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 47typedef enum { 48 I_F_Q = -1, /**< insufficient frame quality */ 49 SILENCE, 50 RATE_OCTAVE, 51 RATE_QUARTER, 52 RATE_HALF, 53 RATE_FULL 54} qcelp_packet_rate; 55 56typedef struct { 57 AVFrame avframe; 58 GetBitContext gb; 59 qcelp_packet_rate bitrate; 60 QCELPFrame frame; /**< unpacked data frame */ 61 62 uint8_t erasure_count; 63 uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */ 64 float prev_lspf[10]; 65 float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */ 66 float pitch_synthesis_filter_mem[303]; 67 float pitch_pre_filter_mem[303]; 68 float rnd_fir_filter_mem[180]; 69 float formant_mem[170]; 70 float last_codebook_gain; 71 int prev_g1[2]; 72 int prev_bitrate; 73 float pitch_gain[4]; 74 uint8_t pitch_lag[4]; 75 uint16_t first16bits; 76 uint8_t warned_buf_mismatch_bitrate; 77 78 /* postfilter */ 79 float postfilter_synth_mem[10]; 80 float postfilter_agc_mem; 81 float postfilter_tilt_mem; 82} QCELPContext; 83 84/** 85 * Initialize the speech codec according to the specification. 86 * 87 * TIA/EIA/IS-733 2.4.9 88 */ 89static av_cold int qcelp_decode_init(AVCodecContext *avctx) 90{ 91 QCELPContext *q = avctx->priv_data; 92 int i; 93 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 avcodec_get_frame_defaults(&q->avframe); 100 avctx->coded_frame = &q->avframe; 101 102 return 0; 103} 104 105/** 106 * Decode the 10 quantized LSP frequencies from the LSPV/LSP 107 * transmission codes of any bitrate and check for badly received packets. 108 * 109 * @param q the context 110 * @param lspf line spectral pair frequencies 111 * 112 * @return 0 on success, -1 if the packet is badly received 113 * 114 * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3 115 */ 116static int decode_lspf(QCELPContext *q, float *lspf) 117{ 118 int i; 119 float tmp_lspf, smooth, erasure_coeff; 120 const float *predictors; 121 122 if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) { 123 predictors = q->prev_bitrate != RATE_OCTAVE && 124 q->prev_bitrate != I_F_Q ? q->prev_lspf 125 : q->predictor_lspf; 126 127 if (q->bitrate == RATE_OCTAVE) { 128 q->octave_count++; 129 130 for (i = 0; i < 10; i++) { 131 q->predictor_lspf[i] = 132 lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR 133 : -QCELP_LSP_SPREAD_FACTOR) + 134 predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR + 135 (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11); 136 } 137 smooth = q->octave_count < 10 ? .875 : 0.1; 138 } else { 139 erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR; 140 141 assert(q->bitrate == I_F_Q); 142 143 if (q->erasure_count > 1) 144 erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7; 145 146 for (i = 0; i < 10; i++) { 147 q->predictor_lspf[i] = 148 lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 + 149 erasure_coeff * predictors[i]; 150 } 151 smooth = 0.125; 152 } 153 154 // Check the stability of the LSP frequencies. 155 lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR); 156 for (i = 1; i < 10; i++) 157 lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR); 158 159 lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR); 160 for (i = 9; i > 0; i--) 161 lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR); 162 163 // Low-pass filter the LSP frequencies. 164 ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10); 165 } else { 166 q->octave_count = 0; 167 168 tmp_lspf = 0.; 169 for (i = 0; i < 5; i++) { 170 lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001; 171 lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001; 172 } 173 174 // Check for badly received packets. 175 if (q->bitrate == RATE_QUARTER) { 176 if (lspf[9] <= .70 || lspf[9] >= .97) 177 return -1; 178 for (i = 3; i < 10; i++) 179 if (fabs(lspf[i] - lspf[i - 2]) < .08) 180 return -1; 181 } else { 182 if (lspf[9] <= .66 || lspf[9] >= .985) 183 return -1; 184 for (i = 4; i < 10; i++) 185 if (fabs(lspf[i] - lspf[i - 4]) < .0931) 186 return -1; 187 } 188 } 189 return 0; 190} 191 192/** 193 * Convert codebook transmission codes to GAIN and INDEX. 