git.sesse.net Git - ffmpeg/blob - libavcodec/qcelpdec.c

   1 /*
   2  * QCELP decoder
   3  * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet
   4  *
   5  * This file is part of Libav.
   6  *
   7  * Libav 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  * Libav 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 Libav; 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 Libav merging spearheaded by Kenan Gillet
  27  * @remark Development mentored by Benjamin Larson
  28  */
  29
  30 #include <stddef.h>
  31
  32 #include "avcodec.h"
  33 #include "internal.h"
  34 #include "get_bits.h"
  35
  36 #include "qcelpdata.h"
  37
  38 #include "celp_math.h"
  39 #include "celp_filters.h"
  40 #include "acelp_filters.h"
  41 #include "acelp_vectors.h"
  42 #include "lsp.h"
  43
  44 #undef NDEBUG
  45 #include <assert.h>
  46
  47 typedef enum
  48 {
  49     I_F_Q = -1,    /**< insufficient frame quality */
  50     SILENCE,
  51     RATE_OCTAVE,
  52     RATE_QUARTER,
  53     RATE_HALF,
  54     RATE_FULL
  55 } qcelp_packet_rate;
  56
  57 typedef struct
  58 {
  59     GetBitContext     gb;
  60     qcelp_packet_rate bitrate;
  61     QCELPFrame        frame;    /**< unpacked data frame */
  62
  63     uint8_t  erasure_count;
  64     uint8_t  octave_count;      /**< count the consecutive RATE_OCTAVE frames */
  65     float    prev_lspf[10];
  66     float    predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */
  67     float    pitch_synthesis_filter_mem[303];
  68     float    pitch_pre_filter_mem[303];
  69     float    rnd_fir_filter_mem[180];
  70     float    formant_mem[170];
  71     float    last_codebook_gain;
  72     int      prev_g1[2];
  73     int      prev_bitrate;
  74     float    pitch_gain[4];
  75     uint8_t  pitch_lag[4];
  76     uint16_t first16bits;
  77     uint8_t  warned_buf_mismatch_bitrate;
  78
  79     /* postfilter */
  80     float    postfilter_synth_mem[10];
  81     float    postfilter_agc_mem;
  82     float    postfilter_tilt_mem;
  83 } QCELPContext;
  84
  85 /**
  86  * Initialize the speech codec according to the specification.
  87  *
  88  * TIA/EIA/IS-733 2.4.9
  89  */
  90 static av_cold int qcelp_decode_init(AVCodecContext *avctx)
  91 {
  92     QCELPContext *q = avctx->priv_data;
  93     int i;
  94
  95     avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
  96
  97     for(i=0; i<10; i++)
  98         q->prev_lspf[i] = (i+1)/11.;
  99
 100     return 0;
 101 }
 102
 103 /**
 104  * Decode the 10 quantized LSP frequencies from the LSPV/LSP
 105  * transmission codes of any bitrate and check for badly received packets.
 106  *
 107  * @param q the context
 108  * @param lspf line spectral pair frequencies
 109  *
 110  * @return 0 on success, -1 if the packet is badly received
 111  *
 112  * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
 113  */
 114 static int decode_lspf(QCELPContext *q, float *lspf)
 115 {
 116     int i;
 117     float tmp_lspf, smooth, erasure_coeff;
 118     const float *predictors;
 119
 120     if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) {
 121         predictors = (q->prev_bitrate != RATE_OCTAVE &&
 122                        q->prev_bitrate != I_F_Q ?
 123                        q->prev_lspf : q->predictor_lspf);
 124
 125         if (q->bitrate == RATE_OCTAVE) {
 126             q->octave_count++;
 127
 128             for (i=0; i<10; i++) {
 129                 q->predictor_lspf[i] =
 130                              lspf[i] = (q->frame.lspv[i] ?  QCELP_LSP_SPREAD_FACTOR
 131                                                          : -QCELP_LSP_SPREAD_FACTOR)
 132                                      + predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR
 133                                      + (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR)/11);
 134             }
 135             smooth = (q->octave_count < 10 ? .875 : 0.1);
 136         } else {
 137             erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
 138
 139             assert(q->bitrate == I_F_Q);
 140
 141             if(q->erasure_count > 1)
 142                 erasure_coeff *= (q->erasure_count < 4 ? 0.9 : 0.7);
 143
 144             for(i = 0; i < 10; i++) {
 145                 q->predictor_lspf[i] =
 146                              lspf[i] = (i + 1) * ( 1 - erasure_coeff)/11
 147                                      + erasure_coeff * predictors[i];
 148             }
 149             smooth = 0.125;
 150         }
 151
 152         // Check the stability of the LSP frequencies.
