| 1 | /* |
| 2 | * Enhanced Variable Rate Codec, Service Option 3 decoder |
| 3 | * Copyright (c) 2013 Paul B Mahol |
| 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 | * Enhanced Variable Rate Codec, Service Option 3 decoder |
| 25 | * @author Paul B Mahol |
| 26 | */ |
| 27 | |
| 28 | #include "libavutil/mathematics.h" |
| 29 | #include "avcodec.h" |
| 30 | #include "internal.h" |
| 31 | #include "get_bits.h" |
| 32 | #include "evrcdata.h" |
| 33 | #include "acelp_vectors.h" |
| 34 | #include "lsp.h" |
| 35 | |
| 36 | #define MIN_LSP_SEP (0.05 / (2.0 * M_PI)) |
| 37 | #define MIN_DELAY 20 |
| 38 | #define MAX_DELAY 120 |
| 39 | #define NB_SUBFRAMES 3 |
| 40 | #define SUBFRAME_SIZE 54 |
| 41 | #define FILTER_ORDER 10 |
| 42 | #define ACB_SIZE 128 |
| 43 | |
| 44 | typedef enum { |
| 45 | RATE_ERRS = -1, |
| 46 | SILENCE, |
| 47 | RATE_QUANT, |
| 48 | RATE_QUARTER, |
| 49 | RATE_HALF, |
| 50 | RATE_FULL, |
| 51 | } evrc_packet_rate; |
| 52 | |
| 53 | /** |
| 54 | * EVRC-A unpacked data frame |
| 55 | */ |
| 56 | typedef struct EVRCAFrame { |
| 57 | uint8_t lpc_flag; ///< spectral change indicator |
| 58 | uint16_t lsp[4]; ///< index into LSP codebook |
| 59 | uint8_t pitch_delay; ///< pitch delay for entire frame |
| 60 | uint8_t delay_diff; ///< delay difference for entire frame |
| 61 | uint8_t acb_gain[3]; ///< adaptive codebook gain |
| 62 | uint16_t fcb_shape[3][4]; ///< fixed codebook shape |
| 63 | uint8_t fcb_gain[3]; ///< fixed codebook gain index |
| 64 | uint8_t energy_gain; ///< frame energy gain index |
| 65 | uint8_t tty; ///< tty baud rate bit |
| 66 | } EVRCAFrame; |
| 67 | |
| 68 | typedef struct EVRCContext { |
| 69 | GetBitContext gb; |
| 70 | evrc_packet_rate bitrate; |
| 71 | evrc_packet_rate last_valid_bitrate; |
| 72 | EVRCAFrame frame; |
| 73 | |
| 74 | float lspf[FILTER_ORDER]; |
| 75 | float prev_lspf[FILTER_ORDER]; |
| 76 | float synthesis[FILTER_ORDER]; |
| 77 | float postfilter_fir[FILTER_ORDER]; |
| 78 | float postfilter_iir[FILTER_ORDER]; |
| 79 | float postfilter_residual[ACB_SIZE + SUBFRAME_SIZE]; |
| 80 | float pitch_delay; |
| 81 | float prev_pitch_delay; |
| 82 | float avg_acb_gain; ///< average adaptive codebook gain |
| 83 | float avg_fcb_gain; ///< average fixed codebook gain |
| 84 | float pitch[ACB_SIZE + FILTER_ORDER + SUBFRAME_SIZE]; |
| 85 | float pitch_back[ACB_SIZE]; |
| 86 | float interpolation_coeffs[136]; |
| 87 | float energy_vector[NB_SUBFRAMES]; |
| 88 | float fade_scale; |
| 89 | float last; |
| 90 | |
| 91 | uint8_t prev_energy_gain; |
| 92 | uint8_t prev_error_flag; |
| 93 | uint8_t warned_buf_mismatch_bitrate; |
| 94 | } EVRCContext; |
| 95 | |
| 96 | /** |
| 97 | * Frame unpacking for RATE_FULL, RATE_HALF and RATE_QUANT |
| 98 | * |
| 99 | * @param e the context |
| 100 | * |
| 101 | * TIA/IS-127 Table 4.21-1 |
| 102 | */ |
| 103 | static void unpack_frame(EVRCContext *e) |
| 104 | { |
| 105 | EVRCAFrame *frame = &e->frame; |
| 106 | GetBitContext *gb = &e->gb; |
| 107 | |
| 108 | switch (e->bitrate) { |
| 109 | case RATE_FULL: |
| 110 | frame->lpc_flag = get_bits1(gb); |
| 111 | frame->lsp[0] = get_bits(gb, 6); |
| 112 | frame->lsp[1] = get_bits(gb, 6); |
| 113 | frame->lsp[2] = get_bits(gb, 9); |
| 114 | frame->lsp[3] = get_bits(gb, 7); |
| 115 | frame->pitch_delay = get_bits(gb, 7); |
| 116 | frame->delay_diff = get_bits(gb, 5); |
| 117 | frame->acb_gain[0] = get_bits(gb, 3); |
| 118 | frame->fcb_shape[0][0] = get_bits(gb, 8); |
| 119 | frame->fcb_shape[0][1] = get_bits(gb, 8); |
| 120 | frame->fcb_shape[0][2] = get_bits(gb, 8); |
| 121 | frame->fcb_shape[0][3] = get_bits(gb, 11); |
| 122 | frame->fcb_gain[0] = get_bits(gb, 5); |
| 123 | frame->acb_gain[1] = get_bits(gb, 3); |
| 124 | frame->fcb_shape[1][0] = get_bits(gb, 8); |
| 125 | frame->fcb_shape[1][1] = get_bits(gb, 8); |
| 126 | frame->fcb_shape[1][2] = get_bits(gb, 8); |
| 127 | frame->fcb_shape[1][3] = get_bits(gb, 11); |
| 128 | frame->fcb_gain [1] = get_bits(gb, 5); |
