1/* 2 * AMR narrowband decoder 3 * Copyright (c) 2006-2007 Robert Swain 4 * Copyright (c) 2009 Colin McQuillan 5 * 6 * This file is part of FFmpeg. 7 * 8 * FFmpeg is free software; you can redistribute it and/or 9 * modify it under the terms of the GNU Lesser General Public 10 * License as published by the Free Software Foundation; either 11 * version 2.1 of the License, or (at your option) any later version. 12 * 13 * FFmpeg is distributed in the hope that it will be useful, 14 * but WITHOUT ANY WARRANTY; without even the implied warranty of 15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 16 * Lesser General Public License for more details. 17 * 18 * You should have received a copy of the GNU Lesser General Public 19 * License along with FFmpeg; if not, write to the Free Software 20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 21 */ 22 23 24/** 25 * @file 26 * AMR narrowband decoder 27 * 28 * This decoder uses floats for simplicity and so is not bit-exact. One 29 * difference is that differences in phase can accumulate. The test sequences 30 * in 3GPP TS 26.074 can still be useful. 31 * 32 * - Comparing this file's output to the output of the ref decoder gives a 33 * PSNR of 30 to 80. Plotting the output samples shows a difference in 34 * phase in some areas. 35 * 36 * - Comparing both decoders against their input, this decoder gives a similar 37 * PSNR. If the test sequence homing frames are removed (this decoder does 38 * not detect them), the PSNR is at least as good as the reference on 140 39 * out of 169 tests. 40 */ 41 42 43#include <string.h> 44#include <math.h> 45 46#include "avcodec.h" 47#include "get_bits.h" 48#include "libavutil/common.h" 49#include "celp_math.h" 50#include "celp_filters.h" 51#include "acelp_filters.h" 52#include "acelp_vectors.h" 53#include "acelp_pitch_delay.h" 54#include "lsp.h" 55 56#include "amrnbdata.h" 57 58#define AMR_BLOCK_SIZE 160 ///< samples per frame 59#define AMR_SAMPLE_BOUND 32768.0 ///< threshold for synthesis overflow 60 61/** 62 * Scale from constructed speech to [-1,1] 63 * 64 * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but 65 * upscales by two (section 6.2.2). 66 * 67 * Fundamentally, this scale is determined by energy_mean through 68 * the fixed vector contribution to the excitation vector. 69 */ 70#define AMR_SAMPLE_SCALE (2.0 / 32768.0) 71 72/** Prediction factor for 12.2kbit/s mode */ 73#define PRED_FAC_MODE_12k2 0.65 74 75#define LSF_R_FAC (8000.0 / 32768.0) ///< LSF residual tables to Hertz 76#define MIN_LSF_SPACING (50.0488 / 8000.0) ///< Ensures stability of LPC filter 77#define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode 78 79/** Initial energy in dB. Also used for bad frames (unimplemented). */ 80#define MIN_ENERGY -14.0 81 82/** Maximum sharpening factor 83 * 84 * The specification says 0.8, which should be 13107, but the reference C code 85 * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in g729dec.c.) 86 */ 87#define SHARP_MAX 0.79449462890625 88 89/** Number of impulse response coefficients used for tilt factor */ 90#define AMR_TILT_RESPONSE 22 91/** Tilt factor = 1st reflection coefficient * gamma_t */ 92#define AMR_TILT_GAMMA_T 0.8 93/** Adaptive gain control factor used in post-filter */ 94#define AMR_AGC_ALPHA 0.9 95 96typedef struct AMRContext { 97 AMRNBFrame frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc) 98 uint8_t bad_frame_indicator; ///< bad frame ? 1 : 0 99 enum Mode cur_frame_mode; 100 101 int16_t prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe 102 double lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame 103 double prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame 104 105 float lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing 106 float lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector 107 108 float lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes 109 110 uint8_t pitch_lag_int; ///< integer part of pitch lag from current subframe 111 112 float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history 113 float *excitation; ///< pointer to the current excitation vector in excitation_buf 114 115 float pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector 116 float fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames) 117 118 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes 119 float pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes 120 float fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes 121 122 float beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX] 