding the 4 pitch subframe gain values
309  * @param cdn_vector array for the generated scaled codebook vector
310  */
311  static void compute_svector(QCELPContext *q, const float *gain,
312  float *cdn_vector)
313 {
314  int i, j, k;
315  uint16_t cbseed, cindex;
316  float *rnd, tmp_gain, fir_filter_value;
317 
318  switch (q->bitrate) {
319  case RATE_FULL:
320  for (i = 0; i < 16; i++) {
321  tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
322  cindex = -q->frame.cindex[i];
323  for (j = 0; j < 10; j++)
324  *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
325  }
326  break;
327  case RATE_HALF:
328  for (i = 0; i < 4; i++) {
329  tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
330  cindex = -q->frame.cindex[i];
331  for (j = 0; j < 40; j++)
332  *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
333  }
334  break;
335  case RATE_QUARTER:
336  cbseed = (0x0003 & q->frame.lspv[4]) << 14 |
337  (0x003F & q->frame.lspv[3]) << 8 |
338  (0x0060 & q->frame.lspv[2]) << 1 |
339  (0x0007 & q->frame.lspv[1]) << 3 |
340  (0x0038 & q->frame.lspv[0]) >> 3;
341  rnd = q->rnd_fir_filter_mem + 20;
342  for (i = 0; i < 8; i++) {
343  tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
344  for (k = 0; k < 20; k++) {
345  cbseed = 521 * cbseed + 259;
346  *rnd = (int16_t) cbseed;
347 
348  // FIR filter
349  fir_filter_value = 0.0;
350  for (j = 0; j < 10; j++)
351  fir_filter_value += qcelp_rnd_fir_coefs[j] *
352  (rnd[-j] + rnd[-20+j]);
353 
354  fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
355  *cdn_vector++ = tmp_gain * fir_filter_value;
356  rnd++;
357  }
358  }
359  memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160,
360  20 * sizeof(float));
361  break;
362  case RATE_OCTAVE:
363  cbseed = q->first16bits;
364  for (i = 0; i < 8; i++) {
365  tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
366  for (j = 0; j < 20; j++) {
367  cbseed = 521 * cbseed + 259;
368  *cdn_vector++ = tmp_gain * (int16_t) cbseed;
369  }
370  }
371  break;
372  case I_F_Q:
373  cbseed = -44; // random codebook index
374  for (i = 0; i < 4; i++) {
375  tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
376  for (j = 0; j < 40; j++)
377  *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
378  }
379  break;
380  case SILENCE:
381  memset(cdn_vector, 0, 160 * sizeof(float));
382  break;
383  }
384 }
385 
386 /**
387  * Apply generic gain control.
388  *
389  * @param v_out output vector
390  * @param v_in gain-controlled vector
391  * @param v_ref vector to control gain of
392  *
393  * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6
394  */
395  static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in)
396 {
397  int i;
398 
399  for (i = 0; i < 160; i += 40) {
400  float res = avpriv_scalarproduct_float_c(v_ref + i, v_ref + i, 40);
401  ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, res, 40);
402  }
403 }
404 
405 /**
406  * Apply filter in pitch-subframe steps.
407  *
408  * @param memory buffer for the previous state of the filter
409  * - must be able to contain 303 elements
410  * - the 143 first elements are from the previous state
411  * - the next 160 are for output
412  * @param v_in input filter vector
413  * @param gain per-subframe gain array, each element is between 0.0 and 2.0
414  * @param lag per-subframe lag array, each element is
415  * - between 16 and 143 if its corresponding pfrac is 0,
416  * - between 16 and 139 otherwise
417  * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
418  * otherwise
419  *
420  * @return filter output vector
421  */
422  static const float *do_pitchfilter(float memory[303], const float v_in[160],
423  const float gain[4], const uint8_t *lag,
424  const uint8_t pfrac[4])
425 {
426  int i, j;
427  float *v_lag, *v_out;
428  const float *v_len;
429 
430  v_out = memory + 143; // Output vector starts at memory[143].
431 
432  for (i = 0; i < 4; i++) {
433  if (gain[i]) {
434  v_lag = memory + 143 + 40 * i - lag[i];
435  for (v_len = v_in + 40; v_in < v_len; v_in++) {
436  if (pfrac[i]) { // If it is a fractional lag...
437  for (j = 0, *v_out = 0.; j < 4; j++)
438  *v_out += qcelp_hammsinc_table[j] * (v_lag[j - 4] + v_lag[3 - j]);
439  } else
440  *v_out = *v_lag;
441 
442  *v_out = *v_in + gain[i] * *v_out;
443 
444  v_lag++;
445  v_out++;
446  }
447  } else {
448  memcpy(v_out, v_in, 40 * sizeof(float));
449  v_in += 40;
450  v_out += 40;
451  }
452  }
453 
454  memmove(memory, memory + 160, 143 * sizeof(float));
455  return memory + 143;
456 }
457 
458 /**
459  * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
460  * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
461  *
462  * @param q the context
463  * @param cdn_vector the scaled codebook vector
464  */
465  static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
466 {
467  int i;
468  const float *v_synthesis_filtered, *v_pre_filtered;
469 
470  if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE ||
471  (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
472 
473  if (q->bitrate >= RATE_HALF) {
474  // Compute gain & lag for the whole frame.
