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 "libavutil/avassert.h"
32 #include "libavutil/float_dsp.h"
33 #include "avcodec.h"
34 #include "codec_internal.h"
35 #include "decode.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 typedef enum {
44  I_F_Q = -1, /**< insufficient frame quality */
51 
52 typedef struct QCELPContext {
55  QCELPFrame frame; /**< unpacked data frame */
56 
57  uint8_t erasure_count;
58  uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */
59  float prev_lspf[10];
60  float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */
63  float rnd_fir_filter_mem[180];
64  float formant_mem[170];
66  int prev_g1[2];
68  float pitch_gain[4];
69  uint8_t pitch_lag[4];
70  uint16_t first16bits;
72 
73  /* postfilter */
77 } QCELPContext;
78 
79 /**
80  * Initialize the speech codec according to the specification.
81  *
82  * TIA/EIA/IS-733 2.4.9
83  */
85 {
86  QCELPContext *q = avctx->priv_data;
87  int i;
88 
92 
93  for (i = 0; i < 10; i++)
94  q->prev_lspf[i] = (i + 1) / 11.0;
95 
96  return 0;
97 }
98 
99 /**
100  * Decode the 10 quantized LSP frequencies from the LSPV/LSP
101  * transmission codes of any bitrate and check for badly received packets.
102  *
103  * @param q the context
104  * @param lspf line spectral pair frequencies
105  *
106  * @return 0 on success, -1 if the packet is badly received
107  *
108  * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
109  */
110 static int decode_lspf(QCELPContext *q, float *lspf)
111 {
112  int i;
113  float tmp_lspf, smooth, erasure_coeff;
114  const float *predictors;
115 
116  if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) {
117  predictors = q->prev_bitrate != RATE_OCTAVE &&
118  q->prev_bitrate != I_F_Q ? q->prev_lspf
119  : q->predictor_lspf;
120 
121  if (q->bitrate == RATE_OCTAVE) {
122  q->octave_count++;
123 
124  for (i = 0; i < 10; i++) {
125  q->predictor_lspf[i] =
126  lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
128  predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR +
129  (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11);
130  }
131  smooth = q->octave_count < 10 ? .875 : 0.1;
132  } else {
133  erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
134 
135  av_assert2(q->bitrate == I_F_Q);
136 
137  if (q->erasure_count > 1)
138  erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7;
139 
140  for (i = 0; i < 10; i++) {
141  q->predictor_lspf[i] =
142  lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 +
143  erasure_coeff * predictors[i];
144  }
145  smooth = 0.125;
146  }
147 
148  // Check the stability of the LSP frequencies.
149  lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
150  for (i = 1; i < 10; i++)
151  lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR);
152 
153  lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR);
154  for (i = 9; i > 0; i--)
155  lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR);
156 
157  // Low-pass filter the LSP frequencies.
158  ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10);
159  } else {
160  q->octave_count = 0;
161 
162  tmp_lspf = 0.0;
163  for (i = 0; i < 5; i++) {
164  lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
165  lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
166  }
167 
168  // Check for badly received packets.
169  if (q->bitrate == RATE_QUARTER) {
170  if (lspf[9] <= .70 || lspf[9] >= .97)
171  return -1;
172  for (i = 3; i < 10; i++)
173  if (fabs(lspf[i] - lspf[i - 2]) < .08)
174  return -1;
175  } else {
176  if (lspf[9] <= .66 || lspf[9] >= .985)
177  return -1;
178  for (i = 4; i < 10; i++)
179  if (fabs(lspf[i] - lspf[i - 4]) < .0931)
180  return -1;
181  }
182  }
183  return 0;
184 }
185 
186 /**
187  * Convert codebook transmission codes to GAIN and INDEX.
188  *
189  * @param q the context
190  * @param gain array holding the decoded gain
191  *
192  * TIA/EIA/IS-733 2.4.6.2
193  */
194 static void decode_gain_and_index(QCELPContext *q, float *gain)
195 {
196  int i, subframes_count, g1[16];
197  float slope;
198 
199  if (q->bitrate >= RATE_QUARTER) {
200  switch (q->bitrate) {
201  case RATE_FULL: subframes_count = 16; break;
202  case RATE_HALF: subframes_count = 4; break;
203  default: subframes_count = 5;
204  }
205  for (i = 0; i < subframes_count; i++) {
206  g1[i] = 4 * q->frame.cbgain[i];
207  if (q->bitrate == RATE_FULL && !((i + 1) & 3)) {
208  g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32);
209  }
210 
211  gain[i] = qcelp_g12ga[g1[i]];
212 
213  if (q->frame.cbsign[i]) {
214  gain[i] = -gain[i];
215  q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127;
216  }
217  }
218 
219  q->prev_g1[0] = g1[i - 2];
220  q->prev_g1[1] = g1[i - 1];
221  q->last_codebook_gain = qcelp_g12ga[g1[i - 1]];
222 
223  if (q->bitrate == RATE_QUARTER) {
224  // Provide smoothing of the unvoiced excitation energy.