123 uint8_t diff_count; ///< the number of subframes for which diff has been above 0.65 124 uint8_t hang_count; ///< the number of subframes since a hangover period started 125 126 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset" 127 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none 128 uint8_t ir_filter_onset; ///< flag for impulse response filter strength 129 130 float postfilter_mem[10]; ///< previous intermediate values in the formant filter 131 float tilt_mem; ///< previous input to tilt compensation filter 132 float postfilter_agc; ///< previous factor used for adaptive gain control 133 float high_pass_mem[2]; ///< previous intermediate values in the high-pass filter 134 135 float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples 136 137} AMRContext; 138 139/** Double version of ff_weighted_vector_sumf() */ 140static void weighted_vector_sumd(double *out, const double *in_a, 141 const double *in_b, double weight_coeff_a, 142 double weight_coeff_b, int length) 143{ 144 int i; 145 146 for (i = 0; i < length; i++) 147 out[i] = weight_coeff_a * in_a[i] 148 + weight_coeff_b * in_b[i]; 149} 150 151static av_cold int amrnb_decode_init(AVCodecContext *avctx) 152{ 153 AMRContext *p = avctx->priv_data; 154 int i; 155 156 avctx->sample_fmt = SAMPLE_FMT_FLT; 157 158 // p->excitation always points to the same position in p->excitation_buf 159 p->excitation = &p->excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1]; 160 161 for (i = 0; i < LP_FILTER_ORDER; i++) { 162 p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15); 163 p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15); 164 } 165 166 for (i = 0; i < 4; i++) 167 p->prediction_error[i] = MIN_ENERGY; 168 169 return 0; 170} 171 172 173/** 174 * Unpack an RFC4867 speech frame into the AMR frame mode and parameters. 175 * 176 * The order of speech bits is specified by 3GPP TS 26.101. 177 * 178 * @param p the context 179 * @param buf pointer to the input buffer 180 * @param buf_size size of the input buffer 181 * 182 * @return the frame mode 183 */ 184static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf, 185 int buf_size) 186{ 187 GetBitContext gb; 188 enum Mode mode; 189 190 init_get_bits(&gb, buf, buf_size * 8); 191 192 // Decode the first octet. 193 skip_bits(&gb, 1); // padding bit 194 mode = get_bits(&gb, 4); // frame type 195 p->bad_frame_indicator = !get_bits1(&gb); // quality bit 196 skip_bits(&gb, 2); // two padding bits 197 198 if (mode < MODE_DTX) { 199 uint16_t *data = (uint16_t *)&p->frame; 200 const uint8_t *order = amr_unpacking_bitmaps_per_mode[mode]; 201 int field_size; 202 203 memset(&p->frame, 0, sizeof(AMRNBFrame)); 204 buf++; 205 while ((field_size = *order++)) { 206 int field = 0; 207 int field_offset = *order++; 208 while (field_size--) { 209 int bit = *order++; 210 field <<= 1; 211 field |= buf[bit >> 3] >> (bit & 7) & 1; 212 } 213 data[field_offset] = field; 214 } 215 } 216 217 return mode; 218} 219 220 221/// @defgroup amr_lpc_decoding AMR pitch LPC coefficient decoding functions 222/// @{ 223 224/** 225 * Convert an lsf vector into an lsp vector. 226 * 227 * @param lsf input lsf vector 228 * @param lsp output lsp vector 229 */ 230static void lsf2lsp(const float *lsf, double *lsp) 231{ 232 int i; 233 234 for (i = 0; i < LP_FILTER_ORDER; i++) 235 lsp[i] = cos(2.0 * M_PI * lsf[i]); 236} 237 238/** 239 * Interpolate the LSF vector (used for fixed gain smoothing). 240 * The interpolation is done over all four subframes even in MODE_12k2. 241 * 242 * @param[in,out] lsf_q LSFs in [0,1] for each subframe 243 * @param[in] lsf_new New LSFs in [0,1] for subframe 4 244 */ 245static void interpolate_lsf(float lsf_q[4][LP_FILTER_ORDER], float *lsf_new) 246{ 247 int i; 248 249 for (i = 0; i < 4; i++) 250 ff_weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new, 251 0.25 * (3 - i), 0.25 * (i + 1), 252 LP_FILTER_ORDER); 253} 254 255/** 256 * Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector. 