 153         lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
 154         for(i=1; i<10; i++)
 155             lspf[i] = FFMAX(lspf[i], (lspf[i-1] + QCELP_LSP_SPREAD_FACTOR));
 156
 157         lspf[9] = FFMIN(lspf[9], (1.0 - QCELP_LSP_SPREAD_FACTOR));
 158         for(i=9; i>0; i--)
 159             lspf[i-1] = FFMIN(lspf[i-1], (lspf[i] - QCELP_LSP_SPREAD_FACTOR));
 160
 161         // Low-pass filter the LSP frequencies.
 162         ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
 163     } else {
 164         q->octave_count = 0;
 165
 166         tmp_lspf = 0.;
 167         for (i = 0; i < 5; i++) {
 168             lspf[2*i+0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
 169             lspf[2*i+1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
 170         }
 171
 172         // Check for badly received packets.
 173         if (q->bitrate == RATE_QUARTER) {
 174             if(lspf[9] <= .70 || lspf[9] >=  .97)
 175                 return -1;
 176             for(i=3; i<10; i++)
 177                 if(fabs(lspf[i] - lspf[i-2]) < .08)
 178                     return -1;
 179         } else {
 180             if(lspf[9] <= .66 || lspf[9] >= .985)
 181                 return -1;
 182             for(i=4; i<10; i++)
 183                 if (fabs(lspf[i] - lspf[i-4]) < .0931)
 184                     return -1;
 185         }
 186     }
 187     return 0;
 188 }
 189
 190 /**
 191  * Convert codebook transmission codes to GAIN and INDEX.
 192  *
 193  * @param q the context
 194  * @param gain array holding the decoded gain
 195  *
 196  * TIA/EIA/IS-733 2.4.6.2
 197  */
 198 static void decode_gain_and_index(QCELPContext  *q,
 199                                   float *gain) {
 200     int   i, subframes_count, g1[16];
 201     float slope;
 202
 203     if (q->bitrate >= RATE_QUARTER) {
 204         switch (q->bitrate) {
 205             case RATE_FULL: subframes_count = 16; break;
 206             case RATE_HALF: subframes_count = 4;  break;
 207             default:        subframes_count = 5;
 208         }
 209         for(i = 0; i < subframes_count; i++) {
 210             g1[i] = 4 * q->frame.cbgain[i];
 211             if (q->bitrate == RATE_FULL && !((i+1) & 3)) {
 212                 g1[i] += av_clip((g1[i-1] + g1[i-2] + g1[i-3]) / 3 - 6, 0, 32);
 213             }
 214
 215             gain[i] = qcelp_g12ga[g1[i]];
 216
 217             if (q->frame.cbsign[i]) {
 218                 gain[i] = -gain[i];
 219                 q->frame.cindex[i] = (q->frame.cindex[i]-89) & 127;
 220             }
 221         }
 222
 223         q->prev_g1[0] = g1[i-2];
 224         q->prev_g1[1] = g1[i-1];
 225         q->last_codebook_gain = qcelp_g12ga[g1[i-1]];
 226
 227         if (q->bitrate == RATE_QUARTER) {
 228             // Provide smoothing of the unvoiced excitation energy.