225  gain[7] = gain[4];
226  gain[6] = 0.4 * gain[3] + 0.6 * gain[4];
227  gain[5] = gain[3];
228  gain[4] = 0.8 * gain[2] + 0.2 * gain[3];
229  gain[3] = 0.2 * gain[1] + 0.8 * gain[2];
230  gain[2] = gain[1];
231  gain[1] = 0.6 * gain[0] + 0.4 * gain[1];
232  }
233  } else if (q->bitrate != SILENCE) {
234  if (q->bitrate == RATE_OCTAVE) {
235  g1[0] = 2 * q->frame.cbgain[0] +
236  av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
237  subframes_count = 8;
238  } else {
239  av_assert2(q->bitrate == I_F_Q);
240 
241  g1[0] = q->prev_g1[1];
242  switch (q->erasure_count) {
243  case 1 : break;
244  case 2 : g1[0] -= 1; break;
245  case 3 : g1[0] -= 2; break;
246  default: g1[0] -= 6;
247  }
248  if (g1[0] < 0)
249  g1[0] = 0;
250  subframes_count = 4;
251  }
252  // This interpolation is done to produce smoother background noise.
253  slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
254  for (i = 1; i <= subframes_count; i++)
255  gain[i - 1] = q->last_codebook_gain + slope * i;
256 
257  q->last_codebook_gain = gain[i - 2];
258  q->prev_g1[0] = q->prev_g1[1];
259  q->prev_g1[1] = g1[0];
260  }
261 }
262 
263 /**
264  * If the received packet is Rate 1/4 a further sanity check is made of the
265  * codebook gain.
266  *
267  * @param cbgain the unpacked cbgain array
268  * @return -1 if the sanity check fails, 0 otherwise
269  *
270  * TIA/EIA/IS-733 2.4.8.7.3
271  */
272 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
273 {
274  int i, diff, prev_diff = 0;
275 
276  for (i = 1; i < 5; i++) {
277  diff = cbgain[i] - cbgain[i-1];
278  if (FFABS(diff) > 10)
279  return -1;
280  else if (FFABS(diff - prev_diff) > 12)
281  return -1;
282  prev_diff = diff;
283  }
284  return 0;
285 }
286 
287 /**
288  * Compute the scaled codebook vector Cdn From INDEX and GAIN
289  * for all rates.
290  *
291  * The specification lacks some information here.
292  *
293  * TIA/EIA/IS-733 has an omission on the codebook index determination
294  * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
295  * you have to subtract the decoded index parameter from the given scaled
296  * codebook vector index 'n' to get the desired circular codebook index, but
297  * it does not mention that you have to clamp 'n' to [0-9] in order to get
298  * RI-compliant results.
299  *
300  * The reason for this mistake seems to be the fact they forgot to mention you
301  * have to do these calculations per codebook subframe and adjust given
302  * equation values accordingly.
303  *
304  * @param q the context
305  * @param gain array holding the 4 pitch subframe gain values
306  * @param cdn_vector array for the generated scaled codebook vector
307  */
308 static void compute_svector(QCELPContext *q, const float *gain,
309  float *cdn_vector)
310 {
311  int i, j, k;
312  uint16_t cbseed, cindex;
313  float *rnd, tmp_gain, fir_filter_value;
314 
315  switch (q->bitrate) {
316  case RATE_FULL:
317  for (i = 0; i < 16; i++) {
318  tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
319  cindex = -q->frame.cindex[i];
320  for (j = 0; j < 10; j++)
321  *cdn_vector++ = tmp_gain *
322  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 *
331  qcelp_rate_half_codebook[cindex++ & 127];
332  }
333  break;
334  case RATE_QUARTER:
335  cbseed = (0x0003 & q->frame.lspv[4]) << 14 |
336  (0x003F & q->frame.lspv[3]) << 8 |
337  (0x0060 & q->frame.lspv[2]) << 1 |
338  (0x0007 & q->frame.lspv[1]) << 3 |
339  (0x0038 & q->frame.lspv[0]) >> 3;
340  rnd = q->rnd_fir_filter_mem + 20;
341  for (i = 0; i < 8; i++) {
342  tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
343  for (k = 0; k < 20; k++) {
344  cbseed = 521 * cbseed + 259;
345  *rnd = (int16_t) cbseed;
346 
347  // FIR filter
348  fir_filter_value = 0.0;
349  for (j = 0; j < 10; j++)
350  fir_filter_value += qcelp_rnd_fir_coefs[j] *
351  (rnd[-j] + rnd[-20+j]);
352 
353  fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
354  *cdn_vector++ = tmp_gain * fir_filter_value;
355  rnd++;
356  }
357  }
358  memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160,
359  20 * sizeof(float));
360  break;
361  case RATE_OCTAVE:
362  cbseed = q->first16bits;
363  for (i = 0; i < 8; i++) {
364  tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
365  for (j = 0; j < 20; j++) {
366  cbseed = 521 * cbseed + 259;
367  *cdn_vector++ = tmp_gain * (int16_t) cbseed;
368  }
369  }
370  break;
371  case I_F_Q:
372  cbseed = -44; // random codebook index
373  for (i = 0; i < 4; i++) {
374  tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
375  for (j = 0; j < 40; j++)
376  *cdn_vector++ = tmp_gain *
377  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 = 0.0; j < 4; j++)
438  *v_out += qcelp_hammsinc_table[j] *