257 * 258 * @param p the context 259 * @param lsp output LSP vector 260 * @param lsf_no_r LSF vector without the residual vector added 261 * @param lsf_quantizer pointers to LSF dictionary tables 262 * @param quantizer_offset offset in tables 263 * @param sign for the 3 dictionary table 264 * @param update store data for computing the next frame's LSFs 265 */ 266static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER], 267 const float lsf_no_r[LP_FILTER_ORDER], 268 const int16_t *lsf_quantizer[5], 269 const int quantizer_offset, 270 const int sign, const int update) 271{ 272 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector 273 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector 274 int i; 275 276 for (i = 0; i < LP_FILTER_ORDER >> 1; i++) 277 memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset], 278 2 * sizeof(*lsf_r)); 279 280 if (sign) { 281 lsf_r[4] *= -1; 282 lsf_r[5] *= -1; 283 } 284 285 if (update) 286 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(float)); 287 288 for (i = 0; i < LP_FILTER_ORDER; i++) 289 lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0); 290 291 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER); 292 293 if (update) 294 interpolate_lsf(p->lsf_q, lsf_q); 295 296 lsf2lsp(lsf_q, lsp); 297} 298 299/** 300 * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors. 301 * 302 * @param p pointer to the AMRContext 303 */ 304static void lsf2lsp_5(AMRContext *p) 305{ 306 const uint16_t *lsf_param = p->frame.lsf; 307 float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector 308 const int16_t *lsf_quantizer[5]; 309 int i; 310 311 lsf_quantizer[0] = lsf_5_1[lsf_param[0]]; 312 lsf_quantizer[1] = lsf_5_2[lsf_param[1]]; 313 lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1]; 314 lsf_quantizer[3] = lsf_5_4[lsf_param[3]]; 315 lsf_quantizer[4] = lsf_5_5[lsf_param[4]]; 316 317 for (i = 0; i < LP_FILTER_ORDER; i++) 318 lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i]; 319 320 lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0); 321 lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1); 322 323 // interpolate LSP vectors at subframes 1 and 3 324 weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER); 325 weighted_vector_sumd(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER); 326} 327 328/** 329 * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector. 330 * 331 * @param p pointer to the AMRContext 332 */ 333static void lsf2lsp_3(AMRContext *p) 334{ 335 const uint16_t *lsf_param = p->frame.lsf; 336 int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector 337 float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector 338 const int16_t *lsf_quantizer; 339 int i, j; 340 341 lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]]; 342 memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r)); 343 344 lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)]; 345 memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r)); 346 347 lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]]; 348 memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r)); 349 350 // calculate mean-removed LSF vector and add mean 351 for (i = 0; i < LP_FILTER_ORDER; i++) 352 lsf_q[i] = (lsf_r[i] + p->prev_lsf_r[i] * pred_fac[i]) * (LSF_R_FAC / 8000.0) + lsf_3_mean[i] * (1.0 / 8000.0); 353 354 ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER); 355 356 // store data for computing the next frame's LSFs 357 interpolate_lsf(p->lsf_q, lsf_q); 358 memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r)); 359 360 lsf2lsp(lsf_q, p->lsp[3]); 361 362 // interpolate LSP vectors at subframes 1, 2 and 3 363 for (i = 1; i <= 3; i++) 364 for(j = 0; j < LP_FILTER_ORDER; j++) 365 p->lsp[i-1][j] = p->prev_lsp_sub4[j] + 366 (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i; 367} 368 369/// @} 370 371 372/// @defgroup amr_pitch_vector_decoding AMR pitch vector decoding functions 373/// @{ 374 375/** 376 * Like ff_decode_pitch_lag(), but with 1/6 resolution 377 */ 378static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index, 379 const int prev_lag_int, const int subframe) 380{ 381 if (subframe == 0 || subframe == 2) { 382 if (pitch_index < 463) { 383 *lag_int = (pitch_index + 107) * 10923 >> 16; 384 *lag_frac = pitch_index - *lag_int * 6 + 105; 385 } else { 386 *lag_int = pitch_index - 368; 387 *lag_frac = 0; 388 } 389 } else { 390 *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1; 391 *lag_frac = pitch_index - *lag_int * 6 - 3; 392 *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2, 393 PITCH_DELAY_MAX - 9); 394 } 395} 396 397static void decode_pitch_vector(AMRContext *p, 398 const AMRNBSubframe *amr_subframe, 399 const int subframe) 400{ 401 int pitch_lag_int, pitch_lag_frac; 402 enum Mode mode = p->cur_frame_mode; 403 404 if (p->cur_frame_mode == MODE_12k2) { 405 decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac, 406 amr_subframe->p_lag, p->pitch_lag_int, 407 subframe); 408 } else 409 ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac, 410 amr_subframe->p_lag, 411 p->pitch_lag_int, subframe, 412 