194 * 195 * @param q the context 196 * @param gain array holding the decoded gain 197 * 198 * TIA/EIA/IS-733 2.4.6.2 199 */ 200static void decode_gain_and_index(QCELPContext *q, float *gain) 201{ 202 int i, subframes_count, g1[16]; 203 float slope; 204 205 if (q->bitrate >= RATE_QUARTER) { 206 switch (q->bitrate) { 207 case RATE_FULL: subframes_count = 16; break; 208 case RATE_HALF: subframes_count = 4; break; 209 default: subframes_count = 5; 210 } 211 for (i = 0; i < subframes_count; i++) { 212 g1[i] = 4 * q->frame.cbgain[i]; 213 if (q->bitrate == RATE_FULL && !((i + 1) & 3)) { 214 g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32); 215 } 216 217 gain[i] = qcelp_g12ga[g1[i]]; 218 219 if (q->frame.cbsign[i]) { 220 gain[i] = -gain[i]; 221 q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127; 222 } 223 } 224 225 q->prev_g1[0] = g1[i - 2]; 226 q->prev_g1[1] = g1[i - 1]; 227 q->last_codebook_gain = qcelp_g12ga[g1[i - 1]]; 228 229 if (q->bitrate == RATE_QUARTER) { 230 // Provide smoothing of the unvoiced excitation energy. 231 gain[7] = gain[4]; 232 gain[6] = 0.4 * gain[3] + 0.6 * gain[4]; 233 gain[5] = gain[3]; 234 gain[4] = 0.8 * gain[2] + 0.2 * gain[3]; 235 gain[3] = 0.2 * gain[1] + 0.8 * gain[2]; 236 gain[2] = gain[1]; 237 gain[1] = 0.6 * gain[0] + 0.4 * gain[1]; 238 } 239 } else if (q->bitrate != SILENCE) { 240 if (q->bitrate == RATE_OCTAVE) { 241 g1[0] = 2 * q->frame.cbgain[0] + 242 av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54); 243 subframes_count = 8; 244 } else { 245 assert(q->bitrate == I_F_Q); 246 247 g1[0] = q->prev_g1[1]; 248 switch (q->erasure_count) { 249 case 1 : break; 250 case 2 : g1[0] -= 1; break; 251 case 3 : g1[0] -= 2; break; 252 default: g1[0] -= 6; 253 } 254 if (g1[0] < 0) 255 g1[0] = 0; 256 subframes_count = 4; 257 } 258 // This interpolation is done to produce smoother background noise. 259 slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count; 260 for (i = 1; i <= subframes_count; i++) 261 gain[i - 1] = q->last_codebook_gain + slope * i; 262 263 q->last_codebook_gain = gain[i - 2]; 264 q->prev_g1[0] = q->prev_g1[1]; 265 q->prev_g1[1] = g1[0]; 266 } 267} 268 269/** 270 * If the received packet is Rate 1/4 a further sanity check is made of the 271 * codebook gain. 272 * 273 * @param cbgain the unpacked cbgain array 274 * @return -1 if the sanity check fails, 0 otherwise 275 * 276 * TIA/EIA/IS-733 2.4.8.7.3 277 */ 278static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain) 279{ 280 int i, diff, prev_diff = 0; 281 282 for (i = 1; i < 5; i++) { 283 diff = cbgain[i] - cbgain[i-1]; 284 if (FFABS(diff) > 10) 285 return -1; 286 else if (FFABS(diff - prev_diff) > 12) 287 return -1; 288 prev_diff = diff; 289 } 290 return 0; 291} 292 293/** 294 * Compute the scaled codebook vector Cdn From INDEX and GAIN 295 * for all rates. 296 * 297 * The specification lacks some information here. 298 * 299 * TIA/EIA/IS-733 has an omission on the codebook index determination 300 * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says 301 * you have to subtract the decoded index parameter from the given scaled 302 * codebook vector index 'n' to get the desired circular codebook index, but 303 * it does not mention that you have to clamp 'n' to [0-9] in order to get 304 * RI-compliant results. 305 * 306 * The reason for this mistake seems to be the fact they forgot to mention you 307 * have to do these calculations per codebook subframe and adjust given 308 * equation values accordingly. 