475  for (i = 0; i < 4; i++) {
476  q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
477 
478  q->pitch_lag[i] = q->frame.plag[i] + 16;
479  }
480  } else {
481  float max_pitch_gain;
482 
483  if (q->bitrate == I_F_Q) {
484  if (q->erasure_count < 3)
485  max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
486  else
487  max_pitch_gain = 0.0;
488  } else {
489  assert(q->bitrate == SILENCE);
490  max_pitch_gain = 1.0;
491  }
492  for (i = 0; i < 4; i++)
493  q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
494 
495  memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
496  }
497 
498  // pitch synthesis filter
499  v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
500  cdn_vector, q->pitch_gain,
501  q->pitch_lag, q->frame.pfrac);
502 
503  // pitch prefilter update
504  for (i = 0; i < 4; i++)
505  q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
506 
507  v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
508  v_synthesis_filtered,
509  q->pitch_gain, q->pitch_lag,
510  q->frame.pfrac);
511 
512  apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
513  } else {
514  memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17, 143 * sizeof(float));
515  memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
516  memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
517  memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
518  }
519 }
520 
521 /**
522  * Reconstruct LPC coefficients from the line spectral pair frequencies
523  * and perform bandwidth expansion.
524  *
525  * @param lspf line spectral pair frequencies
526  * @param lpc linear predictive coding coefficients
527  *
528  * @note: bandwidth_expansion_coeff could be precalculated into a table
529  * but it seems to be slower on x86
530  *
531  * TIA/EIA/IS-733 2.4.3.3.5
532  */
533  static void lspf2lpc(const float *lspf, float *lpc)
534 {
535  double lsp[10];
536  double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
537  int i;
538 
539  for (i = 0; i < 10; i++)
540  lsp[i] = cos(M_PI * lspf[i]);
541 
542  ff_acelp_lspd2lpc(lsp, lpc, 5);
543 
544  for (i = 0; i < 10; i++) {
545  lpc[i] *= bandwidth_expansion_coeff;
546  bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
547  }
548 }
549 
550 /**
551  * Interpolate LSP frequencies and compute LPC coefficients
552  * for a given bitrate & pitch subframe.
553  *
554  * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
555  *
556  * @param q the context
557  * @param curr_lspf LSP frequencies vector of the current frame
558  * @param lpc float vector for the resulting LPC
559  * @param subframe_num frame number in decoded stream
560  */
561  static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
562  float *lpc, const int subframe_num)
563 {
564  float interpolated_lspf[10];
565  float weight;
566 
567  if (q->bitrate >= RATE_QUARTER)
568  weight = 0.25 * (subframe_num + 1);
569  else if (q->bitrate == RATE_OCTAVE && !subframe_num)
570  weight = 0.625;
571  else
572  weight = 1.0;
573 
574  if (weight != 1.0) {
575  ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
576  weight, 1.0 - weight, 10);
577  lspf2lpc(interpolated_lspf, lpc);
578  } else if (q->bitrate >= RATE_QUARTER ||
579  (q->bitrate == I_F_Q && !subframe_num))
580  lspf2lpc(curr_lspf, lpc);
581  else if (q->bitrate == SILENCE && !subframe_num)
582  lspf2lpc(q->prev_lspf, lpc);
583 }
584 
585  static qcelp_packet_rate buf_size2bitrate(const int buf_size)
586 {
587  switch (buf_size) {
588  case 35: return RATE_FULL;
589  case 17: return RATE_HALF;
590  case 8: return RATE_QUARTER;
591  case 4: return RATE_OCTAVE;
592  case 1: return SILENCE;
593  }
594 
595  return I_F_Q;
596 }
597 
598 /**
599  * Determine the bitrate from the frame size and/or the first byte of the frame.
600  *
601  * @param avctx the AV codec context
602  * @param buf_size length of the buffer
603  * @param buf the bufffer
604  *
605  * @return the bitrate on success,
606  * I_F_Q if the bitrate cannot be satisfactorily determined
607  *
608  * TIA/EIA/IS-733 2.4.8.7.1
609  */
610  static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx,
611  const int buf_size,
612  const uint8_t **buf)
613 {
614  qcelp_packet_rate bitrate;
615 
616  if ((bitrate = buf_size2bitrate(buf_size)) >= 0) {
617  if (bitrate > **buf) {
618  QCELPContext *q = avctx->priv_data;
619  if (!q->warned_buf_mismatch_bitrate) {
620  av_log(avctx, AV_LOG_WARNING,
621  "Claimed bitrate and buffer size mismatch.\n");
622  q->warned_buf_mismatch_bitrate = 1;
623  }
624  bitrate = **buf;
625  } else if (bitrate < **buf) {
626  av_log(avctx, AV_LOG_ERROR,
627  "Buffer is too small for the claimed bitrate.\n");
628  return I_F_Q;
629  }
630  (*buf)++;
631  } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) {
632  av_log(avctx, AV_LOG_WARNING,
633  "Bitrate byte is missing, guessing the bitrate from packet size.\n");
634  } else
635  return I_F_Q;
636 
637  if (bitrate == SILENCE) {
638  // FIXME: Remove this warning when tested with samples.