| 129 | frame->acb_gain [2] = get_bits(gb, 3); |
| 130 | frame->fcb_shape[2][0] = get_bits(gb, 8); |
| 131 | frame->fcb_shape[2][1] = get_bits(gb, 8); |
| 132 | frame->fcb_shape[2][2] = get_bits(gb, 8); |
| 133 | frame->fcb_shape[2][3] = get_bits(gb, 11); |
| 134 | frame->fcb_gain [2] = get_bits(gb, 5); |
| 135 | frame->tty = get_bits1(gb); |
| 136 | break; |
| 137 | case RATE_HALF: |
| 138 | frame->lsp [0] = get_bits(gb, 7); |
| 139 | frame->lsp [1] = get_bits(gb, 7); |
| 140 | frame->lsp [2] = get_bits(gb, 8); |
| 141 | frame->pitch_delay = get_bits(gb, 7); |
| 142 | frame->acb_gain [0] = get_bits(gb, 3); |
| 143 | frame->fcb_shape[0][0] = get_bits(gb, 10); |
| 144 | frame->fcb_gain [0] = get_bits(gb, 4); |
| 145 | frame->acb_gain [1] = get_bits(gb, 3); |
| 146 | frame->fcb_shape[1][0] = get_bits(gb, 10); |
| 147 | frame->fcb_gain [1] = get_bits(gb, 4); |
| 148 | frame->acb_gain [2] = get_bits(gb, 3); |
| 149 | frame->fcb_shape[2][0] = get_bits(gb, 10); |
| 150 | frame->fcb_gain [2] = get_bits(gb, 4); |
| 151 | break; |
| 152 | case RATE_QUANT: |
| 153 | frame->lsp [0] = get_bits(gb, 4); |
| 154 | frame->lsp [1] = get_bits(gb, 4); |
| 155 | frame->energy_gain = get_bits(gb, 8); |
| 156 | break; |
| 157 | } |
| 158 | } |
| 159 | |
| 160 | static evrc_packet_rate buf_size2bitrate(const int buf_size) |
| 161 | { |
| 162 | switch (buf_size) { |
| 163 | case 23: return RATE_FULL; |
| 164 | case 11: return RATE_HALF; |
| 165 | case 6: return RATE_QUARTER; |
| 166 | case 3: return RATE_QUANT; |
| 167 | case 1: return SILENCE; |
| 168 | } |
| 169 | |
| 170 | return RATE_ERRS; |
| 171 | } |
| 172 | |
| 173 | /** |
| 174 | * Determine the bitrate from the frame size and/or the first byte of the frame. |
| 175 | * |
| 176 | * @param avctx the AV codec context |
| 177 | * @param buf_size length of the buffer |
| 178 | * @param buf the bufffer |
| 179 | * |
| 180 | * @return the bitrate on success, |
| 181 | * RATE_ERRS if the bitrate cannot be satisfactorily determined |
| 182 | */ |
| 183 | static evrc_packet_rate determine_bitrate(AVCodecContext *avctx, |
| 184 | int *buf_size, |
| 185 | const uint8_t **buf) |
| 186 | { |
| 187 | evrc_packet_rate bitrate; |
| 188 | |
| 189 | if ((bitrate = buf_size2bitrate(*buf_size)) >= 0) { |
| 190 | if (bitrate > **buf) { |
| 191 | EVRCContext *e = avctx->priv_data; |
| 192 | if (!e->warned_buf_mismatch_bitrate) { |
| 193 | av_log(avctx, AV_LOG_WARNING, |
| 194 | "Claimed bitrate and buffer size mismatch.\n"); |
| 195 | e->warned_buf_mismatch_bitrate = 1; |
| 196 | } |
| 197 | bitrate = **buf; |
| 198 | } else if (bitrate < **buf) { |
| 199 | av_log(avctx, AV_LOG_ERROR, |
| 200 | "Buffer is too small for the claimed bitrate.\n"); |
| 201 | return RATE_ERRS; |
| 202 | } |
| 203 | (*buf)++; |
| 204 | *buf_size -= 1; |
| 205 | } else if ((bitrate = buf_size2bitrate(*buf_size + 1)) >= 0) { |
| 206 | av_log(avctx, AV_LOG_DEBUG, |
| 207 | "Bitrate byte is missing, guessing the bitrate from packet size.\n"); |
| 208 | } else |
| 209 | return RATE_ERRS; |
| 210 | |
| 211 | return bitrate; |
| 212 | } |
| 213 | |
| 214 | static void warn_insufficient_frame_quality(AVCodecContext *avctx, |
| 215 | const char *message) |
| 216 | { |
| 217 | av_log(avctx, AV_LOG_WARNING, "Frame #%d, %s\n", |
| 218 | avctx->frame_number, message); |
| 219 | } |
| 220 | |
| 221 | /** |
| 222 | * Initialize the speech codec according to the specification. |
| 223 | * |
| 224 | * TIA/IS-127 5.2 |
| 225 | */ |
| 226 | static av_cold int evrc_decode_init(AVCodecContext *avctx) |
| 227 | { |
| 228 | EVRCContext *e = avctx->priv_data; |
| 229 | int i, n, idx = 0; |
| 230 | float denom = 2.0 / (2.0 * 8.0 + 1.0); |
| 231 | |
| 232 | avctx->channels = 1; |
| 233 | avctx->channel_layout = AV_CH_LAYOUT_MONO; |
| 234 | avctx->sample_fmt = AV_SAMPLE_FMT_FLT; |
| 235 | |
| 236 | for (i = 0; i < FILTER_ORDER; i++) { |
| 237 | e->prev_lspf[i] = (i + 1) * 0.048; |
| 238 | e->synthesis[i] = 0.0; |
| 239 | } |
| 240 | |