309 * 310 * @param q the context 311 * @param gain array holding the 4 pitch subframe gain values 312 * @param cdn_vector array for the generated scaled codebook vector 313 */ 314static void compute_svector(QCELPContext *q, const float *gain, 315 float *cdn_vector) 316{ 317 int i, j, k; 318 uint16_t cbseed, cindex; 319 float *rnd, tmp_gain, fir_filter_value; 320 321 switch (q->bitrate) { 322 case RATE_FULL: 323 for (i = 0; i < 16; i++) { 324 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO; 325 cindex = -q->frame.cindex[i]; 326 for (j = 0; j < 10; j++) 327 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127]; 328 } 329 break; 330 case RATE_HALF: 331 for (i = 0; i < 4; i++) { 332 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO; 333 cindex = -q->frame.cindex[i]; 334 for (j = 0; j < 40; j++) 335 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127]; 336 } 337 break; 338 case RATE_QUARTER: 339 cbseed = (0x0003 & q->frame.lspv[4]) << 14 | 340 (0x003F & q->frame.lspv[3]) << 8 | 341 (0x0060 & q->frame.lspv[2]) << 1 | 342 (0x0007 & q->frame.lspv[1]) << 3 | 343 (0x0038 & q->frame.lspv[0]) >> 3; 344 rnd = q->rnd_fir_filter_mem + 20; 345 for (i = 0; i < 8; i++) { 346 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0); 347 for (k = 0; k < 20; k++) { 348 cbseed = 521 * cbseed + 259; 349 *rnd = (int16_t) cbseed; 350 351 // FIR filter 352 fir_filter_value = 0.0; 353 for (j = 0; j < 10; j++) 354 fir_filter_value += qcelp_rnd_fir_coefs[j] * 355 (rnd[-j] + rnd[-20+j]); 356 357 fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10]; 358 *cdn_vector++ = tmp_gain * fir_filter_value; 359 rnd++; 360 } 361 } 362 memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, 363 20 * sizeof(float)); 364 break; 365 case RATE_OCTAVE: 366 cbseed = q->first16bits; 367 for (i = 0; i < 8; i++) { 368 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0); 369 for (j = 0; j < 20; j++) { 370 cbseed = 521 * cbseed + 259; 371 *cdn_vector++ = tmp_gain * (int16_t) cbseed; 372 } 373 } 374 break; 375 case I_F_Q: 376 cbseed = -44; // random codebook index 377 for (i = 0; i < 4; i++) { 378 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO; 379 for (j = 0; j < 40; j++) 380 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127]; 381 } 382 break; 383 case SILENCE: 384 memset(cdn_vector, 0, 160 * sizeof(float)); 385 break; 386 } 387} 388 389/** 390 * Apply generic gain control. 391 * 392 * @param v_out output vector 393 * @param v_in gain-controlled vector 394 * @param v_ref vector to control gain of 395 * 396 * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6 397 */ 398static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in) 399{ 400 int i; 401 402 for (i = 0; i < 160; i += 40) 403 ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, 404 ff_dot_productf(v_ref + i, 405 v_ref + i, 40), 406 40); 407} 408 409/** 410 * Apply filter in pitch-subframe steps. 411 * 412 * @param memory buffer for the previous state of the filter 413 * - must be able to contain 303 elements 414 * - the 143 first elements are from the previous state 415 * - the next 160 are for output 416 * @param v_in input filter vector 417 * @param gain per-subframe gain array, each element is between 0.0 and 2.0 418 * @param lag per-subframe lag array, each element is 419 * - between 16 and 143 if its corresponding pfrac is 0, 420 * - between 16 and 139 otherwise 421 * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0 422 * otherwise 423 * 424 * @return filter output vector 425 */ 426static const float *do_pitchfilter(float memory[303], const float v_in[160], 427 const float gain[4], const uint8_t *lag, 428 const uint8_t pfrac[4]) 429{ 430 int i, j; 431 float *v_lag, *v_out; 432 const float *v_len; 433 434 v_out = memory + 143; // Output vector starts at memory[143]. 