| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * G.729, G729 Annex D postfilter | ||
| 3 | * Copyright (c) 2008 Vladimir Voroshilov | ||
| 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 | #include <stdint.h> | ||
| 23 | #include <string.h> | ||
| 24 | |||
| 25 | #include "libavutil/common.h" | ||
| 26 | #include "libavutil/intmath.h" | ||
| 27 | |||
| 28 | #include "audiodsp.h" | ||
| 29 | #include "g729.h" | ||
| 30 | #include "g729postfilter.h" | ||
| 31 | #include "celp_math.h" | ||
| 32 | #include "acelp_filters.h" | ||
| 33 | #include "acelp_vectors.h" | ||
| 34 | #include "celp_filters.h" | ||
| 35 | |||
| 36 | #define FRAC_BITS 15 | ||
| 37 | #include "mathops.h" | ||
| 38 | |||
| 39 | /** | ||
| 40 | * short interpolation filter (of length 33, according to spec) | ||
| 41 | * for computing signal with non-integer delay | ||
| 42 | */ | ||
| 43 | static const int16_t ff_g729_interp_filt_short[(ANALYZED_FRAC_DELAYS+1)*SHORT_INT_FILT_LEN] = { | ||
| 44 | 0, 31650, 28469, 23705, 18050, 12266, 7041, 2873, | ||
| 45 | 0, -1597, -2147, -1992, -1492, -933, -484, -188, | ||
| 46 | }; | ||
| 47 | |||
| 48 | /** | ||
| 49 | * long interpolation filter (of length 129, according to spec) | ||
| 50 | * for computing signal with non-integer delay | ||
| 51 | */ | ||
| 52 | static const int16_t ff_g729_interp_filt_long[(ANALYZED_FRAC_DELAYS+1)*LONG_INT_FILT_LEN] = { | ||
| 53 | 0, 31915, 29436, 25569, 20676, 15206, 9639, 4439, | ||
| 54 | 0, -3390, -5579, -6549, -6414, -5392, -3773, -1874, | ||
| 55 | 0, 1595, 2727, 3303, 3319, 2850, 2030, 1023, | ||
| 56 | 0, -887, -1527, -1860, -1876, -1614, -1150, -579, | ||
| 57 | 0, 501, 859, 1041, 1044, 892, 631, 315, | ||
| 58 | 0, -266, -453, -543, -538, -455, -317, -156, | ||
| 59 | 0, 130, 218, 258, 253, 212, 147, 72, | ||
| 60 | 0, -59, -101, -122, -123, -106, -77, -40, | ||
| 61 | }; | ||
| 62 | |||
| 63 | /** | ||
| 64 | * formant_pp_factor_num_pow[i] = FORMANT_PP_FACTOR_NUM^(i+1) | ||
| 65 | */ | ||
| 66 | static const int16_t formant_pp_factor_num_pow[10]= { | ||
| 67 | /* (0.15) */ | ||
| 68 | 18022, 9912, 5451, 2998, 1649, 907, 499, 274, 151, 83 | ||
| 69 | }; | ||
| 70 | |||
| 71 | /** | ||
| 72 | * formant_pp_factor_den_pow[i] = FORMANT_PP_FACTOR_DEN^(i+1) | ||
| 73 | */ | ||
| 74 | static const int16_t formant_pp_factor_den_pow[10] = { | ||
| 75 | /* (0.15) */ | ||
| 76 | 22938, 16057, 11240, 7868, 5508, 3856, 2699, 1889, 1322, 925 | ||
| 77 | }; | ||
| 78 | |||
| 79 | /** | ||
| 80 | * \brief Residual signal calculation (4.2.1 if G.729) | ||
| 81 | * \param[out] out output data filtered through A(z/FORMANT_PP_FACTOR_NUM) | ||
| 82 | * \param filter_coeffs (3.12) A(z/FORMANT_PP_FACTOR_NUM) filter coefficients | ||
| 83 | * \param in input speech data to process | ||
| 84 | * \param subframe_size size of one subframe | ||
| 85 | * | ||
| 86 | * \note in buffer must contain 10 items of previous speech data before top of the buffer | ||
| 87 | * \remark It is safe to pass the same buffer for input and output. | ||
| 88 | */ | ||
| 89 | ✗ | static void residual_filter(int16_t* out, const int16_t* filter_coeffs, const int16_t* in, | |
| 90 | int subframe_size) | ||
| 91 | { | ||
| 92 | int i, n; | ||
| 93 | |||