| 241 | for (i = 0; i < ACB_SIZE; i++) |
| 242 | e->pitch[i] = e->pitch_back[i] = 0.0; |
| 243 | |
| 244 | e->last_valid_bitrate = RATE_QUANT; |
| 245 | e->prev_pitch_delay = 40.0; |
| 246 | e->fade_scale = 1.0; |
| 247 | e->prev_error_flag = 0; |
| 248 | e->avg_acb_gain = e->avg_fcb_gain = 0.0; |
| 249 | |
| 250 | for (i = 0; i < 8; i++) { |
| 251 | float tt = ((float)i - 8.0 / 2.0) / 8.0; |
| 252 | |
| 253 | for (n = -8; n <= 8; n++, idx++) { |
| 254 | float arg1 = M_PI * 0.9 * (tt - n); |
| 255 | float arg2 = M_PI * (tt - n); |
| 256 | |
| 257 | e->interpolation_coeffs[idx] = 0.9; |
| 258 | if (arg1) |
| 259 | e->interpolation_coeffs[idx] *= (0.54 + 0.46 * cos(arg2 * denom)) * |
| 260 | sin(arg1) / arg1; |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | return 0; |
| 265 | } |
| 266 | |
| 267 | /** |
| 268 | * Decode the 10 vector quantized line spectral pair frequencies from the LSP |
| 269 | * transmission codes of any bitrate and check for badly received packets. |
| 270 | * |
| 271 | * @param e the context |
| 272 | * |
| 273 | * @return 0 on success, -1 if the packet is badly received |
| 274 | * |
| 275 | * TIA/IS-127 5.2.1, 5.7.1 |
| 276 | */ |
| 277 | static int decode_lspf(EVRCContext *e) |
| 278 | { |
| 279 | const float * const *codebooks = evrc_lspq_codebooks[e->bitrate]; |
| 280 | int i, j, k = 0; |
| 281 | |
| 282 | for (i = 0; i < evrc_lspq_nb_codebooks[e->bitrate]; i++) { |
| 283 | int row_size = evrc_lspq_codebooks_row_sizes[e->bitrate][i]; |
| 284 | const float *codebook = codebooks[i]; |
| 285 | |
| 286 | for (j = 0; j < row_size; j++) |
| 287 | e->lspf[k++] = codebook[e->frame.lsp[i] * row_size + j]; |
| 288 | } |
| 289 | |
| 290 | // check for monotonic LSPs |
| 291 | for (i = 1; i < FILTER_ORDER; i++) |
| 292 | if (e->lspf[i] <= e->lspf[i - 1]) |
| 293 | return -1; |
| 294 | |
| 295 | // check for minimum separation of LSPs at the splits |
| 296 | for (i = 0, k = 0; i < evrc_lspq_nb_codebooks[e->bitrate] - 1; i++) { |
| 297 | k += evrc_lspq_codebooks_row_sizes[e->bitrate][i]; |
| 298 | if (e->lspf[k] - e->lspf[k - 1] <= MIN_LSP_SEP) |
| 299 | return -1; |
| 300 | } |
| 301 | |
| 302 | return 0; |
| 303 | } |
| 304 | |
| 305 | /* |
| 306 | * Interpolation of LSP parameters. |
| 307 | * |
| 308 | * TIA/IS-127 5.2.3.1, 5.7.3.2 |
| 309 | */ |
| 310 | static void interpolate_lsp(float *ilsp, const float *lsp, |
| 311 | const float *prev, int index) |
| 312 | { |
| 313 | static const float lsp_interpolation_factors[] = { 0.1667, 0.5, 0.8333 }; |
| 314 | ff_weighted_vector_sumf(ilsp, prev, lsp, |
| 315 | 1.0 - lsp_interpolation_factors[index], |
| 316 | lsp_interpolation_factors[index], FILTER_ORDER); |
| 317 | } |
| 318 | |
| 319 | /* |
| 320 | * Reconstruction of the delay contour. |
| 321 | * |
| 322 | * TIA/IS-127 5.2.2.3.2 |
| 323 | */ |
| 324 | static void interpolate_delay(float *dst, float current, float prev, int index) |
| 325 | { |
| 326 | static const float d_interpolation_factors[] = { 0, 0.3313, 0.6625, 1, 1 }; |
| 327 | dst[0] = (1.0 - d_interpolation_factors[index ]) * prev |
| 328 | + d_interpolation_factors[index ] * current; |
| 329 | dst[1] = (1.0 - d_interpolation_factors[index + 1]) * prev |
| 330 | + d_interpolation_factors[index + 1] * current; |
| 331 | dst[2] = (1.0 - d_interpolation_factors[index + 2]) * prev |
| 332 | + d_interpolation_factors[index + 2] * current; |
| 333 | } |
| 334 | |
| 335 | /* |
| 336 | * Convert the quantized, interpolated line spectral frequencies, |
| 337 | * to prediction coefficients. |
| 338 | * |
| 339 | * TIA/IS-127 5.2.3.2, 4.7.2.2 |
| 340 | */ |
| 341 | static void decode_predictor_coeffs(const float *ilspf, float *ilpc) |
| 342 | { |
| 343 | double lsp[FILTER_ORDER]; |
| 344 | float a[FILTER_ORDER / 2 + 1], b[FILTER_ORDER / 2 + 1]; |
| 345 | float a1[FILTER_ORDER / 2] = { 0 }; |
| 346 | float a2[FILTER_ORDER / 2] = { 0 }; |
| 347 | float b1[FILTER_ORDER / 2] = { 0 }; |
| 348 | float b2[FILTER_ORDER / 2] = { 0 }; |
