| 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 "libavutil/channel_layout.h" |
| 47 | #include "libavutil/float_dsp.h" |
| 48 | #include "avcodec.h" |
| 49 | #include "libavutil/common.h" |
| 50 | #include "libavutil/avassert.h" |
| 51 | #include "celp_math.h" |
| 52 | #include "celp_filters.h" |
| 53 | #include "acelp_filters.h" |
| 54 | #include "acelp_vectors.h" |
| 55 | #include "acelp_pitch_delay.h" |
| 56 | #include "lsp.h" |
| 57 | #include "amr.h" |
| 58 | #include "internal.h" |
| 59 | |
| 60 | #include "amrnbdata.h" |
| 61 | |
| 62 | #define AMR_BLOCK_SIZE 160 ///< samples per frame |
| 63 | #define AMR_SAMPLE_BOUND 32768.0 ///< threshold for synthesis overflow |
| 64 | |
| 65 | /** |
| 66 | * Scale from constructed speech to [-1,1] |
| 67 | * |
| 68 | * AMR is designed to produce 16-bit PCM samples (3GPP TS 26.090 4.2) but |
| 69 | * upscales by two (section 6.2.2). |
| 70 | * |
| 71 | * Fundamentally, this scale is determined by energy_mean through |
| 72 | * the fixed vector contribution to the excitation vector. |
| 73 | */ |
| 74 | #define AMR_SAMPLE_SCALE (2.0 / 32768.0) |
| 75 | |
| 76 | /** Prediction factor for 12.2kbit/s mode */ |
| 77 | #define PRED_FAC_MODE_12k2 0.65 |
| 78 | |
| 79 | #define LSF_R_FAC (8000.0 / 32768.0) ///< LSF residual tables to Hertz |
| 80 | #define MIN_LSF_SPACING (50.0488 / 8000.0) ///< Ensures stability of LPC filter |
| 81 | #define PITCH_LAG_MIN_MODE_12k2 18 ///< Lower bound on decoded lag search in 12.2kbit/s mode |
| 82 | |
| 83 | /** Initial energy in dB. Also used for bad frames (unimplemented). */ |
| 84 | #define MIN_ENERGY -14.0 |
| 85 | |
| 86 | /** Maximum sharpening factor |
| 87 | * |
| 88 | * The specification says 0.8, which should be 13107, but the reference C code |
| 89 | * uses 13017 instead. (Amusingly the same applies to SHARP_MAX in g729dec.c.) |
| 90 | */ |
| 91 | #define SHARP_MAX 0.79449462890625 |
| 92 | |
| 93 | /** Number of impulse response coefficients used for tilt factor */ |
| 94 | #define AMR_TILT_RESPONSE 22 |
| 95 | /** Tilt factor = 1st reflection coefficient * gamma_t */ |
| 96 | #define AMR_TILT_GAMMA_T 0.8 |
| 97 | /** Adaptive gain control factor used in post-filter */ |
| 98 | #define AMR_AGC_ALPHA 0.9 |
| 99 | |
| 100 | typedef struct AMRContext { |
| 101 | AMRNBFrame frame; ///< decoded AMR parameters (lsf coefficients, codebook indexes, etc) |
| 102 | uint8_t bad_frame_indicator; ///< bad frame ? 1 : 0 |
| 103 | enum Mode cur_frame_mode; |
| 104 | |
| 105 | int16_t prev_lsf_r[LP_FILTER_ORDER]; ///< residual LSF vector from previous subframe |
| 106 | double lsp[4][LP_FILTER_ORDER]; ///< lsp vectors from current frame |
| 107 | double prev_lsp_sub4[LP_FILTER_ORDER]; ///< lsp vector for the 4th subframe of the previous frame |
| 108 | |
| 109 | float lsf_q[4][LP_FILTER_ORDER]; ///< Interpolated LSF vector for fixed gain smoothing |
| 110 | float lsf_avg[LP_FILTER_ORDER]; ///< vector of averaged lsf vector |
| 111 | |
| 112 | float lpc[4][LP_FILTER_ORDER]; ///< lpc coefficient vectors for 4 subframes |
| 113 | |