| 94 | ✗ | for (n = subframe_size - 1; n >= 0; n--) { | |
| 95 | ✗ | int sum = 0x800; | |
| 96 | ✗ | for (i = 0; i < 10; i++) | |
| 97 | ✗ | sum += filter_coeffs[i] * in[n - i - 1]; | |
| 98 | |||
| 99 | ✗ | out[n] = in[n] + (sum >> 12); | |
| 100 | } | ||
| 101 | ✗ | } | |
| 102 | |||
| 103 | /** | ||
| 104 | * \brief long-term postfilter (4.2.1) | ||
| 105 | * \param dsp initialized DSP context | ||
| 106 | * \param pitch_delay_int integer part of the pitch delay in the first subframe | ||
| 107 | * \param residual filtering input data | ||
| 108 | * \param[out] residual_filt speech signal with applied A(z/FORMANT_PP_FACTOR_NUM) filter | ||
| 109 | * \param subframe_size size of subframe | ||
| 110 | * | ||
| 111 | * \return 0 if long-term prediction gain is less than 3dB, 1 - otherwise | ||
| 112 | */ | ||
| 113 | ✗ | static int16_t long_term_filter(AudioDSPContext *adsp, int pitch_delay_int, | |
| 114 | const int16_t* residual, int16_t *residual_filt, | ||
| 115 | int subframe_size) | ||
| 116 | { | ||
| 117 | int i, k, tmp, tmp2; | ||
| 118 | int sum; | ||
| 119 | int L_temp0; | ||
| 120 | int L_temp1; | ||
| 121 | int64_t L64_temp0; | ||
| 122 | int64_t L64_temp1; | ||
| 123 | int16_t shift; | ||
| 124 | int corr_int_num, corr_int_den; | ||
| 125 | |||
| 126 | int ener; | ||
| 127 | int16_t sh_ener; | ||
| 128 | |||
| 129 | int16_t gain_num,gain_den; //selected signal's gain numerator and denominator | ||
| 130 | int16_t sh_gain_num, sh_gain_den; | ||
| 131 | int gain_num_square; | ||
| 132 | |||
| 133 | int16_t gain_long_num,gain_long_den; //filtered through long interpolation filter signal's gain numerator and denominator | ||
| 134 | int16_t sh_gain_long_num, sh_gain_long_den; | ||
| 135 | |||
| 136 | int16_t best_delay_int, best_delay_frac; | ||
| 137 | |||
| 138 | int16_t delayed_signal_offset; | ||
| 139 | int lt_filt_factor_a, lt_filt_factor_b; | ||
| 140 | |||
| 141 | int16_t * selected_signal; | ||
| 142 | const int16_t * selected_signal_const; //Necessary to avoid compiler warning | ||
| 143 | |||
| 144 | int16_t sig_scaled[SUBFRAME_SIZE + RES_PREV_DATA_SIZE]; | ||
| 145 | int16_t delayed_signal[ANALYZED_FRAC_DELAYS][SUBFRAME_SIZE+1]; | ||
| 146 | int corr_den[ANALYZED_FRAC_DELAYS][2]; | ||
| 147 | |||
| 148 | ✗ | tmp = 0; | |
| 149 | ✗ | for(i=0; i<subframe_size + RES_PREV_DATA_SIZE; i++) | |
| 150 | ✗ | tmp |= FFABS(residual[i]); | |
| 151 | |||
| 152 | ✗ | if(!tmp) | |
| 153 | ✗ | shift = 3; | |
| 154 | else | ||
| 155 | ✗ | shift = av_log2(tmp) - 11; | |
| 156 | |||
| 157 | ✗ | if (shift > 0) | |
| 158 | ✗ | for (i = 0; i < subframe_size + RES_PREV_DATA_SIZE; i++) | |
| 159 | ✗ | sig_scaled[i] = residual[i] >> shift; | |
| 160 | else | ||
| 161 | ✗ | for (i = 0; i < subframe_size + RES_PREV_DATA_SIZE; i++) | |
| 162 | ✗ | sig_scaled[i] = (unsigned)residual[i] << -shift; | |
| 163 | |||
| 164 | /* Start of best delay searching code */ | ||
| 165 | ✗ | gain_num = 0; | |
| 166 | |||
| 167 | ✗ | ener = adsp->scalarproduct_int16(sig_scaled + RES_PREV_DATA_SIZE, | |
| 168 | sig_scaled + RES_PREV_DATA_SIZE, | ||
| 169 | subframe_size); | ||
| 170 | ✗ | if (ener) { | |
| 171 | ✗ | sh_ener = av_log2(ener) - 14; | |
| 172 | ✗ | sh_ener = FFMAX(sh_ener, 0); | |
| 173 | ✗ | ener >>= sh_ener; | |