| 349 | int i, k; |
| 350 | |
| 351 | ff_acelp_lsf2lspd(lsp, ilspf, FILTER_ORDER); |
| 352 | |
| 353 | for (k = 0; k <= FILTER_ORDER; k++) { |
| 354 | a[0] = k < 2 ? 0.25 : 0; |
| 355 | b[0] = k < 2 ? k < 1 ? 0.25 : -0.25 : 0; |
| 356 | |
| 357 | for (i = 0; i < FILTER_ORDER / 2; i++) { |
| 358 | a[i + 1] = a[i] - 2 * lsp[i * 2 ] * a1[i] + a2[i]; |
| 359 | b[i + 1] = b[i] - 2 * lsp[i * 2 + 1] * b1[i] + b2[i]; |
| 360 | a2[i] = a1[i]; |
| 361 | a1[i] = a[i]; |
| 362 | b2[i] = b1[i]; |
| 363 | b1[i] = b[i]; |
| 364 | } |
| 365 | |
| 366 | if (k) |
| 367 | ilpc[k - 1] = 2.0 * (a[FILTER_ORDER / 2] + b[FILTER_ORDER / 2]); |
| 368 | } |
| 369 | } |
| 370 | |
| 371 | static void bl_intrp(EVRCContext *e, float *ex, float delay) |
| 372 | { |
| 373 | float *f; |
| 374 | int offset, i, coef_idx; |
| 375 | int16_t t; |
| 376 | |
| 377 | offset = lrintf(delay); |
| 378 | |
| 379 | t = (offset - delay + 0.5) * 8.0 + 0.5; |
| 380 | if (t == 8) { |
| 381 | t = 0; |
| 382 | offset--; |
| 383 | } |
| 384 | |
| 385 | f = ex - offset - 8; |
| 386 | |
| 387 | coef_idx = t * (2 * 8 + 1); |
| 388 | |
| 389 | ex[0] = 0.0; |
| 390 | for (i = 0; i < 2 * 8 + 1; i++) |
| 391 | ex[0] += e->interpolation_coeffs[coef_idx + i] * f[i]; |
| 392 | } |
| 393 | |
| 394 | /* |
| 395 | * Adaptive codebook excitation. |
| 396 | * |
| 397 | * TIA/IS-127 5.2.2.3.3, 4.12.5.2 |
| 398 | */ |
| 399 | static void acb_excitation(EVRCContext *e, float *excitation, float gain, |
| 400 | const float delay[3], int length) |
| 401 | { |
| 402 | float denom, locdelay, dpr, invl; |
| 403 | int i; |
| 404 | |
| 405 | invl = 1.0 / ((float) length); |
| 406 | dpr = length; |
| 407 | |
| 408 | /* first at-most extra samples */ |
| 409 | denom = (delay[1] - delay[0]) * invl; |
| 410 | for (i = 0; i < dpr; i++) { |
| 411 | locdelay = delay[0] + i * denom; |
| 412 | bl_intrp(e, excitation + i, locdelay); |
| 413 | } |
| 414 | |
| 415 | denom = (delay[2] - delay[1]) * invl; |
| 416 | /* interpolation */ |
| 417 | for (i = dpr; i < dpr + 10; i++) { |
| 418 | locdelay = delay[1] + (i - dpr) * denom; |
| 419 | bl_intrp(e, excitation + i, locdelay); |
| 420 | } |
| 421 | |
| 422 | for (i = 0; i < length; i++) |
| 423 | excitation[i] *= gain; |
| 424 | } |
| 425 | |
| 426 | static void decode_8_pulses_35bits(const uint16_t *fixed_index, float *cod) |
| 427 | { |
| 428 | int i, pos1, pos2, offset; |
| 429 | |
| 430 | offset = (fixed_index[3] >> 9) & 3; |
| 431 | |
| 432 | for (i = 0; i < 3; i++) { |
| 433 | pos1 = ((fixed_index[i] & 0x7f) / 11) * 5 + ((i + offset) % 5); |
| 434 | pos2 = ((fixed_index[i] & 0x7f) % 11) * 5 + ((i + offset) % 5); |
| 435 | |
| 436 | cod[pos1] = (fixed_index[i] & 0x80) ? -1.0 : 1.0; |
| 437 | |
| 438 | if (pos2 < pos1) |
| 439 | cod[pos2] = -cod[pos1]; |
| 440 | else |
| 441 | cod[pos2] += cod[pos1]; |
| 442 | } |
| 443 | |
| 444 | pos1 = ((fixed_index[3] & 0x7f) / 11) * 5 + ((3 + offset) % 5); |
| 445 | pos2 = ((fixed_index[3] & 0x7f) % 11) * 5 + ((4 + offset) % 5); |
| 446 | |
| 447 | cod[pos1] = (fixed_index[3] & 0x100) ? -1.0 : 1.0; |
| 448 | cod[pos2] = (fixed_index[3] & 0x80 ) ? -1.0 : 1.0; |
| 449 | } |
| 450 | |
| 451 | static void decode_3_pulses_10bits(uint16_t fixed_index, float *cod) |
| 452 | { |
| 453 | float sign; |
| 454 | int pos; |
| 455 | |
| 456 | sign = (fixed_index & 0x200) ? -1.0 : 1.0; |
| 457 | |
| 458 | pos = ((fixed_index & 0x7) * 7) + 4; |
| 459 | cod[pos] += sign; |
| 460 | pos = (((fixed_index >> 3) & 0x7) * 7) + 2; |
| 461 | cod[pos] -= sign; |
| 462 | pos = (((fixed_index >> 6) & 0x7) * 7); |
| 463 | cod[pos] += sign; |
| 464 | } |
| 465 | |
| 466 | /* |
| 467 | * Reconstruction of ACELP fixed codebook excitation for full and half rate. |
| 468 | * |
| 469 | * TIA/IS-127 5.2.3.7 |
| 470 | */ |
| 471 | static void fcb_excitation(EVRCContext *e, const uint16_t *codebook, |
| 472 | float *excitation, float pitch_gain, |
| 473 | int pitch_lag, int subframe_size) |
| 474 | { |
| 475 | int i; |
| 476 | |