mode != MODE_4k75 && mode != MODE_5k15, 413 mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6)); 414 415 p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t 416 417 pitch_lag_frac <<= (p->cur_frame_mode != MODE_12k2); 418 419 pitch_lag_int += pitch_lag_frac > 0; 420 421 /* Calculate the pitch vector by interpolating the past excitation at the 422 pitch lag using a b60 hamming windowed sinc function. */ 423 ff_acelp_interpolatef(p->excitation, p->excitation + 1 - pitch_lag_int, 424 ff_b60_sinc, 6, 425 pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0), 426 10, AMR_SUBFRAME_SIZE); 427 428 memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float)); 429} 430 431/// @} 432 433 434/// @defgroup amr_algebraic_code_book AMR algebraic code book (fixed) vector decoding functions 435/// @{ 436 437/** 438 * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame. 439 */ 440static void decode_10bit_pulse(int code, int pulse_position[8], 441 int i1, int i2, int i3) 442{ 443 // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of 444 // the 3 pulses and the upper 7 bits being coded in base 5 445 const uint8_t *positions = base_five_table[code >> 3]; 446 pulse_position[i1] = (positions[2] << 1) + ( code & 1); 447 pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1); 448 pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1); 449} 450 451/** 452 * Decode the algebraic codebook index to pulse positions and signs and 453 * construct the algebraic codebook vector for MODE_10k2. 454 * 455 * @param fixed_index positions of the eight pulses 456 * @param fixed_sparse pointer to the algebraic codebook vector 457 */ 458static void decode_8_pulses_31bits(const int16_t *fixed_index, 459 AMRFixed *fixed_sparse) 460{ 461 int pulse_position[8]; 462 int i, temp; 463 464 decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1); 465 decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5); 466 467 // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of 468 // the 2 pulses and the upper 5 bits being coded in base 5 469 temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5; 470 pulse_position[3] = temp % 5; 471 pulse_position[7] = temp / 5; 472 if (pulse_position[7] & 1) 473 pulse_position[3] = 4 - pulse_position[3]; 474 pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1); 475 pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1); 476 477 fixed_sparse->n = 8; 478 for (i = 0; i < 4; i++) { 479 const int pos1 = (pulse_position[i] << 2) + i; 480 const int pos2 = (pulse_position[i + 4] << 2) + i; 481 const float sign = fixed_index[i] ? -1.0 : 1.0; 482 fixed_sparse->x[i ] = pos1; 483 fixed_sparse->x[i + 4] = pos2; 484 fixed_sparse->y[i ] = sign; 485 fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign; 486 } 487} 488 489/** 490 * Decode the algebraic codebook index to pulse positions and signs, 491 * then construct the algebraic codebook vector. 492 * 493 * nb of pulses | bits encoding pulses 494 * For MODE_4k75 or MODE_5k15, 2 | 1-3, 4-6, 7 495 * MODE_5k9, 2 | 1, 2-4, 5-6, 7-9 496 * MODE_6k7, 3 | 1-3, 4, 5-7, 8, 9-11 497 * MODE_7k4 or MODE_7k95, 4 | 1-3, 4-6, 7-9, 10, 11-13 498 * 499 * @param fixed_sparse pointer to the algebraic codebook vector 500 * @param pulses algebraic codebook indexes 501 * @param mode mode of the current frame 502 * @param subframe current subframe number 503 */ 504static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses, 505 const enum Mode mode, const int subframe) 506{ 507 assert(MODE_4k75 <= mode && mode <= MODE_12k2); 508 509 if (mode == MODE_12k2) { 510 ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3); 511 } else if (mode == MODE_10k2) { 512 decode_8_pulses_31bits(pulses, fixed_sparse); 513 } else { 514 int *pulse_position = fixed_sparse->x; 515 int i, pulse_subset; 516 const int fixed_index = pulses[0]; 517 518 if (mode <= MODE_5k15) { 519 pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1); 520 pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset]; 521 pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1]; 522 fixed_sparse->n = 2; 523 } else if (mode == MODE_5k9) { 524 pulse_subset = ((fixed_index & 1) << 1) + 1; 525 pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset; 526 pulse_subset = (fixed_index >> 4) & 3; 527 pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0); 528 fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2; 529 } else if (mode == MODE_6k7) { 530 pulse_position[0] = (fixed_index & 7) * 5; 531 pulse_subset = (fixed_index >> 2) & 2; 532 pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1; 533 pulse_subset = (fixed_index >> 6) & 2; 534 pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2; 535 fixed_sparse->n = 3; 536 } else { // mode <= MODE_7k95 537 pulse_position[0] = gray_decode[ fixed_index & 7]; 538 pulse_position[1] = gray_decode[(fixed_index >> 3) & 7] + 1; 539 pulse_position[2] = gray_decode[(fixed_index >> 6) & 7] + 2; 540 pulse_subset = (fixed_index >> 9) & 1; 541 pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3; 542 fixed_sparse->n = 4; 543 } 544 for (i = 0; i < fixed_sparse->n; i++) 545 fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0; 546 } 547} 548 549/** 550 * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2) 551 * 552 * @param p the context 553 * @param subframe unpacked amr subframe 554 * @param mode mode of the current frame 555 * @param fixed_sparse sparse respresentation of the fixed vector 556 */ 557static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode, 558 AMRFixed *fixed_sparse) 559{ 560 // The spec suggests the current pitch gain is always used, but in other 561 // modes the pitch and codebook gains are joinly quantized (sec 5.8.2) 562 // so the codebook gain cannot depend on the quantized pitch gain. 563 if (mode == MODE_12k2) 564 p->beta = FFMIN(p->pitch_gain[4], 1.0); 565 566 fixed_sparse->pitch_lag = p->pitch_lag_int; 567 fixed_sparse->pitch_fac = p->beta; 568 569 // Save pitch sharpening factor for the next subframe 570 // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from 571 // the fact that the gains for two subframes are jointly quantized. 572 if (mode != MODE_4k75 || subframe & 1) 573 p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX); 574} 575/// @} 576 577 578/// @defgroup amr_gain_decoding AMR gain decoding functions 579/// @{ 580 581/** 582 * fixed gain smoothing 583 * Note that where the spec specifies the "spectrum in the q domain" 584 * in section 6.1.4, in fact frequencies should be used. 585 * 586 * @param p the context 587 * @param lsf LSFs for the current subframe, in the range [0,1] 588 * @param lsf_avg averaged LSFs 589 * @param mode mode of the current frame 590 * 591 * @return fixed gain smoothed 592 */ 593static float fixed_gain_smooth(AMRContext *p , const float *lsf, 594 const float *lsf_avg, const enum Mode mode) 595{ 596 float diff = 0.0; 597 int i; 598 599 for (i = 0; i < LP_FILTER_ORDER; i++) 600 diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i]; 601 602 // If diff is large for ten subframes, disable smoothing for a 40-subframe 603 // hangover period. 604 p->diff_count++; 605 if (diff <= 0.65) 606 p->diff_count = 0; 607 608 if (p->diff_count > 10) { 609 p->hang_count = 0; 610 p->diff_count--; // don't let diff_count overflow 611 } 612 613 if (p->hang_count < 40) { 614 p->hang_count++; 615 } else if (mode < MODE_7k4 || mode == MODE_10k2) { 616 const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0); 617 const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] + 618 p->fixed_gain[2] + p->fixed_gain[3] + 619 p->fixed_gain[4]) * 0.2; 620 return smoothing_factor * p->fixed_gain[4] + 621 (1.0 - smoothing_factor) * fixed_gain_mean; 622 } 623 return p->fixed_gain[4]; 624} 625 626/** 627 * Decode pitch gain and fixed gain factor (part of section 6.1.3). 628 * 629 * @param p the context 630 * @param amr_subframe unpacked amr subframe 631 * @param mode mode of the current frame 632 * @param subframe current subframe number 633 * @param fixed_gain_factor decoded gain correction factor 634 */ 635static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe, 636 const enum Mode mode, const int subframe, 637 float *fixed_gain_factor) 638{ 639 if (mode == MODE_12k2 || mode == MODE_7k95) { 640 p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ] 641 * (1.0 / 16384.0); 642 *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain] 643 * (1.0 / 2048.0); 644 } else { 645 const uint16_t *gains; 646 647 if (mode >= MODE_6k7) { 648 gains = gains_high[amr_subframe->p_gain]; 649 } else if (mode >= MODE_5k15) { 650 gains = gains_low [amr_subframe->p_gain]; 651 } else { 652 // gain index is only coded in subframes 0,2 for MODE_4k75 653 gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)]; 654 } 655 656 p->pitch_gain[4] = gains[0] * (1.0 / 16384.0); 657 *fixed_gain_factor = gains[1] * (1.0 / 4096.0); 658 } 659} 660 661/// @} 662 663 664/// @defgroup amr_pre_processing AMR pre-processing functions 665/// @{ 666 667/** 668 * Circularly convolve a sparse fixed vector with a phase dispersion impulse 669 * response filter (D.6.2 of G.729 and 6.1.5 of AMR). 