435 436 for (i = 0; i < 4; i++) { 437 if (gain[i]) { 438 v_lag = memory + 143 + 40 * i - lag[i]; 439 for (v_len = v_in + 40; v_in < v_len; v_in++) { 440 if (pfrac[i]) { // If it is a fractional lag... 441 for (j = 0, *v_out = 0.; j < 4; j++) 442 *v_out += qcelp_hammsinc_table[j] * (v_lag[j - 4] + v_lag[3 - j]); 443 } else 444 *v_out = *v_lag; 445 446 *v_out = *v_in + gain[i] * *v_out; 447 448 v_lag++; 449 v_out++; 450 } 451 } else { 452 memcpy(v_out, v_in, 40 * sizeof(float)); 453 v_in += 40; 454 v_out += 40; 455 } 456 } 457 458 memmove(memory, memory + 160, 143 * sizeof(float)); 459 return memory + 143; 460} 461 462/** 463 * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector. 464 * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2 465 * 466 * @param q the context 467 * @param cdn_vector the scaled codebook vector 468 */ 469static void apply_pitch_filters(QCELPContext *q, float *cdn_vector) 470{ 471 int i; 472 const float *v_synthesis_filtered, *v_pre_filtered; 473 474 if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE || 475 (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) { 476 477 if (q->bitrate >= RATE_HALF) { 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, 143 * sizeof(float)); 519 memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float)); 520 memset(q->pitch_gain, 0, sizeof(q->pitch_gain)); 521 memset(q->pitch_lag, 0, sizeof(q->pitch_lag)); 522 } 523} 524 525/** 526 * Reconstruct LPC coefficients from the line spectral pair frequencies 527 * and perform bandwidth expansion. 528 * 529 * @param lspf line spectral pair frequencies 530 * @param lpc linear predictive coding coefficients 531 * 532 * @note: bandwidth_expansion_coeff could be precalculated into a table 533 * but it seems to be slower on x86 534 * 535 * TIA/EIA/IS-733 2.4.3.3.5 536 */ 537static void lspf2lpc(const float *lspf, float *lpc) 538{ 539 double lsp[10]; 540 double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF; 541 int i; 542 543 for (i = 0; i < 10; i++) 544 lsp[i] = cos(M_PI * lspf[i]); 545 546 ff_acelp_lspd2lpc(lsp, lpc, 5); 547 548 for (i = 0; i < 10; i++) { 549 lpc[i] *= bandwidth_expansion_coeff; 550 bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF; 551 } 552} 553 554/** 555 * Interpolate LSP frequencies and compute LPC coefficients 556 * for a given bitrate & pitch subframe. 557 * 558 * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2 559 * 560 * @param q the context 561 * @param curr_lspf LSP frequencies vector of the current frame 562 * @param lpc float vector for the resulting LPC 563 * @param subframe_num frame number in decoded stream 564 */ 565static void interpolate_lpc(QCELPContext *q, const float *curr_lspf, 566 float *lpc, const int subframe_num) 567{ 568 float interpolated_lspf[10]; 569 float weight; 570 571 if (q->bitrate >= RATE_QUARTER) 572 weight = 0.25 * (subframe_num + 1); 573 else if (q->bitrate == RATE_OCTAVE && !subframe_num) 574 weight = 0.625; 575 else 576 weight = 1.0; 577 578 if (weight != 1.0) { 579 ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf, 580 weight, 1.0 - weight, 10); 581 lspf2lpc(interpolated_lspf, lpc); 582 } else if (q->bitrate >= RATE_QUARTER || 583 (q->bitrate == I_F_Q && !subframe_num)) 584 lspf2lpc(curr_lspf, lpc); 585 else if (q->bitrate == SILENCE && !subframe_num) 586 lspf2lpc(q->prev_lspf, lpc); 587} 588 589static qcelp_packet_rate buf_size2bitrate(const int buf_size) 590{ 591 switch (buf_size) { 592 case 35: return RATE_FULL; 593 case 17: return RATE_HALF; 594 case 8: return RATE_QUARTER; 595 case 4: return RATE_OCTAVE; 596 case 1: return SILENCE; 597 } 598 599 return I_F_Q; 600} 601 602/** 603 * Determine the bitrate from the frame size and/or the first byte of the frame. 