| 477 | if (e->bitrate == RATE_FULL) |
| 478 | decode_8_pulses_35bits(codebook, excitation); |
| 479 | else |
| 480 | decode_3_pulses_10bits(*codebook, excitation); |
| 481 | |
| 482 | pitch_gain = av_clipf(pitch_gain, 0.2, 0.9); |
| 483 | |
| 484 | for (i = pitch_lag; i < subframe_size; i++) |
| 485 | excitation[i] += pitch_gain * excitation[i - pitch_lag]; |
| 486 | } |
| 487 | |
| 488 | /** |
| 489 | * Synthesis of the decoder output signal. |
| 490 | * |
| 491 | * param[in] in input signal |
| 492 | * param[in] filter_coeffs LPC coefficients |
| 493 | * param[in/out] memory synthesis filter memory |
| 494 | * param buffer_length amount of data to process |
| 495 | * param[out] samples output samples |
| 496 | * |
| 497 | * TIA/IS-127 5.2.3.15, 5.7.3.4 |
| 498 | */ |
| 499 | static void synthesis_filter(const float *in, const float *filter_coeffs, |
| 500 | float *memory, int buffer_length, float *samples) |
| 501 | { |
| 502 | int i, j; |
| 503 | |
| 504 | for (i = 0; i < buffer_length; i++) { |
| 505 | samples[i] = in[i]; |
| 506 | for (j = FILTER_ORDER - 1; j > 0; j--) { |
| 507 | samples[i] -= filter_coeffs[j] * memory[j]; |
| 508 | memory[j] = memory[j - 1]; |
| 509 | } |
| 510 | samples[i] -= filter_coeffs[0] * memory[0]; |
| 511 | memory[0] = samples[i]; |
| 512 | } |
| 513 | } |
| 514 | |
| 515 | static void bandwidth_expansion(float *coeff, const float *inbuf, float gamma) |
| 516 | { |
| 517 | double fac = gamma; |
| 518 | int i; |
| 519 | |
| 520 | for (i = 0; i < FILTER_ORDER; i++) { |
| 521 | coeff[i] = inbuf[i] * fac; |
| 522 | fac *= gamma; |
| 523 | } |
| 524 | } |
| 525 | |
| 526 | static void residual_filter(float *output, const float *input, |
| 527 | const float *coef, float *memory, int length) |
| 528 | { |
| 529 | float sum; |
| 530 | int i, j; |
| 531 | |
| 532 | for (i = 0; i < length; i++) { |
| 533 | sum = input[i]; |
| 534 | |
| 535 | for (j = FILTER_ORDER - 1; j > 0; j--) { |
| 536 | sum += coef[j] * memory[j]; |
| 537 | memory[j] = memory[j - 1]; |
| 538 | } |
| 539 | sum += coef[0] * memory[0]; |
| 540 | memory[0] = input[i]; |
| 541 | output[i] = sum; |
| 542 | } |
| 543 | } |
| 544 | |
| 545 | /* |
| 546 | * TIA/IS-127 Table 5.9.1-1. |
| 547 | */ |
| 548 | static const struct PfCoeff { |
| 549 | float tilt; |
| 550 | float ltgain; |
| 551 | float p1; |
| 552 | float p2; |
| 553 | } postfilter_coeffs[5] = { |
| 554 | { 0.0 , 0.0 , 0.0 , 0.0 }, |
| 555 | { 0.0 , 0.0 , 0.57, 0.57 }, |
| 556 | { 0.0 , 0.0 , 0.0 , 0.0 }, |
| 557 | { 0.35, 0.50, 0.50, 0.75 }, |
| 558 | { 0.20, 0.50, 0.57, 0.75 }, |
| 559 | }; |
| 560 | |
| 561 | /* |
| 562 | * Adaptive postfilter. |
| 563 | * |
| 564 | * TIA/IS-127 5.9 |
| 565 | */ |
| 566 | static void postfilter(EVRCContext *e, float *in, const float *coeff, |
| 567 | float *out, int idx, const struct PfCoeff *pfc, |
| 568 | int length) |
| 569 | { |
| 570 | float wcoef1[FILTER_ORDER], wcoef2[FILTER_ORDER], |
| 571 | scratch[SUBFRAME_SIZE], temp[SUBFRAME_SIZE], |
| 572 | mem[SUBFRAME_SIZE]; |
| 573 | float sum1 = 0.0, sum2 = 0.0, gamma, gain; |
| 574 | float tilt = pfc->tilt; |
| 575 | int i, n, best; |
| 576 | |
| 577 | bandwidth_expansion(wcoef1, coeff, pfc->p1); |
| 578 | bandwidth_expansion(wcoef2, coeff, pfc->p2); |
| 579 | |
| 580 | /* Tilt compensation filter, TIA/IS-127 5.9.1 */ |
| 581 | for (i = 0; i < length - 1; i++) |
| 582 | sum2 += in[i] * in[i + 1]; |
| 583 | if (sum2 < 0.0) |
| 584 | tilt = 0.0; |
| 585 | |
| 586 | for (i = 0; i < length; i++) { |
| 587 | scratch[i] = in[i] - tilt * e->last; |
| 588 | e->last = in[i]; |
| 589 | } |
| 590 | |
| 591 | /* Short term residual filter, TIA/IS-127 5.9.2 */ |
| 592 | residual_filter(&e->postfilter_residual[ACB_SIZE], scratch, wcoef1, e->postfilter_fir, length); |
| 593 | |
| 594 | /* Long term postfilter */ |
| 595 | best = idx; |
| 596 | for (i = FFMIN(MIN_DELAY, idx - 3); i <= FFMAX(MAX_DELAY, idx + 3); i++) { |
| 597 | for (n = ACB_SIZE, sum2 = 0; n < ACB_SIZE + length; n++) |