| 174 | /* Search for best pitch delay. | ||
| 175 | |||
| 176 | sum{ r(n) * r(k,n) ] }^2 | ||
| 177 | R'(k)^2 := ------------------------- | ||
| 178 | sum{ r(k,n) * r(k,n) } | ||
| 179 | |||
| 180 | |||
| 181 | R(T) := sum{ r(n) * r(n-T) ] } | ||
| 182 | |||
| 183 | |||
| 184 | where | ||
| 185 | r(n-T) is integer delayed signal with delay T | ||
| 186 | r(k,n) is non-integer delayed signal with integer delay best_delay | ||
| 187 | and fractional delay k */ | ||
| 188 | |||
| 189 | /* Find integer delay best_delay which maximizes correlation R(T). | ||
| 190 | |||
| 191 | This is also equals to numerator of R'(0), | ||
| 192 | since the fine search (second step) is done with 1/8 | ||
| 193 | precision around best_delay. */ | ||
| 194 | ✗ | corr_int_num = 0; | |
| 195 | ✗ | best_delay_int = pitch_delay_int - 1; | |
| 196 | ✗ | for (i = pitch_delay_int - 1; i <= pitch_delay_int + 1; i++) { | |
| 197 | ✗ | sum = adsp->scalarproduct_int16(sig_scaled + RES_PREV_DATA_SIZE, | |
| 198 | ✗ | sig_scaled + RES_PREV_DATA_SIZE - i, | |
| 199 | subframe_size); | ||
| 200 | ✗ | if (sum > corr_int_num) { | |
| 201 | ✗ | corr_int_num = sum; | |
| 202 | ✗ | best_delay_int = i; | |
| 203 | } | ||
| 204 | } | ||
| 205 | ✗ | if (corr_int_num) { | |
| 206 | /* Compute denominator of pseudo-normalized correlation R'(0). */ | ||
| 207 | ✗ | corr_int_den = adsp->scalarproduct_int16(sig_scaled + RES_PREV_DATA_SIZE - best_delay_int, | |
| 208 | ✗ | sig_scaled + RES_PREV_DATA_SIZE - best_delay_int, | |
| 209 | subframe_size); | ||
| 210 | |||
| 211 | /* Compute signals with non-integer delay k (with 1/8 precision), | ||
| 212 | where k is in [0;6] range. | ||
| 213 | Entire delay is qual to best_delay+(k+1)/8 | ||
| 214 | This is achieved by applying an interpolation filter of | ||
| 215 | length 33 to source signal. */ | ||
| 216 | ✗ | for (k = 0; k < ANALYZED_FRAC_DELAYS; k++) { | |
| 217 | ✗ | ff_acelp_interpolate(&delayed_signal[k][0], | |
| 218 | ✗ | &sig_scaled[RES_PREV_DATA_SIZE - best_delay_int], | |
| 219 | ff_g729_interp_filt_short, | ||
| 220 | ANALYZED_FRAC_DELAYS+1, | ||
| 221 | 8 - k - 1, | ||
| 222 | SHORT_INT_FILT_LEN, | ||
| 223 | subframe_size + 1); | ||
| 224 | } | ||
| 225 | |||
| 226 | /* Compute denominator of pseudo-normalized correlation R'(k). | ||
| 227 | |||
| 228 | corr_den[k][0] is square root of R'(k) denominator, for int(T) == int(T0) | ||
| 229 | corr_den[k][1] is square root of R'(k) denominator, for int(T) == int(T0)+1 | ||
| 230 | |||
| 231 | Also compute maximum value of above denominators over all k. */ | ||
| 232 | ✗ | tmp = corr_int_den; | |
| 233 | ✗ | for (k = 0; k < ANALYZED_FRAC_DELAYS; k++) { | |
| 234 | ✗ | sum = adsp->scalarproduct_int16(&delayed_signal[k][1], | |
| 235 | ✗ | &delayed_signal[k][1], | |
| 236 | subframe_size - 1); | ||
| 237 | ✗ | corr_den[k][0] = sum + delayed_signal[k][0 ] * delayed_signal[k][0 ]; | |
| 238 | ✗ | corr_den[k][1] = sum + delayed_signal[k][subframe_size] * delayed_signal[k][subframe_size]; | |
| 239 | |||
| 240 | ✗ | tmp = FFMAX3(tmp, corr_den[k][0], corr_den[k][1]); | |
| 241 | } | ||
| 242 | |||
| 243 | ✗ | sh_gain_den = av_log2(tmp) - 14; | |
| 244 | ✗ | if (sh_gain_den >= 0) { | |
| 245 | |||
| 246 | ✗ | sh_gain_num = FFMAX(sh_gain_den, sh_ener); | |