439  (v_lag[j - 4] + v_lag[3 - j]);
440  } else
441  *v_out = *v_lag;
442 
443  *v_out = *v_in + gain[i] * *v_out;
444 
445  v_lag++;
446  v_out++;
447  }
448  } else {
449  memcpy(v_out, v_in, 40 * sizeof(float));
450  v_in += 40;
451  v_out += 40;
452  }
453  }
454 
455  memmove(memory, memory + 160, 143 * sizeof(float));
456  return memory + 143;
457 }
458 
459 /**
460  * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
461  * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
462  *
463  * @param q the context
464  * @param cdn_vector the scaled codebook vector
465  */
466 static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
467 {
468  int i;
469  const float *v_synthesis_filtered, *v_pre_filtered;
470 
471  if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE ||
472  (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
473 
474  if (q->bitrate >= RATE_HALF) {
475  // Compute gain & lag for the whole frame.
476  for (i = 0; i < 4; i++) {
477  q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
478 
479  q->pitch_lag[i] = q->frame.plag[i] + 16;
480  }
481  } else {
482  float max_pitch_gain;
483 
484  if (q->bitrate == I_F_Q) {
485  if (q->erasure_count < 3)
486  max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
487  else
488  max_pitch_gain = 0.0;
489  } else {
490  av_assert2(q->bitrate == SILENCE);
491  max_pitch_gain = 1.0;
492  }
493  for (i = 0; i < 4; i++)
494  q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
495 
496  memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
497  }
498 
499  // pitch synthesis filter
500  v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
501  cdn_vector, q->pitch_gain,
502  q->pitch_lag, q->frame.pfrac);
503 
504  // pitch prefilter update
505  for (i = 0; i < 4; i++)
506  q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
507 
508  v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
509  v_synthesis_filtered,
510  q->pitch_gain, q->pitch_lag,
511  q->frame.pfrac);
512 
513  apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
514  } else {
515  memcpy(q->pitch_synthesis_filter_mem,
516  cdn_vector + 17, 143 * sizeof(float));
517  memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
518  memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
519  memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
520  }
521 }
522 
523 /**
524  * Reconstruct LPC coefficients from the line spectral pair frequencies
525  * and perform bandwidth expansion.
526  *
527  * @param lspf line spectral pair frequencies
528  * @param lpc linear predictive coding coefficients
529  *
530  * @note: bandwidth_expansion_coeff could be precalculated into a table
531  * but it seems to be slower on x86
532  *
533  * TIA/EIA/IS-733 2.4.3.3.5
534  */
535 static void lspf2lpc(const float *lspf, float *lpc)
536 {
537  double lsp[10];
538  double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
539  int i;
540 
541  for (i = 0; i < 10; i++)
542  lsp[i] = cos(M_PI * lspf[i]);
543 
544  ff_acelp_lspd2lpc(lsp, lpc, 5);
545 
546  for (i = 0; i < 10; i++) {
547  lpc[i] *= bandwidth_expansion_coeff;
548  bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
549  }
550 }
551 
552 /**
553  * Interpolate LSP frequencies and compute LPC coefficients
554  * for a given bitrate & pitch subframe.
555  *
556  * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
557  *
558  * @param q the context
559  * @param curr_lspf LSP frequencies vector of the current frame
560  * @param lpc float vector for the resulting LPC
561  * @param subframe_num frame number in decoded stream
562  */
563 static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
564  float *lpc, const int subframe_num)
565 {
566  float interpolated_lspf[10];
567  float weight;
568 
569  if (q->bitrate >= RATE_QUARTER)
570  weight = 0.25 * (subframe_num + 1);
571  else if (q->bitrate == RATE_OCTAVE && !subframe_num)
572  weight = 0.625;
573  else
574  weight = 1.0;
575 
576  if (weight != 1.0) {
577  ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
578  weight, 1.0 - weight, 10);
579  lspf2lpc(interpolated_lspf, lpc);
580  } else if (q->bitrate >= RATE_QUARTER ||
581  (q->bitrate == I_F_Q && !subframe_num))
582  lspf2lpc(curr_lspf, lpc);
583  else if (q->bitrate == SILENCE && !subframe_num)
584  lspf2lpc(q->prev_lspf, lpc);
585 }
586 
587 static qcelp_packet_rate buf_size2bitrate(const int buf_size)
588 {
589  switch (buf_size) {
590  case 35: return RATE_FULL;
591  case 17: return RATE_HALF;
592  case 8: return RATE_QUARTER;
593  case 4: return RATE_OCTAVE;
594  case 1: return SILENCE;
595  }
596 
597  return I_F_Q;
598 }
599 
600 /**
601  * Determine the bitrate from the frame size and/or the first byte of the frame.