| 598 | sum2 += e->postfilter_residual[n] * e->postfilter_residual[n - i]; |
| 599 | if (sum2 > sum1) { |
| 600 | sum1 = sum2; |
| 601 | best = i; |
| 602 | } |
| 603 | } |
| 604 | |
| 605 | for (i = ACB_SIZE, sum1 = 0; i < ACB_SIZE + length; i++) |
| 606 | sum1 += e->postfilter_residual[i - best] * e->postfilter_residual[i - best]; |
| 607 | for (i = ACB_SIZE, sum2 = 0; i < ACB_SIZE + length; i++) |
| 608 | sum2 += e->postfilter_residual[i] * e->postfilter_residual[i - best]; |
| 609 | |
| 610 | if (sum2 * sum1 == 0 || e->bitrate == RATE_QUANT) { |
| 611 | memcpy(temp, e->postfilter_residual + ACB_SIZE, length * sizeof(float)); |
| 612 | } else { |
| 613 | gamma = sum2 / sum1; |
| 614 | if (gamma < 0.5) |
| 615 | memcpy(temp, e->postfilter_residual + ACB_SIZE, length * sizeof(float)); |
| 616 | else { |
| 617 | gamma = FFMIN(gamma, 1.0); |
| 618 | |
| 619 | for (i = 0; i < length; i++) { |
| 620 | temp[i] = e->postfilter_residual[ACB_SIZE + i] + gamma * |
| 621 | pfc->ltgain * e->postfilter_residual[ACB_SIZE + i - best]; |
| 622 | } |
| 623 | } |
| 624 | } |
| 625 | |
| 626 | memcpy(scratch, temp, length * sizeof(float)); |
| 627 | memcpy(mem, e->postfilter_iir, FILTER_ORDER * sizeof(float)); |
| 628 | synthesis_filter(scratch, wcoef2, mem, length, scratch); |
| 629 | |
| 630 | /* Gain computation, TIA/IS-127 5.9.4-2 */ |
| 631 | for (i = 0, sum1 = 0, sum2 = 0; i < length; i++) { |
| 632 | sum1 += in[i] * in[i]; |
| 633 | sum2 += scratch[i] * scratch[i]; |
| 634 | } |
| 635 | gain = sum2 ? sqrt(sum1 / sum2) : 1.0; |
| 636 | |
| 637 | for (i = 0; i < length; i++) |
| 638 | temp[i] *= gain; |
| 639 | |
| 640 | /* Short term postfilter */ |
| 641 | synthesis_filter(temp, wcoef2, e->postfilter_iir, length, out); |
| 642 | |
| 643 | memmove(e->postfilter_residual, |
| 644 | e->postfilter_residual + length, ACB_SIZE * sizeof(float)); |
| 645 | } |
| 646 | |
| 647 | static void frame_erasure(EVRCContext *e, float *samples) |
| 648 | { |
| 649 | float ilspf[FILTER_ORDER], ilpc[FILTER_ORDER], idelay[NB_SUBFRAMES], |
| 650 | tmp[SUBFRAME_SIZE + 6], f; |
| 651 | int i, j; |
| 652 | |
| 653 | for (i = 0; i < FILTER_ORDER; i++) { |
| 654 | if (e->bitrate != RATE_QUANT) |
| 655 | e->lspf[i] = e->prev_lspf[i] * 0.875 + 0.125 * (i + 1) * 0.048; |
| 656 | else |
| 657 | e->lspf[i] = e->prev_lspf[i]; |
| 658 | } |
| 659 | |
| 660 | if (e->prev_error_flag) |
| 661 | e->avg_acb_gain *= 0.75; |
| 662 | if (e->bitrate == RATE_FULL) |
| 663 | memcpy(e->pitch_back, e->pitch, ACB_SIZE * sizeof(float)); |
| 664 | if (e->last_valid_bitrate == RATE_QUANT) |
| 665 | e->bitrate = RATE_QUANT; |
| 666 | else |
| 667 | e->bitrate = RATE_FULL; |
| 668 | |
| 669 | if (e->bitrate == RATE_FULL || e->bitrate == RATE_HALF) { |
| 670 | e->pitch_delay = e->prev_pitch_delay; |
| 671 | } else { |
| 672 | float sum = 0; |
| 673 | |
| 674 | idelay[0] = idelay[1] = idelay[2] = MIN_DELAY; |
| 675 | |
| 676 | for (i = 0; i < NB_SUBFRAMES; i++) |
| 677 | sum += evrc_energy_quant[e->prev_energy_gain][i]; |
| 678 | sum /= (float) NB_SUBFRAMES; |
| 679 | sum = pow(10, sum); |
| 680 | for (i = 0; i < NB_SUBFRAMES; i++) |
| 681 | e->energy_vector[i] = sum; |
| 682 | } |
| 683 | |
| 684 | if (fabs(e->pitch_delay - e->prev_pitch_delay) > 15) |
| 685 | e->prev_pitch_delay = e->pitch_delay; |
| 686 | |
| 687 | for (i = 0; i < NB_SUBFRAMES; i++) { |
| 688 | int subframe_size = subframe_sizes[i]; |
| 689 | int pitch_lag; |
| 690 | |
| 691 | interpolate_lsp(ilspf, e->lspf, e->prev_lspf, i); |
| 692 | |
| 693 | if (e->bitrate != RATE_QUANT) { |
| 694 | if (e->avg_acb_gain < 0.3) { |
| 695 | idelay[0] = estimation_delay[i]; |
| 696 | idelay[1] = estimation_delay[i + 1]; |
| 697 | idelay[2] = estimation_delay[i + 2]; |
| 698 | } else { |
| 699 | interpolate_delay(idelay, e->pitch_delay, e->prev_pitch_delay, i); |
| 700 | } |
| 701 | } |
| 702 | |
| 703 | pitch_lag = lrintf((idelay[1] + idelay[0]) / 2.0); |
| 704 | decode_predictor_coeffs(ilspf, ilpc); |
| 705 | |