| 247 | /* Loop through all k and find delay that maximizes | ||
| 248 | R'(k) correlation. | ||
| 249 | Search is done in [int(T0)-1; intT(0)+1] range | ||
| 250 | with 1/8 precision. */ | ||
| 251 | ✗ | delayed_signal_offset = 1; | |
| 252 | ✗ | best_delay_frac = 0; | |
| 253 | ✗ | gain_den = corr_int_den >> sh_gain_den; | |
| 254 | ✗ | gain_num = corr_int_num >> sh_gain_num; | |
| 255 | ✗ | gain_num_square = gain_num * gain_num; | |
| 256 | ✗ | for (k = 0; k < ANALYZED_FRAC_DELAYS; k++) { | |
| 257 | ✗ | for (i = 0; i < 2; i++) { | |
| 258 | int16_t gain_num_short, gain_den_short; | ||
| 259 | int gain_num_short_square; | ||
| 260 | /* Compute numerator of pseudo-normalized | ||
| 261 | correlation R'(k). */ | ||
| 262 | ✗ | sum = adsp->scalarproduct_int16(&delayed_signal[k][i], | |
| 263 | sig_scaled + RES_PREV_DATA_SIZE, | ||
| 264 | subframe_size); | ||
| 265 | ✗ | gain_num_short = FFMAX(sum >> sh_gain_num, 0); | |
| 266 | |||
| 267 | /* | ||
| 268 | gain_num_short_square gain_num_square | ||
| 269 | R'(T)^2 = -----------------------, max R'(T)^2= -------------- | ||
| 270 | den gain_den | ||
| 271 | */ | ||
| 272 | ✗ | gain_num_short_square = gain_num_short * gain_num_short; | |
| 273 | ✗ | gain_den_short = corr_den[k][i] >> sh_gain_den; | |
| 274 | |||
| 275 | ✗ | tmp = MULL(gain_num_short_square, gain_den, FRAC_BITS); | |
| 276 | ✗ | tmp2 = MULL(gain_num_square, gain_den_short, FRAC_BITS); | |
| 277 | |||
| 278 | // R'(T)^2 > max R'(T)^2 | ||
| 279 | ✗ | if (tmp > tmp2) { | |
| 280 | ✗ | gain_num = gain_num_short; | |
| 281 | ✗ | gain_den = gain_den_short; | |
| 282 | ✗ | gain_num_square = gain_num_short_square; | |
| 283 | ✗ | delayed_signal_offset = i; | |
| 284 | ✗ | best_delay_frac = k + 1; | |
| 285 | } | ||
| 286 | } | ||
| 287 | } | ||
| 288 | |||
| 289 | /* | ||
| 290 | R'(T)^2 | ||
| 291 | 2 * --------- < 1 | ||
| 292 | R(0) | ||
| 293 | */ | ||
| 294 | ✗ | L64_temp0 = (int64_t)gain_num_square << ((sh_gain_num << 1) + 1); | |
| 295 | ✗ | L64_temp1 = ((int64_t)gain_den * ener) << (sh_gain_den + sh_ener); | |
| 296 | ✗ | if (L64_temp0 < L64_temp1) | |
| 297 | ✗ | gain_num = 0; | |
| 298 | } // if(sh_gain_den >= 0) | ||
| 299 | } // if(corr_int_num) | ||
| 300 | } // if(ener) | ||
| 301 | /* End of best delay searching code */ | ||
| 302 | |||
| 303 | ✗ | if (!gain_num) { | |
| 304 | ✗ | memcpy(residual_filt, residual + RES_PREV_DATA_SIZE, subframe_size * sizeof(int16_t)); | |
| 305 | |||
| 306 | /* Long-term prediction gain is less than 3dB. Long-term postfilter is disabled. */ | ||
| 307 | ✗ | return 0; | |
| 308 | } | ||
| 309 | ✗ | if (best_delay_frac) { | |
| 310 | /* Recompute delayed signal with an interpolation filter of length 129. */ | ||
| 311 | ✗ | ff_acelp_interpolate(residual_filt, | |
| 312 | ✗ | &sig_scaled[RES_PREV_DATA_SIZE - best_delay_int + delayed_signal_offset], | |
| 313 | ff_g729_interp_filt_long, | ||
| 314 | ANALYZED_FRAC_DELAYS + 1, | ||
| 315 | 8 - best_delay_frac, | ||
| 316 | LONG_INT_FILT_LEN, | ||
| 317 | subframe_size + 1); | ||
| 318 | /* Compute R'(k) correlation's numerator. */ | ||
| 319 | ✗ | sum = adsp->scalarproduct_int16(residual_filt, | |
| 320 | sig_scaled + RES_PREV_DATA_SIZE, | ||
| 321 | subframe_size); | ||