639  avpriv_request_sample(avctx, "Blank frame handling");
640  }
641  return bitrate;
642 }
643 
644  static void warn_insufficient_frame_quality(AVCodecContext *avctx,
645  const char *message)
646 {
647  av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n",
648  avctx->frame_number, message);
649 }
650 
651  static void postfilter(QCELPContext *q, float *samples, float *lpc)
652 {
653  static const float pow_0_775[10] = {
654  0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
655  0.216676, 0.167924, 0.130141, 0.100859, 0.078166
656  }, pow_0_625[10] = {
657  0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
658  0.059605, 0.037253, 0.023283, 0.014552, 0.009095
659  };
660  float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
661  int n;
662 
663  for (n = 0; n < 10; n++) {
664  lpc_s[n] = lpc[n] * pow_0_625[n];
665  lpc_p[n] = lpc[n] * pow_0_775[n];
666  }
667 
668  ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
669  q->formant_mem + 10, 160, 10);
670  memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10);
671  ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
672  memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
673 
674  ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
675 
676  ff_adaptive_gain_control(samples, pole_out + 10,
677  avpriv_scalarproduct_float_c(q->formant_mem + 10,
678  q->formant_mem + 10,
679  160),
680  160, 0.9375, &q->postfilter_agc_mem);
681 }
682 
683  static int qcelp_decode_frame(AVCodecContext *avctx, void *data,
684  int *got_frame_ptr, AVPacket *avpkt)
685 {
686  const uint8_t *buf = avpkt->data;
687  int buf_size = avpkt->size;
688  QCELPContext *q = avctx->priv_data;
689  AVFrame *frame = data;
690  float *outbuffer;
691  int i, ret;
692  float quantized_lspf[10], lpc[10];
693  float gain[16];
694  float *formant_mem;
695 
696  /* get output buffer */
697  frame->nb_samples = 160;
698  if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
699  return ret;
700  outbuffer = (float *)frame->data[0];
701 
702  if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) {
703  warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
704  goto erasure;
705  }
706 
707  if (q->bitrate == RATE_OCTAVE &&
708  (q->first16bits = AV_RB16(buf)) == 0xFFFF) {
709  warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
710  goto erasure;
711  }
712 
713  if (q->bitrate > SILENCE) {
714  const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
715  const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] +
716  qcelp_unpacking_bitmaps_lengths[q->bitrate];
717  uint8_t *unpacked_data = (uint8_t *)&q->frame;
718 
719  init_get_bits(&q->gb, buf, 8 * buf_size);
720 
721  memset(&q->frame, 0, sizeof(QCELPFrame));
722 
723  for (; bitmaps < bitmaps_end; bitmaps++)
724  unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
725 
726  // Check for erasures/blanks on rates 1, 1/4 and 1/8.
727  if (q->frame.reserved) {
728  warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
729  goto erasure;
730  }
731  if (q->bitrate == RATE_QUARTER &&
732  codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) {
733  warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
734  goto erasure;
735  }
736 
737  if (q->bitrate >= RATE_HALF) {
738  for (i = 0; i < 4; i++) {
739  if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) {
740  warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
741  goto erasure;
742  }
743  }
744  }
745  }
746 
747  decode_gain_and_index(q, gain);
748  compute_svector(q, gain, outbuffer);
749 
750  if (decode_lspf(q, quantized_lspf) < 0) {
751  warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
752  goto erasure;
753  }
754 
755  apply_pitch_filters(q, outbuffer);
756 
757  if (q->bitrate == I_F_Q) {
758 erasure:
759  q->bitrate = I_F_Q;
760  q->erasure_count++;
761  decode_gain_and_index(q, gain);
762  compute_svector(q, gain, outbuffer);
763  decode_lspf(q, quantized_lspf);
764  apply_pitch_filters(q, outbuffer);
765  } else
766  q->erasure_count = 0;
767 
768  formant_mem = q->formant_mem + 10;
769  for (i = 0; i < 4; i++) {
770  interpolate_lpc(q, quantized_lspf, lpc, i);
771  ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40, 10);
772  formant_mem += 40;
773  }
774 
775  // postfilter, as per TIA/EIA/IS-733 2.4.8.6
776  postfilter(q, outbuffer, lpc);
777 
778  memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
779 
780  memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
781  q->prev_bitrate = q->bitrate;
782 
783  *got_frame_ptr = 1;
784 
785  return buf_size;
786 }
787 
788  AVCodec ff_qcelp_decoder = {
789  .name = "qcelp",
790  .type = AVMEDIA_TYPE_AUDIO,
791  .id = AV_CODEC_ID_QCELP,
792  .init = qcelp_decode_init,
793  .decode = qcelp_decode_frame,
794  .capabilities = CODEC_CAP_DR1,
795  .priv_data_size = sizeof(QCELPContext),
796  .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
797 };
Generated on Wed Jul 10 2013 23:48:03 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) /