602  *
603  * @param avctx the AV codec context
604  * @param buf_size length of the buffer
605  * @param buf the buffer
606  *
607  * @return the bitrate on success,
608  * I_F_Q if the bitrate cannot be satisfactorily determined
609  *
610  * TIA/EIA/IS-733 2.4.8.7.1
611  */
613  const int buf_size,
614  const uint8_t **buf)
615 {
617 
618  if ((bitrate = buf_size2bitrate(buf_size)) >= 0) {
619  if (bitrate > **buf) {
620  QCELPContext *q = avctx->priv_data;
621  if (!q->warned_buf_mismatch_bitrate) {
622  av_log(avctx, AV_LOG_WARNING,
623  "Claimed bitrate and buffer size mismatch.\n");
625  }
626  bitrate = **buf;
627  } else if (bitrate < **buf) {
628  av_log(avctx, AV_LOG_ERROR,
629  "Buffer is too small for the claimed bitrate.\n");
630  return I_F_Q;
631  }
632  (*buf)++;
633  } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) {
634  av_log(avctx, AV_LOG_WARNING,
635  "Bitrate byte missing, guessing bitrate from packet size.\n");
636  } else
637  return I_F_Q;
638 
639  if (bitrate == SILENCE) {
640  // FIXME: Remove this warning when tested with samples.
641  avpriv_request_sample(avctx, "Blank frame handling");
642  }
643  return bitrate;
644 }
645 
647  const char *message)
648 {
649  av_log(avctx, AV_LOG_WARNING, "Frame #%"PRId64", IFQ: %s\n",
650  avctx->frame_num, message);
651 }
652 
653 static void postfilter(QCELPContext *q, float *samples, float *lpc)
654 {
655  static const float pow_0_775[10] = {
656  0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
657  0.216676, 0.167924, 0.130141, 0.100859, 0.078166
658  }, pow_0_625[10] = {
659  0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
660  0.059605, 0.037253, 0.023283, 0.014552, 0.009095
661  };
662  float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
663  int n;
664 
665  for (n = 0; n < 10; n++) {
666  lpc_s[n] = lpc[n] * pow_0_625[n];
667  lpc_p[n] = lpc[n] * pow_0_775[n];
668  }
669 
670  ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
671  q->formant_mem + 10, 160, 10);
672  memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10);
673  ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
674  memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
675 
676  ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
677 
678  ff_adaptive_gain_control(samples, pole_out + 10,
680  q->formant_mem + 10,
681  160),
682  160, 0.9375, &q->postfilter_agc_mem);
683 }
684 
686  int *got_frame_ptr, AVPacket *avpkt)
687 {
688  const uint8_t *buf = avpkt->data;
689  int buf_size = avpkt->size;
690  QCELPContext *q = avctx->priv_data;
691  float *outbuffer;
692  int i, ret;
693  float quantized_lspf[10], lpc[10];
694  float gain[16];
695  float *formant_mem;
696 
697  /* get output buffer */
698  frame->nb_samples = 160;
699  if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
700  return ret;
701  outbuffer = (float *)frame->data[0];
702 
703  if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) {
704  warn_insufficient_frame_quality(avctx, "Bitrate cannot be determined.");
705  goto erasure;
706  }
707 
708  if (q->bitrate == RATE_OCTAVE &&
709  (q->first16bits = AV_RB16(buf)) == 0xFFFF) {
710  warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
711  goto erasure;
712  }
713 
714  if (q->bitrate > SILENCE) {
716  const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] +
718  uint8_t *unpacked_data = (uint8_t *)&q->frame;
719 
720  if ((ret = init_get_bits8(&q->gb, buf, buf_size)) < 0)
721  return ret;
722 
723  memset(&q->frame, 0, sizeof(QCELPFrame));
724 
725  for (; bitmaps < bitmaps_end; bitmaps++)
726  unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
727 
728  // Check for erasures/blanks on rates 1, 1/4 and 1/8.