670 * 671 * @param out vector with filter applied 672 * @param in source vector 673 * @param filter phase filter coefficients 674 * 675 * out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] } 676 */ 677static void apply_ir_filter(float *out, const AMRFixed *in, 678 const float *filter) 679{ 680 float filter1[AMR_SUBFRAME_SIZE], //!< filters at pitch lag*1 and *2 681 filter2[AMR_SUBFRAME_SIZE]; 682 int lag = in->pitch_lag; 683 float fac = in->pitch_fac; 684 int i; 685 686 if (lag < AMR_SUBFRAME_SIZE) { 687 ff_celp_circ_addf(filter1, filter, filter, lag, fac, 688 AMR_SUBFRAME_SIZE); 689 690 if (lag < AMR_SUBFRAME_SIZE >> 1) 691 ff_celp_circ_addf(filter2, filter, filter1, lag, fac, 692 AMR_SUBFRAME_SIZE); 693 } 694 695 memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE); 696 for (i = 0; i < in->n; i++) { 697 int x = in->x[i]; 698 float y = in->y[i]; 699 const float *filterp; 700 701 if (x >= AMR_SUBFRAME_SIZE - lag) { 702 filterp = filter; 703 } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) { 704 filterp = filter1; 705 } else 706 filterp = filter2; 707 708 ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE); 709 } 710} 711 712/** 713 * Reduce fixed vector sparseness by smoothing with one of three IR filters. 714 * Also know as "adaptive phase dispersion". 715 * 716 * This implements 3GPP TS 26.090 section 6.1(5). 717 * 718 * @param p the context 719 * @param fixed_sparse algebraic codebook vector 720 * @param fixed_vector unfiltered fixed vector 721 * @param fixed_gain smoothed gain 722 * @param out space for modified vector if necessary 723 */ 724static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse, 725 const float *fixed_vector, 726 float fixed_gain, float *out) 727{ 728 int ir_filter_nr; 729 730 if (p->pitch_gain[4] < 0.6) { 731 ir_filter_nr = 0; // strong filtering 732 } else if (p->pitch_gain[4] < 0.9) { 733 ir_filter_nr = 1; // medium filtering 734 } else 735 ir_filter_nr = 2; // no filtering 736 737 // detect 'onset' 738 if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) { 739 p->ir_filter_onset = 2; 740 } else if (p->ir_filter_onset) 741 p->ir_filter_onset--; 742 743 if (!p->ir_filter_onset) { 744 int i, count = 0; 745 746 for (i = 0; i < 5; i++) 747 if (p->pitch_gain[i] < 0.6) 748 count++; 749 if (count > 2) 750 ir_filter_nr = 0; 751 752 if (ir_filter_nr > p->prev_ir_filter_nr + 1) 753 ir_filter_nr--; 754 } else if (ir_filter_nr < 2) 755 ir_filter_nr++; 756 757 // Disable filtering for very low level of fixed_gain. 758 // Note this step is not specified in the technical description but is in 759 // the reference source in the function Ph_disp. 760 if (fixed_gain < 5.0) 761 ir_filter_nr = 2; 762 763 if (p->cur_frame_mode != MODE_7k4 && p->cur_frame_mode < MODE_10k2 764 && ir_filter_nr < 2) { 765 apply_ir_filter(out, fixed_sparse, 766 (p->cur_frame_mode == MODE_7k95 ? 767 ir_filters_lookup_MODE_7k95 : 768 ir_filters_lookup)[ir_filter_nr]); 769 fixed_vector = out; 770 } 771 772 // update ir filter strength history 773 p->prev_ir_filter_nr = ir_filter_nr; 774 p->prev_sparse_fixed_gain = fixed_gain; 775 776 return fixed_vector; 777} 778 779/// @} 780 781 782/// @defgroup amr_synthesis AMR synthesis functions 783/// @{ 784 785/** 786 * Conduct 10th order linear predictive coding synthesis. 787 * 788 * @param p pointer to the AMRContext 789 * @param lpc pointer to the LPC coefficients 790 * @param fixed_gain fixed codebook gain for synthesis 791 * @param fixed_vector algebraic codebook vector 792 * @param samples pointer to the output speech samples 793 * @param overflow 16-bit overflow flag 794 */ 795static int synthesis(AMRContext *p, float *lpc, 796 float fixed_gain, const float *fixed_vector, 797 float *samples, uint8_t overflow) 798{ 799 int i; 800 float excitation[AMR_SUBFRAME_SIZE]; 801 802 // if an overflow has been detected, the pitch vector is scaled down by a 803 // factor of 4 804 if (overflow) 805 for (i = 0; i < AMR_SUBFRAME_SIZE; i++) 806 p->pitch_vector[i] *= 0.25; 807 808 ff_weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector, 809 p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE); 810 811 // emphasize pitch vector contribution 812 if (p->pitch_gain[4] > 0.5 && !overflow) { 813 float energy = ff_dot_productf(excitation, excitation, 814 AMR_SUBFRAME_SIZE); 815 float pitch_factor = 816 p->pitch_gain[4] * 817 (p->cur_frame_mode == MODE_12k2 ? 