| 706 | if (e->bitrate != RATE_QUANT) { |
| 707 | acb_excitation(e, e->pitch + ACB_SIZE, |
| 708 | e->avg_acb_gain, idelay, subframe_size); |
| 709 | for (j = 0; j < subframe_size; j++) |
| 710 | e->pitch[ACB_SIZE + j] *= e->fade_scale; |
| 711 | e->fade_scale = FFMAX(e->fade_scale - 0.05, 0.0); |
| 712 | } else { |
| 713 | for (j = 0; j < subframe_size; j++) |
| 714 | e->pitch[ACB_SIZE + j] = e->energy_vector[i]; |
| 715 | } |
| 716 | |
| 717 | memmove(e->pitch, e->pitch + subframe_size, ACB_SIZE * sizeof(float)); |
| 718 | |
| 719 | if (e->bitrate != RATE_QUANT && e->avg_acb_gain < 0.4) { |
| 720 | f = 0.1 * e->avg_fcb_gain; |
| 721 | for (j = 0; j < subframe_size; j++) |
| 722 | e->pitch[ACB_SIZE + j] += f; |
| 723 | } else if (e->bitrate == RATE_QUANT) { |
| 724 | for (j = 0; j < subframe_size; j++) |
| 725 | e->pitch[ACB_SIZE + j] = e->energy_vector[i]; |
| 726 | } |
| 727 | |
| 728 | synthesis_filter(e->pitch + ACB_SIZE, ilpc, |
| 729 | e->synthesis, subframe_size, tmp); |
| 730 | postfilter(e, tmp, ilpc, samples, pitch_lag, |
| 731 | &postfilter_coeffs[e->bitrate], subframe_size); |
| 732 | |
| 733 | samples += subframe_size; |
| 734 | } |
| 735 | } |
| 736 | |
| 737 | static int evrc_decode_frame(AVCodecContext *avctx, void *data, |
| 738 | int *got_frame_ptr, AVPacket *avpkt) |
| 739 | { |
| 740 | const uint8_t *buf = avpkt->data; |
| 741 | AVFrame *frame = data; |
| 742 | EVRCContext *e = avctx->priv_data; |
| 743 | int buf_size = avpkt->size; |
| 744 | float ilspf[FILTER_ORDER], ilpc[FILTER_ORDER], idelay[NB_SUBFRAMES]; |
| 745 | float *samples; |
| 746 | int i, j, ret, error_flag = 0; |
| 747 | |
| 748 | frame->nb_samples = 160; |
| 749 | if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) |
| 750 | return ret; |
| 751 | samples = (float *)frame->data[0]; |
| 752 | |
| 753 | if ((e->bitrate = determine_bitrate(avctx, &buf_size, &buf)) == RATE_ERRS) { |
| 754 | warn_insufficient_frame_quality(avctx, "bitrate cannot be determined."); |
| 755 | goto erasure; |
| 756 | } |
| 757 | if (e->bitrate <= SILENCE || e->bitrate == RATE_QUARTER) |
| 758 | goto erasure; |
| 759 | if (e->bitrate == RATE_QUANT && e->last_valid_bitrate == RATE_FULL |
| 760 | && !e->prev_error_flag) |
| 761 | goto erasure; |
| 762 | |
| 763 | init_get_bits(&e->gb, buf, 8 * buf_size); |
| 764 | memset(&e->frame, 0, sizeof(EVRCAFrame)); |
| 765 | |
| 766 | unpack_frame(e); |
| 767 | |
| 768 | if (e->bitrate != RATE_QUANT) { |
| 769 | uint8_t *p = (uint8_t *) &e->frame; |
| 770 | for (i = 0; i < sizeof(EVRCAFrame); i++) { |
| 771 | if (p[i]) |
| 772 | break; |
| 773 | } |
| 774 | if (i == sizeof(EVRCAFrame)) |
| 775 | goto erasure; |
| 776 | } else if (e->frame.lsp[0] == 0xf && |
| 777 | e->frame.lsp[1] == 0xf && |
| 778 | e->frame.energy_gain == 0xff) { |
| 779 | goto erasure; |
| 780 | } |
| 781 | |
| 782 | if (decode_lspf(e) < 0) |
| 783 | goto erasure; |
| 784 | |
| 785 | if (e->bitrate == RATE_FULL || e->bitrate == RATE_HALF) { |
| 786 | /* Pitch delay parameter checking as per TIA/IS-127 5.1.5.1 */ |
| 787 | if (e->frame.pitch_delay > MAX_DELAY - MIN_DELAY) |
| 788 | goto erasure; |
| 789 | |
| 790 | e->pitch_delay = e->frame.pitch_delay + MIN_DELAY; |
| 791 | |
| 792 | /* Delay diff parameter checking as per TIA/IS-127 5.1.5.2 */ |
| 793 | if (e->frame.delay_diff) { |
| 794 | int p = e->pitch_delay - e->frame.delay_diff + 16; |
| 795 | if (p < MIN_DELAY || p > MAX_DELAY) |
| 796 | goto erasure; |
| 797 | } |
| 798 | |
| 799 | /* Delay contour reconstruction as per TIA/IS-127 5.2.2.2 */ |
| 800 | if (e->frame.delay_diff && |
| 801 | e->bitrate == RATE_FULL && e->prev_error_flag) { |
| 802 | float delay; |
| 803 | |
| 804 | memcpy(e->pitch, e->pitch_back, ACB_SIZE * sizeof(float)); |
| 805 | |
| 806 | delay = e->prev_pitch_delay; |
| 807 | e->prev_pitch_delay = delay - e->frame.delay_diff + 16.0; |
| 808 | |
| 809 | if (fabs(e->pitch_delay - delay) > 15) |
| 810 | delay = e->pitch_delay; |
| 811 | |
| 812 | for (i = 0; i < NB_SUBFRAMES; i++) { |