729  if (q->frame.reserved) {
730  warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
731  goto erasure;
732  }
733  if (q->bitrate == RATE_QUARTER &&
735  warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
736  goto erasure;
737  }
738 
739  if (q->bitrate >= RATE_HALF) {
740  for (i = 0; i < 4; i++) {
741  if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) {
742  warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
743  goto erasure;
744  }
745  }
746  }
747  }
748 
749  decode_gain_and_index(q, gain);
750  compute_svector(q, gain, outbuffer);
751 
752  if (decode_lspf(q, quantized_lspf) < 0) {
753  warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
754  goto erasure;
755  }
756 
757  apply_pitch_filters(q, outbuffer);
758 
759  if (q->bitrate == I_F_Q) {
760 erasure:
761  q->bitrate = I_F_Q;
762  q->erasure_count++;
763  decode_gain_and_index(q, gain);
764  compute_svector(q, gain, outbuffer);
765  decode_lspf(q, quantized_lspf);
766  apply_pitch_filters(q, outbuffer);
767  } else
768  q->erasure_count = 0;
769 
770  formant_mem = q->formant_mem + 10;
771  for (i = 0; i < 4; i++) {
772  interpolate_lpc(q, quantized_lspf, lpc, i);
773  ff_celp_lp_synthesis_filterf(formant_mem, lpc,
774  outbuffer + i * 40, 40, 10);
775  formant_mem += 40;
776  }
777 
778  // postfilter, as per TIA/EIA/IS-733 2.4.8.6
779  postfilter(q, outbuffer, lpc);
780 
781  memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
782 
783  memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
784  q->prev_bitrate = q->bitrate;
785 
786  *got_frame_ptr = 1;
787 
788  return buf_size;
789 }
790 
792  .p.name = "qcelp",
793  CODEC_LONG_NAME("QCELP / PureVoice"),
794  .p.type = AVMEDIA_TYPE_AUDIO,
795  .p.id = AV_CODEC_ID_QCELP,
796  .init = qcelp_decode_init,
798  .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
799  .priv_data_size = sizeof(QCELPContext),
800 };
AV_LOG_WARNING
#define AV_LOG_WARNING
Something somehow does not look correct.
Definition: log.h:186
codebook_sanity_check_for_rate_quarter
static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
If the received packet is Rate 1/4 a further sanity check is made of the codebook gain.
Definition: qcelpdec.c:272
av_clip
#define av_clip
Definition: common.h:98
SILENCE
@ SILENCE
Definition: qcelpdec.c:45
acelp_vectors.h
qcelp_lspvq
static const qcelp_vector *const qcelp_lspvq[5]
Definition: qcelpdata.h:414
QCELPContext::erasure_count
uint8_t erasure_count
Definition: qcelpdec.c:57
message
Definition: api-threadmessage-test.c:46
QCELPContext::pitch_lag
uint8_t pitch_lag[4]
Definition: qcelpdec.c:69
QCELPContext::gb
GetBitContext gb
Definition: qcelpdec.c:53
qcelp_unpacking_bitmaps_per_rate
static const QCELPBitmap *const qcelp_unpacking_bitmaps_per_rate[5]
Bitmapping data position for each packet type in the QCELPContext.
Definition: qcelpdata.h:268
QCELPFrame::cindex
uint8_t cindex[16]
codebook index for each codebook subframe
Definition: qcelpdata.h:45
qcelp_rate_full_codebook
static const int16_t qcelp_rate_full_codebook[128]
Circular codebook for rate 1 frames in x*100 form.
Definition: qcelpdata.h:459
QCELP_LSP_OCTAVE_PREDICTOR
#define QCELP_LSP_OCTAVE_PREDICTOR
Predictor coefficient for the conversion of LSP codes to LSP frequencies for 1/8 and I_F_Q.
Definition: qcelpdata.h:541
qcelp_decode_init
static av_cold int qcelp_decode_init(AVCodecContext *avctx)
Initialize the speech codec according to the specification.
Definition: qcelpdec.c:84
AVFrame
This structure describes decoded (raw) audio or video data.
Definition: frame.h:340
AVPacket::data
uint8_t * data
Definition: packet.h:522
qcelp_decode_frame
static int qcelp_decode_frame(AVCodecContext *avctx, AVFrame *frame, int *got_frame_ptr, AVPacket *avpkt)
Definition: qcelpdec.c:685
FFCodec
Definition: codec_internal.h:127
FFMAX
#define FFMAX(a, b)
Definition: macros.h:47
QCELPContext::postfilter_synth_mem
float postfilter_synth_mem[10]
Definition: qcelpdec.c:74
qcelp_packet_rate
qcelp_packet_rate
Definition: qcelpdec.c:43
QCELPFrame::lspv
uint8_t lspv[10]
line spectral pair frequencies (LSP) for RATE_OCTAVE, line spectral pair frequencies grouped into fiv...
Definition: qcelpdata.h:60
ff_celp_lp_synthesis_filterf
void ff_celp_lp_synthesis_filterf(float *out, const float *filter_coeffs, const float *in, int buffer_length, int filter_length)
LP synthesis filter.
Definition: celp_filters.c:85
AVFrame::data
uint8_t * data[AV_NUM_DATA_POINTERS]
pointer to the picture/channel planes.
Definition: frame.h:361
RATE_FULL
@ RATE_FULL
Definition: qcelpdec.c:49
get_bits
static unsigned int get_bits(GetBitContext *s, int n)
Read 1-25 bits.