604 * 605 * @param avctx the AV codec context 606 * @param buf_size length of the buffer 607 * @param buf the bufffer 608 * 609 * @return the bitrate on success, 610 * I_F_Q if the bitrate cannot be satisfactorily determined 611 * 612 * TIA/EIA/IS-733 2.4.8.7.1 613 */ 614static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, 615 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 648static void warn_insufficient_frame_quality(AVCodecContext *avctx, 649 const char *message) 650{ 651 av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", 652 avctx->frame_number, message); 653} 654 655static 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, 682 q->formant_mem + 10, 160), 683 160, 0.9375, &q->postfilter_agc_mem); 684} 685 686static int qcelp_decode_frame(AVCodecContext *avctx, void *data, 687 int *got_frame_ptr, AVPacket *avpkt) 688{ 689 const uint8_t *buf = avpkt->data; 690 int buf_size = avpkt->size; 691 QCELPContext *q = avctx->priv_data; 692 float *outbuffer; 693 int i, ret; 694 float quantized_lspf[10], lpc[10]; 695 float gain[16]; 696 float *formant_mem; 697 698 /* get output buffer */ 699 q->avframe.nb_samples = 160; 700 if ((ret = avctx->get_buffer(avctx, &q->avframe)) < 0) { 701 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n"); 702 return ret; 703 } 704 outbuffer = (float *)q->avframe.data[0]; 705 706 if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) { 707 warn_insufficient_frame_quality(avctx, "bitrate cannot be determined."); 708 goto erasure; 709 } 710 711 if (q->bitrate == RATE_OCTAVE && 712 (q->first16bits = AV_RB16(buf)) == 0xFFFF) { 713 warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on."); 714 goto erasure; 715 } 716 717 if (q->bitrate > SILENCE) { 718 const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate]; 719 const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] + 720 qcelp_unpacking_bitmaps_lengths[q->bitrate]; 721 uint8_t *unpacked_data = (uint8_t *)&q->frame; 722 723 init_get_bits(&q->gb, buf, 8 * buf_size); 724 725 memset(&q->frame, 0, sizeof(QCELPFrame)); 726 727 for (; bitmaps < bitmaps_end; bitmaps++) 728 unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos; 729 730 // Check for erasures/blanks on rates 1, 1/4 and 1/8. 731 if (q->frame.reserved) { 732 warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area."); 733 goto erasure; 734 } 735 if (q->bitrate == RATE_QUARTER && 736 codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) { 737 warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed."); 738 goto erasure; 739 } 740 741 if (q->bitrate >= RATE_HALF) { 742 for (i = 0; i < 4; i++) { 743 if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) { 744 warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter."); 745 goto erasure; 746 } 747 } 748 } 749 } 750 751 decode_gain_and_index(q, gain); 752 compute_svector(q, gain, outbuffer); 753 754 if (decode_lspf(q, quantized_lspf) < 0) { 755 warn_insufficient_frame_quality(avctx, "Badly received packets in frame."); 756 goto erasure; 757 } 758 759 apply_pitch_filters(q, outbuffer); 760 761 if (q->bitrate == I_F_Q) { 762erasure: 763 q->bitrate = I_F_Q; 764 q->erasure_count++; 765 decode_gain_and_index(q, gain); 766 compute_svector(q, gain, outbuffer); 767 decode_lspf(q, quantized_lspf); 768 apply_pitch_filters(q, outbuffer); 769 } else 770 q->erasure_count = 0; 771 772 formant_mem = q->formant_mem + 10; 773 for (i = 0; i < 4; i++) { 774 interpolate_lpc(q, quantized_lspf, lpc, i); 775 ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40, 10); 776 formant_mem += 40; 777 } 778 779 // postfilter, as per TIA/EIA/IS-733 2.4.8.6 780 postfilter(q, outbuffer, lpc); 781 782 memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float)); 783 784 memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf)); 785 q->prev_bitrate = q->bitrate; 786 787 *got_frame_ptr = 1; 788 *(AVFrame *)data = q->avframe; 789 790 return buf_size; 791} 792 793AVCodec ff_qcelp_decoder = { 794 .name = "qcelp", 795 .type = AVMEDIA_TYPE_AUDIO, 796 .id = CODEC_ID_QCELP, 797 .init = qcelp_decode_init, 798 .decode = qcelp_decode_frame, 799 .capabilities = CODEC_CAP_DR1, 800 .priv_data_size = sizeof(QCELPContext), 801 .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"), 802}; 803