| 114 | uint8_t pitch_lag_int; ///< integer part of pitch lag from current subframe |
| 115 | |
| 116 | float excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1 + AMR_SUBFRAME_SIZE]; ///< current excitation and all necessary excitation history |
| 117 | float *excitation; ///< pointer to the current excitation vector in excitation_buf |
| 118 | |
| 119 | float pitch_vector[AMR_SUBFRAME_SIZE]; ///< adaptive code book (pitch) vector |
| 120 | float fixed_vector[AMR_SUBFRAME_SIZE]; ///< algebraic codebook (fixed) vector (must be kept zero between frames) |
| 121 | |
| 122 | float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes |
| 123 | float pitch_gain[5]; ///< quantified pitch gains for the current and previous four subframes |
| 124 | float fixed_gain[5]; ///< quantified fixed gains for the current and previous four subframes |
| 125 | |
| 126 | float beta; ///< previous pitch_gain, bounded by [0.0,SHARP_MAX] |
| 127 | uint8_t diff_count; ///< the number of subframes for which diff has been above 0.65 |
| 128 | uint8_t hang_count; ///< the number of subframes since a hangover period started |
| 129 | |
| 130 | float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness processing to determine "onset" |
| 131 | uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none |
| 132 | uint8_t ir_filter_onset; ///< flag for impulse response filter strength |
| 133 | |
| 134 | float postfilter_mem[10]; ///< previous intermediate values in the formant filter |
| 135 | float tilt_mem; ///< previous input to tilt compensation filter |
| 136 | float postfilter_agc; ///< previous factor used for adaptive gain control |
| 137 | float high_pass_mem[2]; ///< previous intermediate values in the high-pass filter |
| 138 | |
| 139 | float samples_in[LP_FILTER_ORDER + AMR_SUBFRAME_SIZE]; ///< floating point samples |
| 140 | |
| 141 | ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs |
| 142 | ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs |
| 143 | CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs |
| 144 | CELPMContext celpm_ctx; ///< context for fixed point math operations |
| 145 | |
| 146 | } AMRContext; |
| 147 | |
| 148 | /** Double version of ff_weighted_vector_sumf() */ |
| 149 | static void weighted_vector_sumd(double *out, const double *in_a, |
| 150 | const double *in_b, double weight_coeff_a, |
| 151 | double weight_coeff_b, int length) |
| 152 | { |
| 153 | int i; |
| 154 | |
| 155 | for (i = 0; i < length; i++) |
| 156 | out[i] = weight_coeff_a * in_a[i] |
| 157 | + weight_coeff_b * in_b[i]; |
| 158 | } |
| 159 | |
| 160 | static av_cold int amrnb_decode_init(AVCodecContext *avctx) |
| 161 | { |
| 162 | AMRContext *p = avctx->priv_data; |
| 163 | int i; |
| 164 | |
| 165 | if (avctx->channels > 1) { |
| 166 | avpriv_report_missing_feature(avctx, "multi-channel AMR"); |
| 167 | return AVERROR_PATCHWELCOME; |
| 168 | } |
| 169 | |
| 170 | avctx->channels = 1; |
| 171 | avctx->channel_layout = AV_CH_LAYOUT_MONO; |
| 172 | if (!avctx->sample_rate) |
| 173 | avctx->sample_rate = 8000; |
| 174 | avctx->sample_fmt = AV_SAMPLE_FMT_FLT; |
| 175 | |