| 322 | |||
| 323 | ✗ | if (sum < 0) { | |
| 324 | ✗ | gain_long_num = 0; | |
| 325 | ✗ | sh_gain_long_num = 0; | |
| 326 | } else { | ||
| 327 | ✗ | tmp = av_log2(sum) - 14; | |
| 328 | ✗ | tmp = FFMAX(tmp, 0); | |
| 329 | ✗ | sum >>= tmp; | |
| 330 | ✗ | gain_long_num = sum; | |
| 331 | ✗ | sh_gain_long_num = tmp; | |
| 332 | } | ||
| 333 | |||
| 334 | /* Compute R'(k) correlation's denominator. */ | ||
| 335 | ✗ | sum = adsp->scalarproduct_int16(residual_filt, residual_filt, subframe_size); | |
| 336 | |||
| 337 | ✗ | tmp = av_log2(sum) - 14; | |
| 338 | ✗ | tmp = FFMAX(tmp, 0); | |
| 339 | ✗ | sum >>= tmp; | |
| 340 | ✗ | gain_long_den = sum; | |
| 341 | ✗ | sh_gain_long_den = tmp; | |
| 342 | |||
| 343 | /* Select between original and delayed signal. | ||
| 344 | Delayed signal will be selected if it increases R'(k) | ||
| 345 | correlation. */ | ||
| 346 | ✗ | L_temp0 = gain_num * gain_num; | |
| 347 | ✗ | L_temp0 = MULL(L_temp0, gain_long_den, FRAC_BITS); | |
| 348 | |||
| 349 | ✗ | L_temp1 = gain_long_num * gain_long_num; | |
| 350 | ✗ | L_temp1 = MULL(L_temp1, gain_den, FRAC_BITS); | |
| 351 | |||
| 352 | ✗ | tmp = ((sh_gain_long_num - sh_gain_num) * 2) - (sh_gain_long_den - sh_gain_den); | |
| 353 | ✗ | if (tmp > 0) | |
| 354 | ✗ | L_temp0 >>= tmp; | |
| 355 | else | ||
| 356 | ✗ | L_temp1 >>= FFMIN(-tmp, 31); | |
| 357 | |||
| 358 | /* Check if longer filter increases the values of R'(k). */ | ||
| 359 | ✗ | if (L_temp1 > L_temp0) { | |
| 360 | /* Select long filter. */ | ||
| 361 | ✗ | selected_signal = residual_filt; | |
| 362 | ✗ | gain_num = gain_long_num; | |
| 363 | ✗ | gain_den = gain_long_den; | |
| 364 | ✗ | sh_gain_num = sh_gain_long_num; | |
| 365 | ✗ | sh_gain_den = sh_gain_long_den; | |
| 366 | } else | ||
| 367 | /* Select short filter. */ | ||
| 368 | ✗ | selected_signal = &delayed_signal[best_delay_frac-1][delayed_signal_offset]; | |
| 369 | |||
| 370 | /* Rescale selected signal to original value. */ | ||
| 371 | ✗ | if (shift > 0) | |
| 372 | ✗ | for (i = 0; i < subframe_size; i++) | |
| 373 | ✗ | selected_signal[i] *= 1 << shift; | |
| 374 | else | ||
| 375 | ✗ | for (i = 0; i < subframe_size; i++) | |
| 376 | ✗ | selected_signal[i] >>= -shift; | |
| 377 | |||
| 378 | /* necessary to avoid compiler warning */ | ||
| 379 | ✗ | selected_signal_const = selected_signal; | |
| 380 | } // if(best_delay_frac) | ||
| 381 | else | ||
| 382 | ✗ | selected_signal_const = residual + RES_PREV_DATA_SIZE - (best_delay_int + 1 - delayed_signal_offset); | |
| 383 | #ifdef G729_BITEXACT | ||
| 384 | tmp = sh_gain_num - sh_gain_den; | ||
| 385 | if (tmp > 0) | ||
| 386 | gain_den >>= tmp; | ||
| 387 | else | ||
| 388 | gain_num >>= -tmp; | ||
| 389 | |||
| 390 | if (gain_num > gain_den) | ||
| 391 | lt_filt_factor_a = MIN_LT_FILT_FACTOR_A; | ||
| 392 | else { | ||
| 393 | gain_num >>= 2; | ||
| 394 | gain_den >>= 1; | ||
| 395 | lt_filt_factor_a = (gain_den << 15) / (gain_den + gain_num); | ||
| 396 | } | ||
| 397 | #else | ||
| 398 | ✗ | L64_temp0 = (((int64_t)gain_num) << sh_gain_num) >> 1; | |
| 399 | ✗ | L64_temp1 = ((int64_t)gain_den) << sh_gain_den; | |
| 400 | ✗ | lt_filt_factor_a = FFMAX((L64_temp1 << 15) / (L64_temp1 + L64_temp0), MIN_LT_FILT_FACTOR_A); | |