Definition: get_bits.h:335
FFCodec::p
AVCodec p
The public AVCodec.
Definition: codec_internal.h:131
QCELPContext::prev_g1
int prev_g1[2]
Definition: qcelpdec.c:66
QCELP_SQRT1887
#define QCELP_SQRT1887
sqrt(1.887) is the maximum of the pseudorandom white sequence used to generate the scaled codebook ve...
Definition: qcelpdata.h:511
AVCodecContext::ch_layout
AVChannelLayout ch_layout
Audio channel layout.
Definition: avcodec.h:1065
QCELPBitmap::index
uint8_t index
index into the QCELPContext structure
Definition: qcelpdata.h:77
GetBitContext
Definition: get_bits.h:108
QCELPFrame::plag
uint8_t plag[4]
pitch lag for each pitch subframe
Definition: qcelpdata.h:50
RATE_OCTAVE
@ RATE_OCTAVE
Definition: qcelpdec.c:46
ff_adaptive_gain_control
void ff_adaptive_gain_control(float *out, const float *in, float speech_energ, int size, float alpha, float *gain_mem)
Adaptive gain control (as used in AMR postfiltering)
Definition: acelp_vectors.c:192
avassert.h
rnd
#define rnd()
Definition: checkasm.h:163
AV_LOG_ERROR
#define AV_LOG_ERROR
Something went wrong and cannot losslessly be recovered.
Definition: log.h:180
av_cold
#define av_cold
Definition: attributes.h:90
do_pitchfilter
static const float * do_pitchfilter(float memory[303], const float v_in[160], const float gain[4], const uint8_t *lag, const uint8_t pfrac[4])
Apply filter in pitch-subframe steps.
Definition: qcelpdec.c:422
init_get_bits8
static int init_get_bits8(GetBitContext *s, const uint8_t *buffer, int byte_size)
Initialize GetBitContext.
Definition: get_bits.h:545
QCELP_LSP_SPREAD_FACTOR
#define QCELP_LSP_SPREAD_FACTOR
This spread factor is used, for bitrate 1/8 and I_F_Q, to force LSP frequencies to be at least 80 Hz ...
Definition: qcelpdata.h:533
FF_CODEC_DECODE_CB
#define FF_CODEC_DECODE_CB(func)
Definition: codec_internal.h:287
buf_size2bitrate
static qcelp_packet_rate buf_size2bitrate(const int buf_size)
Definition: qcelpdec.c:587
QCELPContext::formant_mem
float formant_mem[170]
Definition: qcelpdec.c:64
bitrate
int64_t bitrate
Definition: av1_levels.c:47
qcelp_unpacking_bitmaps_lengths
static const uint16_t qcelp_unpacking_bitmaps_lengths[5]
Definition: qcelpdata.h:276
AVMEDIA_TYPE_AUDIO
@ AVMEDIA_TYPE_AUDIO
Definition: avutil.h:202
decode.h
get_bits.h
QCELPContext::first16bits
uint16_t first16bits
Definition: qcelpdec.c:70
qcelp_rate_half_codebook
static const int8_t qcelp_rate_half_codebook[128]
Circular codebook for rate 1/2 frames in x*2 form.
Definition: qcelpdata.h:484
qcelp_hammsinc_table
static const float qcelp_hammsinc_table[4]
Pre-calculated table for hammsinc function.
Definition: qcelpdata.h:74
I_F_Q
@ I_F_Q
insufficient frame quality
Definition: qcelpdec.c:44
CODEC_LONG_NAME
#define CODEC_LONG_NAME(str)
Definition: codec_internal.h:272
frame
static AVFrame * frame
Definition: demux_decode.c:54
FFABS
#define FFABS(a)
Absolute value, Note, INT_MIN / INT64_MIN result in undefined behavior as they are not representable ...
Definition: common.h:72
QCELPContext::pitch_gain
float pitch_gain[4]
Definition: qcelpdec.c:68
if
if(ret)
Definition: filter_design.txt:179
QCELPContext::rnd_fir_filter_mem
float rnd_fir_filter_mem[180]
Definition: qcelpdec.c:63
fabs
static __device__ float fabs(float a)
Definition: cuda_runtime.h:182
apply_pitch_filters
static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
Definition: qcelpdec.c:466
QCELPFrame::cbsign
uint8_t cbsign[16]
sign of the codebook gain for each codebook subframe
Definition: qcelpdata.h:43
QCELPContext::predictor_lspf
float predictor_lspf[10]
LSP predictor for RATE_OCTAVE and I_F_Q.
Definition: qcelpdec.c:60
QCELPContext::warned_buf_mismatch_bitrate
uint8_t warned_buf_mismatch_bitrate
Definition: qcelpdec.c:71
QCELPContext::prev_bitrate
int prev_bitrate
Definition: qcelpdec.c:67
celp_filters.h
QCELPContext::postfilter_agc_mem
float postfilter_agc_mem
Definition: qcelpdec.c:75
lspf2lpc
static void lspf2lpc(const float *lspf, float *lpc)
Reconstruct LPC coefficients from the line spectral pair frequencies and perform bandwidth expansion.