 229             gain[7] =     gain[4];
 230             gain[6] = 0.4*gain[3] + 0.6*gain[4];
 231             gain[5] =     gain[3];
 232             gain[4] = 0.8*gain[2] + 0.2*gain[3];
 233             gain[3] = 0.2*gain[1] + 0.8*gain[2];
 234             gain[2] =     gain[1];
 235             gain[1] = 0.6*gain[0] + 0.4*gain[1];
 236         }
 237     } else if (q->bitrate != SILENCE) {
 238         if (q->bitrate == RATE_OCTAVE) {
 239             g1[0] = 2 * q->frame.cbgain[0]
 240                   + av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
 241             subframes_count = 8;
 242         } else {
 243             assert(q->bitrate == I_F_Q);
 244
 245             g1[0] = q->prev_g1[1];
 246             switch (q->erasure_count) {
 247                 case 1 : break;
 248                 case 2 : g1[0] -= 1; break;
 249                 case 3 : g1[0] -= 2; break;
 250                 default: g1[0] -= 6;
 251             }
 252             if(g1[0] < 0)
 253                 g1[0] = 0;
 254             subframes_count = 4;
 255         }
 256         // This interpolation is done to produce smoother background noise.
 257         slope = 0.5*(qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
 258         for(i=1; i<=subframes_count; i++)
 259             gain[i-1] = q->last_codebook_gain + slope * i;
 260
 261         q->last_codebook_gain = gain[i-2];
 262         q->prev_g1[0] = q->prev_g1[1];
 263         q->prev_g1[1] = g1[0];
 264     }
 265 }
 266
 267 /**
 268  * If the received packet is Rate 1/4 a further sanity check is made of the
 269  * codebook gain.
 270  *
 271  * @param cbgain the unpacked cbgain array
 272  * @return -1 if the sanity check fails, 0 otherwise
 273  *
 274  * TIA/EIA/IS-733 2.4.8.7.3
 275  */
 276 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
 277 {
 278     int i, diff, prev_diff=0;
 279
 280     for(i=1; i<5; i++) {
 281         diff = cbgain[i] - cbgain[i-1];
 282         if(FFABS(diff) > 10)
 283             return -1;
 284         else if(FFABS(diff - prev_diff) > 12)
 285             return -1;
 286         prev_diff = diff;
 287     }
 288     return 0;
 289 }
 290
 291 /**
 292  * Compute the scaled codebook vector Cdn From INDEX and GAIN
 293  * for all rates.
 294  *
 295  * The specification lacks some information here.
 296  *
 297  * TIA/EIA/IS-733 has an omission on the codebook index determination
 298  * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
 299  * you have to subtract the decoded index parameter from the given scaled
 300  * codebook vector index 'n' to get the desired circular codebook index, but
 301  * it does not mention that you have to clamp 'n' to [0-9] in order to get
 302  * RI-compliant results.
 303  *
 304  * The reason for this mistake seems to be the fact they forgot to mention you
 305  * have to do these calculations per codebook subframe and adjust given
 306  * equation values accordingly.
 307  *
 308  * @param q the context
 309  * @param gain array holding the 4 pitch subframe gain values
 310  * @param cdn_vector array for the generated scaled codebook vector
 311  */
 312 static void compute_svector(QCELPContext *q, const float *gain,
 313                             float *cdn_vector)
 314 {
 315     int      i, j, k;
 316     uint16_t cbseed, cindex;
 317     float    *rnd, tmp_gain, fir_filter_value;
 318
 319     switch (q->bitrate) {
 320         case RATE_FULL:
 321             for (i = 0; i < 16; i++) {
 322                 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
 323                 cindex = -q->frame.cindex[i];
 324                 for(j=0; j<10; j++)
 325                     *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
 326             }
 327         break;
 328         case RATE_HALF:
 329             for (i = 0; i < 4; i++) {
 330                 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
 331                 cindex = -q->frame.cindex[i];
 332                 for (j = 0; j < 40; j++)
 333                 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
 334             }
 335         break;
 336         case RATE_QUARTER:
 337             cbseed = (0x0003 & q->frame.lspv[4])<<14 |
 338                      (0x003F & q->frame.lspv[3])<< 8 |
 339                      (0x0060 & q->frame.lspv[2])<< 1 |
 340                      (0x0007 & q->frame.lspv[1])<< 3 |
 341                      (0x0038 & q->frame.lspv[0])>> 3 ;
 342             rnd = q->rnd_fir_filter_mem + 20;
 343             for (i = 0; i < 8; i++) {
 344                 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
 345                 for (k = 0; k < 20; k++) {
 346                     cbseed = 521 * cbseed + 259;
 347                     *rnd = (int16_t)cbseed;
 348
 349                     // FIR filter
 350                     fir_filter_value = 0.0;
 351                     for(j=0; j<10; j++)
 352                         fir_filter_value += qcelp_rnd_fir_coefs[j ]
 353                                           * (rnd[-j ] + rnd[-20+j]);
 354
 355                     fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
 356                     *cdn_vector++ = tmp_gain * fir_filter_value;
 357                     rnd++;
 358                 }
 359             }
 360             memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, 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,
 396                             const float *v_in)
 397 {
 398     int i;
 399
 400     for (i = 0; i < 160; i += 40)
 401         ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i,
 402                                                 ff_dot_productf(v_ref + i,
 403                                                                 v_ref + i, 40),
 404                                                 40);
 405 }
 406
 407 /**
 408  * Apply filter in pitch-subframe steps.