818 0.25 * FFMIN(p->pitch_gain[4], 1.0) : 819 0.5 * FFMIN(p->pitch_gain[4], SHARP_MAX)); 820 821 for (i = 0; i < AMR_SUBFRAME_SIZE; i++) 822 excitation[i] += pitch_factor * p->pitch_vector[i]; 823 824 ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy, 825 AMR_SUBFRAME_SIZE); 826 } 827 828 ff_celp_lp_synthesis_filterf(samples, lpc, excitation, AMR_SUBFRAME_SIZE, 829 LP_FILTER_ORDER); 830 831 // detect overflow 832 for (i = 0; i < AMR_SUBFRAME_SIZE; i++) 833 if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) { 834 return 1; 835 } 836 837 return 0; 838} 839 840/// @} 841 842 843/// @defgroup amr_update AMR update functions 844/// @{ 845 846/** 847 * Update buffers and history at the end of decoding a subframe. 848 * 849 * @param p pointer to the AMRContext 850 */ 851static void update_state(AMRContext *p) 852{ 853 memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0])); 854 855 memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE], 856 (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float)); 857 858 memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float)); 859 memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float)); 860 861 memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE], 862 LP_FILTER_ORDER * sizeof(float)); 863} 864 865/// @} 866 867 868/// @defgroup amr_postproc AMR Post processing functions 869/// @{ 870 871/** 872 * Get the tilt factor of a formant filter from its transfer function 873 * 874 * @param lpc_n LP_FILTER_ORDER coefficients of the numerator 875 * @param lpc_d LP_FILTER_ORDER coefficients of the denominator 876 */ 877static float tilt_factor(float *lpc_n, float *lpc_d) 878{ 879 float rh0, rh1; // autocorrelation at lag 0 and 1 880 881 // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf 882 float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 }; 883 float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response 884 885 hf[0] = 1.0; 886 memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER); 887 ff_celp_lp_synthesis_filterf(hf, lpc_d, hf, AMR_TILT_RESPONSE, 888 LP_FILTER_ORDER); 889 890 rh0 = ff_dot_productf(hf, hf, AMR_TILT_RESPONSE); 891 rh1 = ff_dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1); 892 893 // The spec only specifies this check for 12.2 and 10.2 kbit/s 894 // modes. But in the ref source the tilt is always non-negative. 895 return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0; 896} 897 898/** 899 * Perform adaptive post-filtering to enhance the quality of the speech. 900 * See section 6.2.1. 901 * 902 * @param p pointer to the AMRContext 903 * @param lpc interpolated LP coefficients for this subframe 904 * @param buf_out output of the filter 905 */ 906static void postfilter(AMRContext *p, float *lpc, float *buf_out) 907{ 908 int i; 909 float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input 910 911 float speech_gain = ff_dot_productf(samples, samples, 912 AMR_SUBFRAME_SIZE); 913 914 float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter 915 const float *gamma_n, *gamma_d; // Formant filter factor table 916 float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients 917 918 if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) { 919 gamma_n = ff_pow_0_7; 920 gamma_d = ff_pow_0_75; 921 } else { 922 gamma_n = ff_pow_0_55; 923 gamma_d = ff_pow_0_7; 924 } 925 926 for (i = 0; i < LP_FILTER_ORDER; i++) { 927 lpc_n[i] = lpc[i] * gamma_n[i]; 928 lpc_d[i] = lpc[i] * gamma_d[i]; 929 } 930 931 memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER); 932 ff_celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples, 933 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER); 934 memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE, 935 sizeof(float) * LP_FILTER_ORDER); 936 937 ff_celp_lp_zero_synthesis_filterf(buf_out, lpc_n, 938 pole_out + LP_FILTER_ORDER, 939 AMR_SUBFRAME_SIZE, LP_FILTER_ORDER); 940 941 ff_tilt_compensation(&p->tilt_mem, tilt_factor(lpc_n, lpc_d), buf_out, 942 AMR_SUBFRAME_SIZE); 943 944 ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE, 945 AMR_AGC_ALPHA, &p->postfilter_agc); 946} 947 948/// @} 949 950static int amrnb_decode_frame(AVCodecContext *avctx, void *data, int *data_size, 951 AVPacket *avpkt) 952{ 953 954 AMRContext *p = avctx->priv_data; // pointer to private data 955 const uint8_t *buf = avpkt->data; 956 int buf_size = avpkt->size; 957 float *buf_out = data; // pointer to the output data buffer 958 int i, subframe; 959 float fixed_gain_factor; 960 AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing 961 float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing 962 float synth_fixed_gain; // the fixed