Definition: qcelpdec.c:535
weight
static int weight(int i, int blen, int offset)
Definition: diracdec.c:1562
float_dsp.h
QCELPContext::prev_lspf
float prev_lspf[10]
Definition: qcelpdec.c:59
AV_CODEC_CAP_CHANNEL_CONF
#define AV_CODEC_CAP_CHANNEL_CONF
Codec should fill in channel configuration and samplerate instead of container.
Definition: codec.h:106
warn_insufficient_frame_quality
static void warn_insufficient_frame_quality(AVCodecContext *avctx, const char *message)
Definition: qcelpdec.c:646
AV_CODEC_ID_QCELP
@ AV_CODEC_ID_QCELP
Definition: codec_id.h:464
QCELPFrame::pgain
uint8_t pgain[4]
pitch gain for each pitch subframe
Definition: qcelpdata.h:52
ff_get_buffer
int ff_get_buffer(AVCodecContext *avctx, AVFrame *frame, int flags)
Get a buffer for a frame.
Definition: decode.c:1568
AV_CODEC_CAP_DR1
#define AV_CODEC_CAP_DR1
Codec uses get_buffer() or get_encode_buffer() for allocating buffers and supports custom allocators.
Definition: codec.h:52
QCELPContext::bitrate
qcelp_packet_rate bitrate
Definition: qcelpdec.c:54
AVPacket::size
int size
Definition: packet.h:523
AVChannelLayout
An AVChannelLayout holds information about the channel layout of audio data.
Definition: channel_layout.h:303
codec_internal.h
AVCodecContext::sample_fmt
enum AVSampleFormat sample_fmt
audio sample format
Definition: avcodec.h:1057
QCELPBitmap::bitlen
uint8_t bitlen
number of bits to read
Definition: qcelpdata.h:79
diff
static av_always_inline int diff(const struct color_info *a, const struct color_info *b, const int trans_thresh)
Definition: vf_paletteuse.c:164
interpolate_lpc
static void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc, const int subframe_num)
Interpolate LSP frequencies and compute LPC coefficients for a given bitrate & pitch subframe.
Definition: qcelpdec.c:563
QCELPFrame
QCELP unpacked data frame.
Definition: qcelpdata.h:40
M_PI
#define M_PI
Definition: mathematics.h:67
ff_tilt_compensation
void ff_tilt_compensation(float *mem, float tilt, float *samples, int size)
Apply tilt compensation filter, 1 - tilt * z-1.
Definition: acelp_filters.c:138
QCELPBitmap
Definition: qcelpdata.h:76
av_assert2
#define av_assert2(cond)
assert() equivalent, that does lie in speed critical code.
Definition: avassert.h:67
AVFrame::nb_samples
int nb_samples
number of audio samples (per channel) described by this frame
Definition: frame.h:420
qcelpdata.h
i
#define i(width, name, range_min, range_max)
Definition: cbs_h2645.c:255
apply_gain_ctrl
static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in)
Apply generic gain control.
Definition: qcelpdec.c:395
QCELPBitmap::bitpos
uint8_t bitpos
position of the lowest bit in the value's byte
Definition: qcelpdata.h:78
QCELPContext::octave_count
uint8_t octave_count
count the consecutive RATE_OCTAVE frames
Definition: qcelpdec.c:58
decode_lspf
static int decode_lspf(QCELPContext *q, float *lspf)
Decode the 10 quantized LSP frequencies from the LSPV/LSP transmission codes of any bitrate and check...
Definition: qcelpdec.c:110
determine_bitrate
static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, const int buf_size, const uint8_t **buf)
Determine the bitrate from the frame size and/or the first byte of the frame.
Definition: qcelpdec.c:612
FFMIN
#define FFMIN(a, b)
Definition: macros.h:49
acelp_filters.h
ff_weighted_vector_sumf
void ff_weighted_vector_sumf(float *out, const float *in_a, const float *in_b, float weight_coeff_a, float weight_coeff_b, int length)
float implementation of weighted sum of two vectors.
Definition: acelp_vectors.c:182
qcelp_g12ga
static const float qcelp_g12ga[61]
Table for computing Ga (decoded linear codebook gain magnitude)
Definition: qcelpdata.h:436
AVCodec::name
const char * name
Name of the codec implementation.
Definition: codec.h:194
QCELPContext::pitch_pre_filter_mem
float pitch_pre_filter_mem[303]
Definition: qcelpdec.c:62
avcodec.h
AVCodecContext::frame_num
int64_t frame_num
Frame counter, set by libavcodec.
Definition: avcodec.h:2030
ret
ret
Definition: filter_design.txt:187
lsp.h
ff_celp_lp_zero_synthesis_filterf
void ff_celp_lp_zero_synthesis_filterf(float *out, const float *filter_coeffs, const float *in, int buffer_length, int filter_length)
LP zero synthesis filter.