 409  *
 410  * @param memory buffer for the previous state of the filter
 411  *        - must be able to contain 303 elements
 412  *        - the 143 first elements are from the previous state
 413  *        - the next 160 are for output
 414  * @param v_in input filter vector
 415  * @param gain per-subframe gain array, each element is between 0.0 and 2.0
 416  * @param lag per-subframe lag array, each element is
 417  *        - between 16 and 143 if its corresponding pfrac is 0,
 418  *        - between 16 and 139 otherwise
 419  * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
 420  *        otherwise
 421  *
 422  * @return filter output vector
 423  */
 424 static const float *do_pitchfilter(float memory[303], const float v_in[160],
 425                                    const float gain[4], const uint8_t *lag,
 426                                    const uint8_t pfrac[4])
 427 {
 428     int         i, j;
 429     float       *v_lag, *v_out;
 430     const float *v_len;
 431
 432     v_out = memory + 143; // Output vector starts at memory[143].
 433
 434     for (i = 0; i < 4; i++) {
 435         if (gain[i]) {
 436             v_lag = memory + 143 + 40 * i - lag[i];
 437             for (v_len = v_in + 40; v_in < v_len; v_in++) {
 438                 if (pfrac[i]) { // If it is a fractional lag...
 439                     for(j=0, *v_out=0.; j<4; j++)
 440                         *v_out += qcelp_hammsinc_table[j] * (v_lag[j-4] + v_lag[3-j]);
 441                 }else
 442                     *v_out = *v_lag;
 443
 444                 *v_out = *v_in + gain[i] * *v_out;
 445
 446                 v_lag++;
 447                 v_out++;
 448             }
 449         } else {
 450             memcpy(v_out, v_in, 40 * sizeof(float));
 451             v_in  += 40;
 452             v_out += 40;
 453         }
 454     }
 455
 456     memmove(memory, memory + 160, 143 * sizeof(float));
 457     return memory + 143;
 458 }
 459
 460 /**
 461  * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
 462  * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
 463  *
 464  * @param q the context
 465  * @param cdn_vector the scaled codebook vector
 466  */
 467 static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
 468 {
 469     int         i;
 470     const float *v_synthesis_filtered, *v_pre_filtered;
 471
 472     if(q->bitrate >= RATE_HALF ||
 473        q->bitrate == SILENCE ||
 474       (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
 475
 476         if(q->bitrate >= RATE_HALF) {
 477
 478             // Compute gain & lag for the whole frame.
 479             for (i = 0; i < 4; i++) {
 480                 q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
 481
 482                 q->pitch_lag[i] = q->frame.plag[i] + 16;
 483             }
 484         } else {
 485             float max_pitch_gain;
 486
 487             if (q->bitrate == I_F_Q) {
 488                   if (q->erasure_count < 3)
 489                       max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
 490                   else
 491                       max_pitch_gain = 0.0;
 492             } else {
 493                 assert(q->bitrate == SILENCE);
 494                 max_pitch_gain = 1.0;
 495             }
 496             for(i=0; i<4; i++)
 497                 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
 498
 499             memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
 500         }
 501
 502         // pitch synthesis filter
 503         v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
 504                                               cdn_vector, q->pitch_gain,
 505                                               q->pitch_lag, q->frame.pfrac);
 506
 507         // pitch prefilter update
 508         for(i=0; i<4; i++)
 509             q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
 510
 511         v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
 512                                         v_synthesis_filtered,
 513                                         q->pitch_gain, q->pitch_lag,
 514                                         q->frame.pfrac);
 515
 516         apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
 517     } else {
 518         memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17,
 519                143 * sizeof(float));
 520         memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
 521         memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
 522         memset(q->pitch_lag,  0, sizeof(q->pitch_lag));
 523     }
 524 }
 525
 526 /**
 527  * Reconstruct LPC coefficients from the line spectral pair frequencies
 528  * and perform bandwidth expansion.