| 176 | // p->excitation always points to the same position in p->excitation_buf |
| 177 | p->excitation = &p->excitation_buf[PITCH_DELAY_MAX + LP_FILTER_ORDER + 1]; |
| 178 | |
| 179 | for (i = 0; i < LP_FILTER_ORDER; i++) { |
| 180 | p->prev_lsp_sub4[i] = lsp_sub4_init[i] * 1000 / (float)(1 << 15); |
| 181 | p->lsf_avg[i] = p->lsf_q[3][i] = lsp_avg_init[i] / (float)(1 << 15); |
| 182 | } |
| 183 | |
| 184 | for (i = 0; i < 4; i++) |
| 185 | p->prediction_error[i] = MIN_ENERGY; |
| 186 | |
| 187 | ff_acelp_filter_init(&p->acelpf_ctx); |
| 188 | ff_acelp_vectors_init(&p->acelpv_ctx); |
| 189 | ff_celp_filter_init(&p->celpf_ctx); |
| 190 | ff_celp_math_init(&p->celpm_ctx); |
| 191 | |
| 192 | return 0; |
| 193 | } |
| 194 | |
| 195 | |
| 196 | /** |
| 197 | * Unpack an RFC4867 speech frame into the AMR frame mode and parameters. |
| 198 | * |
| 199 | * The order of speech bits is specified by 3GPP TS 26.101. |
| 200 | * |
| 201 | * @param p the context |
| 202 | * @param buf pointer to the input buffer |
| 203 | * @param buf_size size of the input buffer |
| 204 | * |
| 205 | * @return the frame mode |
| 206 | */ |
| 207 | static enum Mode unpack_bitstream(AMRContext *p, const uint8_t *buf, |
| 208 | int buf_size) |
| 209 | { |
| 210 | enum Mode mode; |
| 211 | |
| 212 | // Decode the first octet. |
| 213 | mode = buf[0] >> 3 & 0x0F; // frame type |
| 214 | p->bad_frame_indicator = (buf[0] & 0x4) != 0x4; // quality bit |
| 215 | |
| 216 | if (mode >= N_MODES || buf_size < frame_sizes_nb[mode] + 1) { |
| 217 | return NO_DATA; |
| 218 | } |
| 219 | |
| 220 | if (mode < MODE_DTX) |
| 221 | ff_amr_bit_reorder((uint16_t *) &p->frame, sizeof(AMRNBFrame), buf + 1, |
| 222 | amr_unpacking_bitmaps_per_mode[mode]); |
| 223 | |
| 224 | return mode; |
| 225 | } |
| 226 | |
| 227 | |
| 228 | /// @name AMR pitch LPC coefficient decoding functions |
| 229 | /// @{ |
| 230 | |
| 231 | /** |
| 232 | * Interpolate the LSF vector (used for fixed gain smoothing). |
| 233 | * The interpolation is done over all four subframes even in MODE_12k2. |
| 234 | * |
| 235 | * @param[in] ctx The Context |
| 236 | * @param[in,out] lsf_q LSFs in [0,1] for each subframe |
| 237 | * @param[in] lsf_new New LSFs in [0,1] for subframe 4 |
| 238 | */ |
| 239 | static void interpolate_lsf(ACELPVContext *ctx, float lsf_q[4][LP_FILTER_ORDER], float *lsf_new) |
| 240 | { |
| 241 | int i; |
| 242 | |
| 243 | for (i = 0; i < 4; i++) |
| 244 | ctx->weighted_vector_sumf(lsf_q[i], lsf_q[3], lsf_new, |
| 245 | 0.25 * (3 - i), 0.25 * (i + 1), |
| 246 | LP_FILTER_ORDER); |
| 247 | } |
| 248 | |
| 249 | /** |
| 250 | * Decode a set of 5 split-matrix quantized lsf indexes into an lsp vector. |
| 251 | * |
| 252 | * @param p the context |
| 253 | * @param lsp output LSP vector |
| 254 | * @param lsf_no_r LSF vector without the residual vector added |
| 255 | * @param lsf_quantizer pointers to LSF dictionary tables |
| 256 | * @param quantizer_offset offset in tables |
| 257 | * @param sign for the 3 dictionary table |
| 258 | * @param update store data for computing the next frame's LSFs |
| 259 | */ |