| 401 | #endif | ||
| 402 | |||
| 403 | /* Filter through selected filter. */ | ||
| 404 | ✗ | lt_filt_factor_b = 32767 - lt_filt_factor_a + 1; | |
| 405 | |||
| 406 | ✗ | ff_acelp_weighted_vector_sum(residual_filt, residual + RES_PREV_DATA_SIZE, | |
| 407 | selected_signal_const, | ||
| 408 | lt_filt_factor_a, lt_filt_factor_b, | ||
| 409 | 1<<14, 15, subframe_size); | ||
| 410 | |||
| 411 | // Long-term prediction gain is larger than 3dB. | ||
| 412 | ✗ | return 1; | |
| 413 | } | ||
| 414 | |||
| 415 | /** | ||
| 416 | * \brief Calculate reflection coefficient for tilt compensation filter (4.2.3). | ||
| 417 | * \param dsp initialized DSP context | ||
| 418 | * \param lp_gn (3.12) coefficients of A(z/FORMANT_PP_FACTOR_NUM) filter | ||
| 419 | * \param lp_gd (3.12) coefficients of A(z/FORMANT_PP_FACTOR_DEN) filter | ||
| 420 | * \param speech speech to update | ||
| 421 | * \param subframe_size size of subframe | ||
| 422 | * | ||
| 423 | * \return (3.12) reflection coefficient | ||
| 424 | * | ||
| 425 | * \remark The routine also calculates the gain term for the short-term | ||
| 426 | * filter (gf) and multiplies the speech data by 1/gf. | ||
| 427 | * | ||
| 428 | * \note All members of lp_gn, except 10-19 must be equal to zero. | ||
| 429 | */ | ||
| 430 | ✗ | static int16_t get_tilt_comp(AudioDSPContext *adsp, int16_t *lp_gn, | |
| 431 | const int16_t *lp_gd, int16_t* speech, | ||
| 432 | int subframe_size) | ||
| 433 | { | ||
| 434 | int rh1,rh0; // (3.12) | ||
| 435 | int temp; | ||
| 436 | int i; | ||
| 437 | int gain_term; | ||
| 438 | |||
| 439 | ✗ | lp_gn[10] = 4096; //1.0 in (3.12) | |
| 440 | |||
| 441 | /* Apply 1/A(z/FORMANT_PP_FACTOR_DEN) filter to hf. */ | ||
| 442 | ✗ | ff_celp_lp_synthesis_filter(lp_gn + 11, lp_gd + 1, lp_gn + 11, 22, 10, 0, 0, 0x800); | |
| 443 | /* Now lp_gn (starting with 10) contains impulse response | ||
| 444 | of A(z/FORMANT_PP_FACTOR_NUM)/A(z/FORMANT_PP_FACTOR_DEN) filter. */ | ||
| 445 | |||
| 446 | ✗ | rh0 = adsp->scalarproduct_int16(lp_gn + 10, lp_gn + 10, 20); | |
| 447 | ✗ | rh1 = adsp->scalarproduct_int16(lp_gn + 10, lp_gn + 11, 20); | |
| 448 | |||
| 449 | /* downscale to avoid overflow */ | ||
| 450 | ✗ | temp = av_log2(rh0) - 14; | |
| 451 | ✗ | if (temp > 0) { | |
| 452 | ✗ | rh0 >>= temp; | |
| 453 | ✗ | rh1 >>= temp; | |
| 454 | } | ||
| 455 | |||
| 456 | ✗ | if (FFABS(rh1) > rh0 || !rh0) | |
| 457 | ✗ | return 0; | |
| 458 | |||
| 459 | ✗ | gain_term = 0; | |
| 460 | ✗ | for (i = 0; i < 20; i++) | |
| 461 | ✗ | gain_term += FFABS(lp_gn[i + 10]); | |
| 462 | ✗ | gain_term >>= 2; // (3.12) -> (5.10) | |
| 463 | |||
| 464 | ✗ | if (gain_term > 0x400) { // 1.0 in (5.10) | |
| 465 | ✗ | temp = 0x2000000 / gain_term; // 1.0/gain_term in (0.15) | |
| 466 | ✗ | for (i = 0; i < subframe_size; i++) | |
| 467 | ✗ | speech[i] = (speech[i] * temp + 0x4000) >> 15; | |
| 468 | } | ||
| 469 | |||
| 470 | ✗ | return -(rh1 * (1 << 15)) / rh0; | |
| 471 | } | ||
| 472 | |||
| 473 | /** | ||
| 474 | * \brief Apply tilt compensation filter (4.2.3). | ||
| 475 | * \param[in,out] res_pst residual signal (partially filtered) | ||
| 476 | * \param k1 (3.12) reflection coefficient | ||
| 477 | * \param subframe_size size of subframe | ||