 529  *
 530  * @param lspf line spectral pair frequencies
 531  * @param lpc linear predictive coding coefficients
 532  *
 533  * @note: bandwidth_expansion_coeff could be precalculated into a table
 534  *        but it seems to be slower on x86
 535  *
 536  * TIA/EIA/IS-733 2.4.3.3.5
 537  */
 538 static void lspf2lpc(const float *lspf, float *lpc)
 539 {
 540     double lsp[10];
 541     double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
 542     int   i;
 543
 544     for (i=0; i<10; i++)
 545         lsp[i] = cos(M_PI * lspf[i]);
 546
 547     ff_acelp_lspd2lpc(lsp, lpc, 5);
 548
 549     for (i = 0; i < 10; i++) {
 550         lpc[i] *= bandwidth_expansion_coeff;
 551         bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
 552     }
 553 }
 554
 555 /**
 556  * Interpolate LSP frequencies and compute LPC coefficients
 557  * for a given bitrate & pitch subframe.
 558  *
 559  * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
 560  *
 561  * @param q the context
 562  * @param curr_lspf LSP frequencies vector of the current frame
 563  * @param lpc float vector for the resulting LPC
 564  * @param subframe_num frame number in decoded stream
 565  */
 566 static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
 567                             float *lpc, const int subframe_num)
 568 {
 569     float interpolated_lspf[10];
 570     float weight;
 571
 572     if(q->bitrate >= RATE_QUARTER)
 573         weight = 0.25 * (subframe_num + 1);
 574     else if(q->bitrate == RATE_OCTAVE && !subframe_num)
 575         weight = 0.625;
 576     else
 577         weight = 1.0;
 578
 579     if (weight != 1.0) {
 580         ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
 581                                 weight, 1.0 - weight, 10);
 582         lspf2lpc(interpolated_lspf, lpc);
 583     }else if(q->bitrate >= RATE_QUARTER ||
 584              (q->bitrate == I_F_Q && !subframe_num))
 585         lspf2lpc(curr_lspf, lpc);
 586     else if(q->bitrate == SILENCE && !subframe_num)
 587         lspf2lpc(q->prev_lspf, lpc);
 588 }
 589
 590 static qcelp_packet_rate buf_size2bitrate(const int buf_size)
 591 {
 592     switch (buf_size) {
 593         case 35: return RATE_FULL;
 594         case 17: return RATE_HALF;
 595         case  8: return RATE_QUARTER;
 596         case  4: return RATE_OCTAVE;
 597         case  1: return SILENCE;
 598     }
 599
 600     return I_F_Q;
 601 }
 602
 603 /**
 604  * Determine the bitrate from the frame size and/or the first byte of the frame.
 605  *
 606  * @param avctx the AV codec context
 607  * @param buf_size length of the buffer
 608  * @param buf the bufffer
 609  *
 610  * @return the bitrate on success,
 611  *         I_F_Q  if the bitrate cannot be satisfactorily determined
 612  *
 613  * TIA/EIA/IS-733 2.4.8.7.1
 614  */
 615 static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, const int buf_size,
 616                              const uint8_t **buf)
 617 {
 618     qcelp_packet_rate bitrate;
 619
 620     if ((bitrate = buf_size2bitrate(buf_size)) >= 0) {
 621         if (bitrate > **buf) {
 622             QCELPContext *q = avctx->priv_data;
 623             if (!q->warned_buf_mismatch_bitrate) {
 624             av_log(avctx, AV_LOG_WARNING,
 625                    "Claimed bitrate and buffer size mismatch.\n");
 626                 q->warned_buf_mismatch_bitrate = 1;
 627             }
 628             bitrate = **buf;
 629         } else if (bitrate < **buf) {
 630             av_log(avctx, AV_LOG_ERROR,
 631                    "Buffer is too small for the claimed bitrate.\n");
 632             return I_F_Q;
 633         }
 634         (*buf)++;
 635     } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) {
 636         av_log(avctx, AV_LOG_WARNING,
 637                "Bitrate byte is missing, guessing the bitrate from packet size.\n");
 638     }else
 639         return I_F_Q;
 640
 641     if (bitrate == SILENCE) {
 642         //FIXME: Remove experimental warning when tested with samples.