| 260 | static void lsf2lsp_for_mode12k2(AMRContext *p, double lsp[LP_FILTER_ORDER], |
| 261 | const float lsf_no_r[LP_FILTER_ORDER], |
| 262 | const int16_t *lsf_quantizer[5], |
| 263 | const int quantizer_offset, |
| 264 | const int sign, const int update) |
| 265 | { |
| 266 | int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector |
| 267 | float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector |
| 268 | int i; |
| 269 | |
| 270 | for (i = 0; i < LP_FILTER_ORDER >> 1; i++) |
| 271 | memcpy(&lsf_r[i << 1], &lsf_quantizer[i][quantizer_offset], |
| 272 | 2 * sizeof(*lsf_r)); |
| 273 | |
| 274 | if (sign) { |
| 275 | lsf_r[4] *= -1; |
| 276 | lsf_r[5] *= -1; |
| 277 | } |
| 278 | |
| 279 | if (update) |
| 280 | memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r)); |
| 281 | |
| 282 | for (i = 0; i < LP_FILTER_ORDER; i++) |
| 283 | lsf_q[i] = lsf_r[i] * (LSF_R_FAC / 8000.0) + lsf_no_r[i] * (1.0 / 8000.0); |
| 284 | |
| 285 | ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER); |
| 286 | |
| 287 | if (update) |
| 288 | interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q); |
| 289 | |
| 290 | ff_acelp_lsf2lspd(lsp, lsf_q, LP_FILTER_ORDER); |
| 291 | } |
| 292 | |
| 293 | /** |
| 294 | * Decode a set of 5 split-matrix quantized lsf indexes into 2 lsp vectors. |
| 295 | * |
| 296 | * @param p pointer to the AMRContext |
| 297 | */ |
| 298 | static void lsf2lsp_5(AMRContext *p) |
| 299 | { |
| 300 | const uint16_t *lsf_param = p->frame.lsf; |
| 301 | float lsf_no_r[LP_FILTER_ORDER]; // LSFs without the residual vector |
| 302 | const int16_t *lsf_quantizer[5]; |
| 303 | int i; |
| 304 | |
| 305 | lsf_quantizer[0] = lsf_5_1[lsf_param[0]]; |
| 306 | lsf_quantizer[1] = lsf_5_2[lsf_param[1]]; |
| 307 | lsf_quantizer[2] = lsf_5_3[lsf_param[2] >> 1]; |
| 308 | lsf_quantizer[3] = lsf_5_4[lsf_param[3]]; |
| 309 | lsf_quantizer[4] = lsf_5_5[lsf_param[4]]; |
| 310 | |
| 311 | for (i = 0; i < LP_FILTER_ORDER; i++) |
| 312 | lsf_no_r[i] = p->prev_lsf_r[i] * LSF_R_FAC * PRED_FAC_MODE_12k2 + lsf_5_mean[i]; |
| 313 | |
| 314 | lsf2lsp_for_mode12k2(p, p->lsp[1], lsf_no_r, lsf_quantizer, 0, lsf_param[2] & 1, 0); |
| 315 | lsf2lsp_for_mode12k2(p, p->lsp[3], lsf_no_r, lsf_quantizer, 2, lsf_param[2] & 1, 1); |
| 316 | |
| 317 | // interpolate LSP vectors at subframes 1 and 3 |
| 318 | weighted_vector_sumd(p->lsp[0], p->prev_lsp_sub4, p->lsp[1], 0.5, 0.5, LP_FILTER_ORDER); |
| 319 | weighted_vector_sumd(p->lsp[2], p->lsp[1] , p->lsp[3], 0.5, 0.5, LP_FILTER_ORDER); |
| 320 | } |
| 321 | |
| 322 | /** |
| 323 | * Decode a set of 3 split-matrix quantized lsf indexes into an lsp vector. |
| 324 | * |
| 325 | * @param p pointer to the AMRContext |
| 326 | */ |
| 327 | static void lsf2lsp_3(AMRContext *p) |
| 328 | { |
| 329 | const uint16_t *lsf_param = p->frame.lsf; |
| 330 | int16_t lsf_r[LP_FILTER_ORDER]; // residual LSF vector |
| 331 | float lsf_q[LP_FILTER_ORDER]; // quantified LSF vector |
| 332 | const int16_t *lsf_quantizer; |
| 333 | int i, j; |
| 334 | |
| 335 | lsf_quantizer = (p->cur_frame_mode == MODE_7k95 ? lsf_3_1_MODE_7k95 : lsf_3_1)[lsf_param[0]]; |