| 478 | * \param ht_prev_data previous data for 4.2.3, equation 86 | ||
| 479 | * | ||
| 480 | * \return new value for ht_prev_data | ||
| 481 | */ | ||
| 482 | ✗ | static int16_t apply_tilt_comp(int16_t* out, int16_t* res_pst, int refl_coeff, | |
| 483 | int subframe_size, int16_t ht_prev_data) | ||
| 484 | { | ||
| 485 | int tmp, tmp2; | ||
| 486 | int i; | ||
| 487 | int gt, ga; | ||
| 488 | int fact, sh_fact; | ||
| 489 | |||
| 490 | ✗ | if (refl_coeff > 0) { | |
| 491 | ✗ | gt = (refl_coeff * G729_TILT_FACTOR_PLUS + 0x4000) >> 15; | |
| 492 | ✗ | fact = 0x2000; // 0.5 in (0.15) | |
| 493 | ✗ | sh_fact = 14; | |
| 494 | } else { | ||
| 495 | ✗ | gt = (refl_coeff * G729_TILT_FACTOR_MINUS + 0x4000) >> 15; | |
| 496 | ✗ | fact = 0x400; // 0.5 in (3.12) | |
| 497 | ✗ | sh_fact = 11; | |
| 498 | } | ||
| 499 | ✗ | ga = (fact << 16) / av_clip_int16(32768 - FFABS(gt)); | |
| 500 | ✗ | gt >>= 1; | |
| 501 | |||
| 502 | /* Apply tilt compensation filter to signal. */ | ||
| 503 | ✗ | tmp = res_pst[subframe_size - 1]; | |
| 504 | |||
| 505 | ✗ | for (i = subframe_size - 1; i >= 1; i--) { | |
| 506 | ✗ | tmp2 = (gt * res_pst[i-1]) * 2 + 0x4000; | |
| 507 | ✗ | tmp2 = res_pst[i] + (tmp2 >> 15); | |
| 508 | |||
| 509 | ✗ | tmp2 = (tmp2 * ga + fact) >> sh_fact; | |
| 510 | ✗ | out[i] = tmp2; | |
| 511 | } | ||
| 512 | ✗ | tmp2 = (gt * ht_prev_data) * 2 + 0x4000; | |
| 513 | ✗ | tmp2 = res_pst[0] + (tmp2 >> 15); | |
| 514 | ✗ | tmp2 = (tmp2 * ga + fact) >> sh_fact; | |
| 515 | ✗ | out[0] = tmp2; | |
| 516 | |||
| 517 | ✗ | return tmp; | |
| 518 | } | ||
| 519 | |||
| 520 | ✗ | void ff_g729_postfilter(AudioDSPContext *adsp, int16_t* ht_prev_data, int* voicing, | |
| 521 | const int16_t *lp_filter_coeffs, int pitch_delay_int, | ||
| 522 | int16_t* residual, int16_t* res_filter_data, | ||
| 523 | int16_t* pos_filter_data, int16_t *speech, int subframe_size) | ||
| 524 | { | ||
| 525 | int16_t residual_filt_buf[SUBFRAME_SIZE+11]; | ||
| 526 | int16_t lp_gn[33]; // (3.12) | ||
| 527 | int16_t lp_gd[11]; // (3.12) | ||
| 528 | int tilt_comp_coeff; | ||
| 529 | int i; | ||
| 530 | |||
| 531 | /* Zero-filling is necessary for tilt-compensation filter. */ | ||
| 532 | ✗ | memset(lp_gn, 0, 33 * sizeof(int16_t)); | |
| 533 | |||
| 534 | /* Calculate A(z/FORMANT_PP_FACTOR_NUM) filter coefficients. */ | ||
| 535 | ✗ | for (i = 0; i < 10; i++) | |
| 536 | ✗ | lp_gn[i + 11] = (lp_filter_coeffs[i + 1] * formant_pp_factor_num_pow[i] + 0x4000) >> 15; | |
| 537 | |||
| 538 | /* Calculate A(z/FORMANT_PP_FACTOR_DEN) filter coefficients. */ | ||
| 539 | ✗ | for (i = 0; i < 10; i++) | |
| 540 | ✗ | lp_gd[i + 1] = (lp_filter_coeffs[i + 1] * formant_pp_factor_den_pow[i] + 0x4000) >> 15; | |
| 541 | |||
| 542 | /* residual signal calculation (one-half of short-term postfilter) */ | ||
| 543 | ✗ | memcpy(speech - 10, res_filter_data, 10 * sizeof(int16_t)); | |
| 544 | ✗ | residual_filter(residual + RES_PREV_DATA_SIZE, lp_gn + 11, speech, subframe_size); | |
| 545 | /* Save data to use it in the next subframe. */ | ||
| 546 | ✗ | memcpy(res_filter_data, speech + subframe_size - 10, 10 * sizeof(int16_t)); | |
| 547 | |||
| 548 | /* long-term filter. If long-term prediction gain is larger than 3dB (returned value is | ||