Definition: celp_filters.c:200
compute_svector
static void compute_svector(QCELPContext *q, const float *gain, float *cdn_vector)
Compute the scaled codebook vector Cdn From INDEX and GAIN for all rates.
Definition: qcelpdec.c:308
QCELPContext::postfilter_tilt_mem
float postfilter_tilt_mem
Definition: qcelpdec.c:76
AVCodecContext
main external API structure.
Definition: avcodec.h:445
RATE_HALF
@ RATE_HALF
Definition: qcelpdec.c:48
channel_layout.h
decode_gain_and_index
static void decode_gain_and_index(QCELPContext *q, float *gain)
Convert codebook transmission codes to GAIN and INDEX.
Definition: qcelpdec.c:194
QCELPContext
Definition: qcelpdec.c:52
av_channel_layout_uninit
void av_channel_layout_uninit(AVChannelLayout *channel_layout)
Free any allocated data in the channel layout and reset the channel count to 0.
Definition: channel_layout.c:432
QCELP_BANDWIDTH_EXPANSION_COEFF
#define QCELP_BANDWIDTH_EXPANSION_COEFF
Initial coefficient to perform bandwidth expansion on LPC.
Definition: qcelpdata.h:550
RATE_QUARTER
@ RATE_QUARTER
Definition: qcelpdec.c:47
samples
Filter the word “frame” indicates either a video frame or a group of audio samples
Definition: filter_design.txt:8
avpriv_scalarproduct_float_c
float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len)
Return the scalar product of two vectors.
Definition: float_dsp.c:124
avpriv_request_sample
#define avpriv_request_sample(...)
Definition: tableprint_vlc.h:36
QCELPContext::frame
QCELPFrame frame
unpacked data frame
Definition: qcelpdec.c:55
QCELP_RATE_HALF_CODEBOOK_RATIO
#define QCELP_RATE_HALF_CODEBOOK_RATIO
Definition: qcelpdata.h:502
AV_CHANNEL_LAYOUT_MONO
#define AV_CHANNEL_LAYOUT_MONO
Definition: channel_layout.h:378
smooth
static float smooth(DeshakeOpenCLContext *deshake_ctx, float *gauss_kernel, int length, float max_val, AVFifo *values)
Definition: vf_deshake_opencl.c:888
AVPacket
This structure stores compressed data.
Definition: packet.h:499
AVCodecContext::priv_data
void * priv_data
Definition: avcodec.h:472
ff_qcelp_decoder
const FFCodec ff_qcelp_decoder
Definition: qcelpdec.c:791
ff_scale_vector_to_given_sum_of_squares
void ff_scale_vector_to_given_sum_of_squares(float *out, const float *in, float sum_of_squares, const int n)
Set the sum of squares of a signal by scaling.
Definition: acelp_vectors.c:213
QCELPFrame::reserved
uint8_t reserved
reserved bits only present in bitrate 1, 1/4 and 1/8 packets
Definition: qcelpdata.h:65
av_log
#define av_log(a,...)
Definition: tableprint_vlc.h:27
ff_acelp_lspd2lpc
void ff_acelp_lspd2lpc(const double *lsp, float *lpc, int lp_half_order)
Reconstruct LPC coefficients from the line spectral pair frequencies.
Definition: lsp.c:220
qcelp_rnd_fir_coefs
static const double qcelp_rnd_fir_coefs[11]
Table for impulse response of BPF used to filter the white excitation for bitrate 1/4 synthesis.
Definition: qcelpdata.h:521
QCELPContext::last_codebook_gain
float last_codebook_gain
Definition: qcelpdec.c:65
QCELPFrame::pfrac
uint8_t pfrac[4]
fractional pitch lag for each pitch subframe
Definition: qcelpdata.h:51
AV_SAMPLE_FMT_FLT
@ AV_SAMPLE_FMT_FLT
float
Definition: samplefmt.h:60
QCELPFrame::cbgain
uint8_t cbgain[16]
unsigned codebook gain for each codebook subframe
Definition: qcelpdata.h:44
QCELP_RATE_FULL_CODEBOOK_RATIO
#define QCELP_RATE_FULL_CODEBOOK_RATIO
Definition: qcelpdata.h:477
postfilter
static void postfilter(QCELPContext *q, float *samples, float *lpc)
Definition: qcelpdec.c:653
AV_RB16
uint64_t_TMPL AV_WL64 unsigned int_TMPL AV_WL32 unsigned int_TMPL AV_WL24 unsigned int_TMPL AV_WL16 uint64_t_TMPL AV_WB64 unsigned int_TMPL AV_WB32 unsigned int_TMPL AV_WB24 unsigned int_TMPL AV_RB16
Definition: bytestream.h:98
QCELPContext::pitch_synthesis_filter_mem
float pitch_synthesis_filter_mem[303]
Definition: qcelpdec.c:61