| 336 | memcpy(lsf_r, lsf_quantizer, 3 * sizeof(*lsf_r)); |
| 337 | |
| 338 | lsf_quantizer = lsf_3_2[lsf_param[1] << (p->cur_frame_mode <= MODE_5k15)]; |
| 339 | memcpy(lsf_r + 3, lsf_quantizer, 3 * sizeof(*lsf_r)); |
| 340 | |
| 341 | lsf_quantizer = (p->cur_frame_mode <= MODE_5k15 ? lsf_3_3_MODE_5k15 : lsf_3_3)[lsf_param[2]]; |
| 342 | memcpy(lsf_r + 6, lsf_quantizer, 4 * sizeof(*lsf_r)); |
| 343 | |
| 344 | // calculate mean-removed LSF vector and add mean |
| 345 | for (i = 0; i < LP_FILTER_ORDER; i++) |
| 346 | 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); |
| 347 | |
| 348 | ff_set_min_dist_lsf(lsf_q, MIN_LSF_SPACING, LP_FILTER_ORDER); |
| 349 | |
| 350 | // store data for computing the next frame's LSFs |
| 351 | interpolate_lsf(&p->acelpv_ctx, p->lsf_q, lsf_q); |
| 352 | memcpy(p->prev_lsf_r, lsf_r, LP_FILTER_ORDER * sizeof(*lsf_r)); |
| 353 | |
| 354 | ff_acelp_lsf2lspd(p->lsp[3], lsf_q, LP_FILTER_ORDER); |
| 355 | |
| 356 | // interpolate LSP vectors at subframes 1, 2 and 3 |
| 357 | for (i = 1; i <= 3; i++) |
| 358 | for(j = 0; j < LP_FILTER_ORDER; j++) |
| 359 | p->lsp[i-1][j] = p->prev_lsp_sub4[j] + |
| 360 | (p->lsp[3][j] - p->prev_lsp_sub4[j]) * 0.25 * i; |
| 361 | } |
| 362 | |
| 363 | /// @} |
| 364 | |
| 365 | |
| 366 | /// @name AMR pitch vector decoding functions |
| 367 | /// @{ |
| 368 | |
| 369 | /** |
| 370 | * Like ff_decode_pitch_lag(), but with 1/6 resolution |
| 371 | */ |
| 372 | static void decode_pitch_lag_1_6(int *lag_int, int *lag_frac, int pitch_index, |
| 373 | const int prev_lag_int, const int subframe) |
| 374 | { |
| 375 | if (subframe == 0 || subframe == 2) { |
| 376 | if (pitch_index < 463) { |
| 377 | *lag_int = (pitch_index + 107) * 10923 >> 16; |
| 378 | *lag_frac = pitch_index - *lag_int * 6 + 105; |
| 379 | } else { |
| 380 | *lag_int = pitch_index - 368; |
| 381 | *lag_frac = 0; |
| 382 | } |
| 383 | } else { |
| 384 | *lag_int = ((pitch_index + 5) * 10923 >> 16) - 1; |
| 385 | *lag_frac = pitch_index - *lag_int * 6 - 3; |
| 386 | *lag_int += av_clip(prev_lag_int - 5, PITCH_LAG_MIN_MODE_12k2, |
| 387 | PITCH_DELAY_MAX - 9); |
| 388 | } |
| 389 | } |
| 390 | |
| 391 | static void decode_pitch_vector(AMRContext *p, |
| 392 | const AMRNBSubframe *amr_subframe, |
| 393 | const int subframe) |
| 394 | { |
| 395 | int pitch_lag_int, pitch_lag_frac; |
| 396 | enum Mode mode = p->cur_frame_mode; |
| 397 | |
| 398 | if (p->cur_frame_mode == MODE_12k2) { |
| 399 | decode_pitch_lag_1_6(&pitch_lag_int, &pitch_lag_frac, |
| 400 | amr_subframe->p_lag, p->pitch_lag_int, |
| 401 | subframe); |
| 402 | } else |
| 403 | ff_decode_pitch_lag(&pitch_lag_int, &pitch_lag_frac, |
| 404 | amr_subframe->p_lag, |
| 405 | p->pitch_lag_int, subframe, |
| 406 | mode != MODE_4k75 && mode != MODE_5k15, |
| 407 | mode <= MODE_6k7 ? 4 : (mode == MODE_7k95 ? 5 : 6)); |
| 408 | |
| 409 | p->pitch_lag_int = pitch_lag_int; // store previous lag in a uint8_t |
| 410 | |
| 411 | pitch_lag_frac <<= (p->cur_frame_mode != MODE_12k2); |
| 412 | |
| 413 | pitch_lag_int += pitch_lag_frac > 0; |
| 414 | |