gain that synthesis should use 963 const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use 964 965 p->cur_frame_mode = unpack_bitstream(p, buf, buf_size); 966 if (p->cur_frame_mode == MODE_DTX) { 967 av_log_missing_feature(avctx, "dtx mode", 1); 968 return -1; 969 } 970 971 if (p->cur_frame_mode == MODE_12k2) { 972 lsf2lsp_5(p); 973 } else 974 lsf2lsp_3(p); 975 976 for (i = 0; i < 4; i++) 977 ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5); 978 979 for (subframe = 0; subframe < 4; subframe++) { 980 const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe]; 981 982 decode_pitch_vector(p, amr_subframe, subframe); 983 984 decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses, 985 p->cur_frame_mode, subframe); 986 987 // The fixed gain (section 6.1.3) depends on the fixed vector 988 // (section 6.1.2), but the fixed vector calculation uses 989 // pitch sharpening based on the on the pitch gain (section 6.1.3). 990 // So the correct order is: pitch gain, pitch sharpening, fixed gain. 991 decode_gains(p, amr_subframe, p->cur_frame_mode, subframe, 992 &fixed_gain_factor); 993 994 pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse); 995 996 ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0, 997 AMR_SUBFRAME_SIZE); 998 999 p->fixed_gain[4] = 1000 ff_amr_set_fixed_gain(fixed_gain_factor, 1001 ff_dot_productf(p->fixed_vector, p->fixed_vector, 1002 AMR_SUBFRAME_SIZE)/AMR_SUBFRAME_SIZE, 1003 p->prediction_error, 1004 energy_mean[p->cur_frame_mode], energy_pred_fac); 1005 1006 // The excitation feedback is calculated without any processing such 1007 // as fixed gain smoothing. This isn't mentioned in the specification. 1008 for (i = 0; i < AMR_SUBFRAME_SIZE; i++) 1009 p->excitation[i] *= p->pitch_gain[4]; 1010 ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4], 1011 AMR_SUBFRAME_SIZE); 1012 1013 // In the ref decoder, excitation is stored with no fractional bits. 1014 // This step prevents buzz in silent periods. The ref encoder can 1015 // emit long sequences with pitch factor greater than one. This 1016 // creates unwanted feedback if the excitation vector is nonzero. 1017 // (e.g. test sequence T19_795.COD in 3GPP TS 26.074) 1018 for (i = 0; i < AMR_SUBFRAME_SIZE; i++) 1019 p->excitation[i] = truncf(p->excitation[i]); 1020 1021 // Smooth fixed gain. 1022 // The specification is ambiguous, but in the reference source, the 1023 // smoothed value is NOT fed back into later fixed gain smoothing. 1024 synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe], 1025 p->lsf_avg, p->cur_frame_mode); 1026 1027 synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector, 1028 synth_fixed_gain, spare_vector); 1029 1030 if (synthesis(p, p->lpc[subframe], synth_fixed_gain, 1031 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0)) 1032 // overflow detected -> rerun synthesis scaling pitch vector down 1033 // by a factor of 4, skipping pitch vector contribution emphasis 1034 // and adaptive gain control 1035 synthesis(p, p->lpc[subframe], synth_fixed_gain, 1036 synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1); 1037 1038 postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE); 1039 1040 // update buffers and history 1041 ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE); 1042 update_state(p); 1043 } 1044 1045 ff_acelp_apply_order_2_transfer_function(buf_out, buf_out, highpass_zeros, 1046 highpass_poles, 1047 highpass_gain * AMR_SAMPLE_SCALE, 1048 p->high_pass_mem, AMR_BLOCK_SIZE); 1049 1050 /* Update averaged lsf vector (used for fixed gain smoothing). 1051 * 1052 * Note that lsf_avg should not incorporate the current frame's LSFs 1053 * for fixed_gain_smooth. 1054 * The specification has an incorrect formula: the reference decoder uses 1055 * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */ 1056 ff_weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3], 1057 0.84, 0.16, LP_FILTER_ORDER); 1058 1059 /* report how many samples we got */ 1060 *data_size = AMR_BLOCK_SIZE * sizeof(float); 1061 1062 /* return the amount of bytes consumed if everything was OK */ 1063 return frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC 1064} 1065 1066 1067AVCodec amrnb_decoder = { 1068 .name = "amrnb", 1069 .type = AVMEDIA_TYPE_AUDIO, 1070 .id = CODEC_ID_AMR_NB, 1071 .priv_data_size = sizeof(AMRContext), 1072 .init = amrnb_decode_init, 1073 .decode = amrnb_decode_frame, 1074 .long_name = NULL_IF_CONFIG_SMALL("Adaptive Multi-Rate NarrowBand"), 1075 .sample_fmts = (enum SampleFormat[]){SAMPLE_FMT_FLT,SAMPLE_FMT_NONE}, 1076}; 1077