| 813 | int subframe_size = subframe_sizes[i]; |
| 814 | |
| 815 | interpolate_delay(idelay, delay, e->prev_pitch_delay, i); |
| 816 | acb_excitation(e, e->pitch + ACB_SIZE, e->avg_acb_gain, idelay, subframe_size); |
| 817 | memmove(e->pitch, e->pitch + subframe_size, ACB_SIZE * sizeof(float)); |
| 818 | } |
| 819 | } |
| 820 | |
| 821 | /* Smoothing of the decoded delay as per TIA/IS-127 5.2.2.5 */ |
| 822 | if (fabs(e->pitch_delay - e->prev_pitch_delay) > 15) |
| 823 | e->prev_pitch_delay = e->pitch_delay; |
| 824 | |
| 825 | e->avg_acb_gain = e->avg_fcb_gain = 0.0; |
| 826 | } else { |
| 827 | idelay[0] = idelay[1] = idelay[2] = MIN_DELAY; |
| 828 | |
| 829 | /* Decode frame energy vectors as per TIA/IS-127 5.7.2 */ |
| 830 | for (i = 0; i < NB_SUBFRAMES; i++) |
| 831 | e->energy_vector[i] = pow(10, evrc_energy_quant[e->frame.energy_gain][i]); |
| 832 | e->prev_energy_gain = e->frame.energy_gain; |
| 833 | } |
| 834 | |
| 835 | for (i = 0; i < NB_SUBFRAMES; i++) { |
| 836 | float tmp[SUBFRAME_SIZE + 6] = { 0 }; |
| 837 | int subframe_size = subframe_sizes[i]; |
| 838 | int pitch_lag; |
| 839 | |
| 840 | interpolate_lsp(ilspf, e->lspf, e->prev_lspf, i); |
| 841 | |
| 842 | if (e->bitrate != RATE_QUANT) |
| 843 | interpolate_delay(idelay, e->pitch_delay, e->prev_pitch_delay, i); |
| 844 | |
| 845 | pitch_lag = lrintf((idelay[1] + idelay[0]) / 2.0); |
| 846 | decode_predictor_coeffs(ilspf, ilpc); |
| 847 | |
| 848 | /* Bandwidth expansion as per TIA/IS-127 5.2.3.3 */ |
| 849 | if (e->frame.lpc_flag && e->prev_error_flag) |
| 850 | bandwidth_expansion(ilpc, ilpc, 0.75); |
| 851 | |
| 852 | if (e->bitrate != RATE_QUANT) { |
| 853 | float acb_sum, f; |
| 854 | |
| 855 | f = exp((e->bitrate == RATE_HALF ? 0.5 : 0.25) |
| 856 | * (e->frame.fcb_gain[i] + 1)); |
| 857 | acb_sum = pitch_gain_vq[e->frame.acb_gain[i]]; |
| 858 | e->avg_acb_gain += acb_sum / NB_SUBFRAMES; |
| 859 | e->avg_fcb_gain += f / NB_SUBFRAMES; |
| 860 | |
| 861 | acb_excitation(e, e->pitch + ACB_SIZE, |
| 862 | acb_sum, idelay, subframe_size); |
| 863 | fcb_excitation(e, e->frame.fcb_shape[i], tmp, |
| 864 | acb_sum, pitch_lag, subframe_size); |
| 865 | |
| 866 | /* Total excitation generation as per TIA/IS-127 5.2.3.9 */ |
| 867 | for (j = 0; j < subframe_size; j++) |
| 868 | e->pitch[ACB_SIZE + j] += f * tmp[j]; |
| 869 | e->fade_scale = FFMIN(e->fade_scale + 0.2, 1.0); |
| 870 | } else { |
| 871 | for (j = 0; j < subframe_size; j++) |
| 872 | e->pitch[ACB_SIZE + j] = e->energy_vector[i]; |
| 873 | } |
| 874 | |
| 875 | memmove(e->pitch, e->pitch + subframe_size, ACB_SIZE * sizeof(float)); |
| 876 | |
| 877 | synthesis_filter(e->pitch + ACB_SIZE, ilpc, |
| 878 | e->synthesis, subframe_size, tmp); |
| 879 | postfilter(e, tmp, ilpc, samples, pitch_lag, |
| 880 | &postfilter_coeffs[e->bitrate], subframe_size); |
| 881 | |
| 882 | samples += subframe_size; |
| 883 | } |
| 884 | |
| 885 | if (error_flag) { |
| 886 | erasure: |
| 887 | error_flag = 1; |
| 888 | av_log(avctx, AV_LOG_WARNING, "frame erasure\n"); |
| 889 | frame_erasure(e, samples); |
| 890 | } |
| 891 | |
| 892 | memcpy(e->prev_lspf, e->lspf, sizeof(e->prev_lspf)); |
| 893 | e->prev_error_flag = error_flag; |
| 894 | e->last_valid_bitrate = e->bitrate; |
| 895 | |
| 896 | if (e->bitrate != RATE_QUANT) |
| 897 | e->prev_pitch_delay = e->pitch_delay; |
| 898 | |
| 899 | samples = (float *)frame->data[0]; |
| 900 | for (i = 0; i < 160; i++) |
| 901 | samples[i] /= 32768; |
| 902 | |
| 903 | *got_frame_ptr = 1; |
| 904 | |
| 905 | return avpkt->size; |
| 906 | } |
| 907 | |
| 908 | AVCodec ff_evrc_decoder = { |
| 909 | .name = "evrc", |
| 910 | .long_name = NULL_IF_CONFIG_SMALL("EVRC (Enhanced Variable Rate Codec)"), |
| 911 | .type = AVMEDIA_TYPE_AUDIO, |
| 912 | .id = AV_CODEC_ID_EVRC, |
| 913 | .init = evrc_decode_init, |
| 914 | .decode = evrc_decode_frame, |
| 915 | .capabilities = CODEC_CAP_DR1, |
| 916 | .priv_data_size = sizeof(EVRCContext), |
| 917 | }; |