| 415 | /* Calculate the pitch vector by interpolating the past excitation at the |
| 416 | pitch lag using a b60 hamming windowed sinc function. */ |
| 417 | p->acelpf_ctx.acelp_interpolatef(p->excitation, |
| 418 | p->excitation + 1 - pitch_lag_int, |
| 419 | ff_b60_sinc, 6, |
| 420 | pitch_lag_frac + 6 - 6*(pitch_lag_frac > 0), |
| 421 | 10, AMR_SUBFRAME_SIZE); |
| 422 | |
| 423 | memcpy(p->pitch_vector, p->excitation, AMR_SUBFRAME_SIZE * sizeof(float)); |
| 424 | } |
| 425 | |
| 426 | /// @} |
| 427 | |
| 428 | |
| 429 | /// @name AMR algebraic code book (fixed) vector decoding functions |
| 430 | /// @{ |
| 431 | |
| 432 | /** |
| 433 | * Decode a 10-bit algebraic codebook index from a 10.2 kbit/s frame. |
| 434 | */ |
| 435 | static void decode_10bit_pulse(int code, int pulse_position[8], |
| 436 | int i1, int i2, int i3) |
| 437 | { |
| 438 | // coded using 7+3 bits with the 3 LSBs being, individually, the LSB of 1 of |
| 439 | // the 3 pulses and the upper 7 bits being coded in base 5 |
| 440 | const uint8_t *positions = base_five_table[code >> 3]; |
| 441 | pulse_position[i1] = (positions[2] << 1) + ( code & 1); |
| 442 | pulse_position[i2] = (positions[1] << 1) + ((code >> 1) & 1); |
| 443 | pulse_position[i3] = (positions[0] << 1) + ((code >> 2) & 1); |
| 444 | } |
| 445 | |
| 446 | /** |
| 447 | * Decode the algebraic codebook index to pulse positions and signs and |
| 448 | * construct the algebraic codebook vector for MODE_10k2. |
| 449 | * |
| 450 | * @param fixed_index positions of the eight pulses |
| 451 | * @param fixed_sparse pointer to the algebraic codebook vector |
| 452 | */ |
| 453 | static void decode_8_pulses_31bits(const int16_t *fixed_index, |
| 454 | AMRFixed *fixed_sparse) |
| 455 | { |
| 456 | int pulse_position[8]; |
| 457 | int i, temp; |
| 458 | |
| 459 | decode_10bit_pulse(fixed_index[4], pulse_position, 0, 4, 1); |
| 460 | decode_10bit_pulse(fixed_index[5], pulse_position, 2, 6, 5); |
| 461 | |
| 462 | // coded using 5+2 bits with the 2 LSBs being, individually, the LSB of 1 of |
| 463 | // the 2 pulses and the upper 5 bits being coded in base 5 |
| 464 | temp = ((fixed_index[6] >> 2) * 25 + 12) >> 5; |
| 465 | pulse_position[3] = temp % 5; |
| 466 | pulse_position[7] = temp / 5; |
| 467 | if (pulse_position[7] & 1) |
| 468 | pulse_position[3] = 4 - pulse_position[3]; |
| 469 | pulse_position[3] = (pulse_position[3] << 1) + ( fixed_index[6] & 1); |
| 470 | pulse_position[7] = (pulse_position[7] << 1) + ((fixed_index[6] >> 1) & 1); |
| 471 | |
| 472 | fixed_sparse->n = 8; |
| 473 | for (i = 0; i < 4; i++) { |
| 474 | const int pos1 = (pulse_position[i] << 2) + i; |
| 475 | const int pos2 = (pulse_position[i + 4] << 2) + i; |
| 476 | const float sign = fixed_index[i] ? -1.0 : 1.0; |
| 477 | fixed_sparse->x[i ] = pos1; |
| 478 | fixed_sparse->x[i + 4] = pos2; |
| 479 | fixed_sparse->y[i ] = sign; |
| 480 | fixed_sparse->y[i + 4] = pos2 < pos1 ? -sign : sign; |
| 481 | } |
| 482 | } |
| 483 | |
| 484 | /** |
| 485 | * Decode the algebraic codebook index to pulse positions and signs, |
| 486 | * then construct the algebraic codebook vector. |