| 549 | nonzero) then declare current subframe as periodic. */ | ||
| 550 | ✗ | i = long_term_filter(adsp, pitch_delay_int, | |
| 551 | residual, residual_filt_buf + 10, | ||
| 552 | subframe_size); | ||
| 553 | ✗ | *voicing = FFMAX(*voicing, i); | |
| 554 | |||
| 555 | /* shift residual for using in next subframe */ | ||
| 556 | ✗ | memmove(residual, residual + subframe_size, RES_PREV_DATA_SIZE * sizeof(int16_t)); | |
| 557 | |||
| 558 | /* short-term filter tilt compensation */ | ||
| 559 | ✗ | tilt_comp_coeff = get_tilt_comp(adsp, lp_gn, lp_gd, residual_filt_buf + 10, subframe_size); | |
| 560 | |||
| 561 | /* Apply second half of short-term postfilter: 1/A(z/FORMANT_PP_FACTOR_DEN) */ | ||
| 562 | ✗ | ff_celp_lp_synthesis_filter(pos_filter_data + 10, lp_gd + 1, | |
| 563 | residual_filt_buf + 10, | ||
| 564 | subframe_size, 10, 0, 0, 0x800); | ||
| 565 | ✗ | memcpy(pos_filter_data, pos_filter_data + subframe_size, 10 * sizeof(int16_t)); | |
| 566 | |||
| 567 | ✗ | *ht_prev_data = apply_tilt_comp(speech, pos_filter_data + 10, tilt_comp_coeff, | |
| 568 | ✗ | subframe_size, *ht_prev_data); | |
| 569 | ✗ | } | |
| 570 | |||
| 571 | /** | ||
| 572 | * \brief Adaptive gain control (4.2.4) | ||
| 573 | * \param gain_before gain of speech before applying postfilters | ||
| 574 | * \param gain_after gain of speech after applying postfilters | ||
| 575 | * \param[in,out] speech signal buffer | ||
| 576 | * \param subframe_size length of subframe | ||
| 577 | * \param gain_prev (3.12) previous value of gain coefficient | ||
| 578 | * | ||
| 579 | * \return (3.12) last value of gain coefficient | ||
| 580 | */ | ||
| 581 | ✗ | int16_t ff_g729_adaptive_gain_control(int gain_before, int gain_after, int16_t *speech, | |
| 582 | int subframe_size, int16_t gain_prev) | ||
| 583 | { | ||
| 584 | unsigned gain; // (3.12) | ||
| 585 | int n; | ||
| 586 | int exp_before, exp_after; | ||
| 587 | |||
| 588 | ✗ | if(!gain_after && gain_before) | |
| 589 | ✗ | return 0; | |
| 590 | |||
| 591 | ✗ | if (gain_before) { | |
| 592 | |||
| 593 | ✗ | exp_before = 14 - av_log2(gain_before); | |
| 594 | ✗ | gain_before = bidir_sal(gain_before, exp_before); | |
| 595 | |||
| 596 | ✗ | exp_after = 14 - av_log2(gain_after); | |
| 597 | ✗ | gain_after = bidir_sal(gain_after, exp_after); | |
| 598 | |||
| 599 | ✗ | if (gain_before < gain_after) { | |
| 600 | ✗ | gain = (gain_before << 15) / gain_after; | |
| 601 | ✗ | gain = bidir_sal(gain, exp_after - exp_before - 1); | |
| 602 | } else { | ||
| 603 | ✗ | gain = ((gain_before - gain_after) << 14) / gain_after + 0x4000; | |
| 604 | ✗ | gain = bidir_sal(gain, exp_after - exp_before); | |
| 605 | } | ||
| 606 | ✗ | gain = FFMIN(gain, 32767); | |
| 607 | ✗ | gain = (gain * G729_AGC_FAC1 + 0x4000) >> 15; // gain * (1-0.9875) | |
| 608 | } else | ||
| 609 | ✗ | gain = 0; | |
| 610 | |||
| 611 | ✗ | for (n = 0; n < subframe_size; n++) { | |
| 612 | // gain_prev = gain + 0.9875 * gain_prev | ||
| 613 | ✗ | gain_prev = (G729_AGC_FACTOR * gain_prev + 0x4000) >> 15; | |
| 614 | ✗ | gain_prev = av_clip_int16(gain + gain_prev); | |
| 615 | ✗ | speech[n] = av_clip_int16((speech[n] * gain_prev + 0x2000) >> 14); | |
| 616 | } | ||
| 617 | ✗ | return gain_prev; | |
| 618 | } | ||
| 619 |