 643         av_log_ask_for_sample(avctx, "'Blank frame handling is experimental.");
 644     }
 645     return bitrate;
 646 }
 647
 648 static void warn_insufficient_frame_quality(AVCodecContext *avctx,
 649                                             const char *message)
 650 {
 651     av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", avctx->frame_number,
 652            message);
 653 }
 654
 655 static void postfilter(QCELPContext *q, float *samples, float *lpc)
 656 {
 657     static const float pow_0_775[10] = {
 658         0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
 659         0.216676, 0.167924, 0.130141, 0.100859, 0.078166
 660     }, pow_0_625[10] = {
 661         0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
 662         0.059605, 0.037253, 0.023283, 0.014552, 0.009095
 663     };
 664     float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
 665     int n;
 666
 667     for (n = 0; n < 10; n++) {
 668         lpc_s[n] = lpc[n] * pow_0_625[n];
 669         lpc_p[n] = lpc[n] * pow_0_775[n];
 670     }
 671
 672     ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
 673                                       q->formant_mem + 10, 160, 10);
 674     memcpy(pole_out, q->postfilter_synth_mem,       sizeof(float) * 10);
 675     ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
 676     memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
 677
 678     ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
 679
 680     ff_adaptive_gain_control(samples, pole_out + 10,
 681         ff_dot_productf(q->formant_mem + 10, q->formant_mem + 10, 160),
 682         160, 0.9375, &q->postfilter_agc_mem);
 683 }
 684
 685 static int qcelp_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
 686                               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 = data;
 692     int   i, out_size;
 693     float quantized_lspf[10], lpc[10];
 694     float gain[16];
 695     float *formant_mem;
 696
 697     out_size = 160 * av_get_bytes_per_sample(avctx->sample_fmt);
 698     if (*data_size < out_size) {
 699         av_log(avctx, AV_LOG_ERROR, "Output buffer is too small\n");
 700         return AVERROR(EINVAL);
 701     }
 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) {
 715         const QCELPBitmap *bitmaps     = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
 716         const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate]
 717                                        + qcelp_unpacking_bitmaps_lengths[q->bitrate];
 718         uint8_t           *unpacked_data = (uint8_t *)&q->frame;
 719
 720         init_get_bits(&q->gb, buf, 8*buf_size);
 721
 722         memset(&q->frame, 0, sizeof(QCELPFrame));
 723
 724         for(; bitmaps < bitmaps_end; bitmaps++)
 725             unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
 726
 727         // Check for erasures/blanks on rates 1, 1/4 and 1/8.
 728         if (q->frame.reserved) {
 729             warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
 730             goto erasure;
 731         }
 732         if(q->bitrate == RATE_QUARTER &&
 733            codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) {
 734             warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
 735             goto erasure;
 736         }
 737
 738         if (q->bitrate >= RATE_HALF) {
 739             for (i = 0; i < 4; i++) {
 740                 if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) {
 741                     warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
 742                     goto erasure;
 743                 }
 744             }
 745         }
 746     }
 747
 748     decode_gain_and_index(q, gain);
 749     compute_svector(q, gain, outbuffer);
 750
 751     if (decode_lspf(q, quantized_lspf) < 0) {
 752         warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
 753         goto erasure;
 754     }
 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, outbuffer + i * 40, 40,
 774                                      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     *data_size = out_size;
 787
 788     return buf_size;
 789 }
 790
 791 AVCodec ff_qcelp_decoder =
 792 {
 793     .name   = "qcelp",
 794     .type   = AVMEDIA_TYPE_AUDIO,
 795     .id     = CODEC_ID_QCELP,
 796     .init   = qcelp_decode_init,
 797     .decode = qcelp_decode_frame,
 798     .priv_data_size = sizeof(QCELPContext),
 799     .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
 800 };