| 487 | * |
| 488 | * nb of pulses | bits encoding pulses |
| 489 | * For MODE_4k75 or MODE_5k15, 2 | 1-3, 4-6, 7 |
| 490 | * MODE_5k9, 2 | 1, 2-4, 5-6, 7-9 |
| 491 | * MODE_6k7, 3 | 1-3, 4, 5-7, 8, 9-11 |
| 492 | * MODE_7k4 or MODE_7k95, 4 | 1-3, 4-6, 7-9, 10, 11-13 |
| 493 | * |
| 494 | * @param fixed_sparse pointer to the algebraic codebook vector |
| 495 | * @param pulses algebraic codebook indexes |
| 496 | * @param mode mode of the current frame |
| 497 | * @param subframe current subframe number |
| 498 | */ |
| 499 | static void decode_fixed_sparse(AMRFixed *fixed_sparse, const uint16_t *pulses, |
| 500 | const enum Mode mode, const int subframe) |
| 501 | { |
| 502 | av_assert1(MODE_4k75 <= (signed)mode && mode <= MODE_12k2); |
| 503 | |
| 504 | if (mode == MODE_12k2) { |
| 505 | ff_decode_10_pulses_35bits(pulses, fixed_sparse, gray_decode, 5, 3); |
| 506 | } else if (mode == MODE_10k2) { |
| 507 | decode_8_pulses_31bits(pulses, fixed_sparse); |
| 508 | } else { |
| 509 | int *pulse_position = fixed_sparse->x; |
| 510 | int i, pulse_subset; |
| 511 | const int fixed_index = pulses[0]; |
| 512 | |
| 513 | if (mode <= MODE_5k15) { |
| 514 | pulse_subset = ((fixed_index >> 3) & 8) + (subframe << 1); |
| 515 | pulse_position[0] = ( fixed_index & 7) * 5 + track_position[pulse_subset]; |
| 516 | pulse_position[1] = ((fixed_index >> 3) & 7) * 5 + track_position[pulse_subset + 1]; |
| 517 | fixed_sparse->n = 2; |
| 518 | } else if (mode == MODE_5k9) { |
| 519 | pulse_subset = ((fixed_index & 1) << 1) + 1; |
| 520 | pulse_position[0] = ((fixed_index >> 1) & 7) * 5 + pulse_subset; |
| 521 | pulse_subset = (fixed_index >> 4) & 3; |
| 522 | pulse_position[1] = ((fixed_index >> 6) & 7) * 5 + pulse_subset + (pulse_subset == 3 ? 1 : 0); |
| 523 | fixed_sparse->n = pulse_position[0] == pulse_position[1] ? 1 : 2; |
| 524 | } else if (mode == MODE_6k7) { |
| 525 | pulse_position[0] = (fixed_index & 7) * 5; |
| 526 | pulse_subset = (fixed_index >> 2) & 2; |
| 527 | pulse_position[1] = ((fixed_index >> 4) & 7) * 5 + pulse_subset + 1; |
| 528 | pulse_subset = (fixed_index >> 6) & 2; |
| 529 | pulse_position[2] = ((fixed_index >> 8) & 7) * 5 + pulse_subset + 2; |
| 530 | fixed_sparse->n = 3; |
| 531 | } else { // mode <= MODE_7k95 |
| 532 | pulse_position[0] = gray_decode[ fixed_index & 7]; |
| 533 | pulse_position[1] = gray_decode[(fixed_index >> 3) & 7] + 1; |
| 534 | pulse_position[2] = gray_decode[(fixed_index >> 6) & 7] + 2; |
| 535 | pulse_subset = (fixed_index >> 9) & 1; |
| 536 | pulse_position[3] = gray_decode[(fixed_index >> 10) & 7] + pulse_subset + 3; |
| 537 | fixed_sparse->n = 4; |
| 538 | } |
| 539 | for (i = 0; i < fixed_sparse->n; i++) |
| 540 | fixed_sparse->y[i] = (pulses[1] >> i) & 1 ? 1.0 : -1.0; |
| 541 | } |
| 542 | } |
| 543 | |
| 544 | /** |
| 545 | * Apply pitch lag to obtain the sharpened fixed vector (section 6.1.2) |
| 546 | * |
| 547 | * @param p the context |
| 548 | * @param subframe unpacked amr subframe |
| 549 | * @param mode mode of the current frame |
| 550 | * @param fixed_sparse sparse respresentation of the fixed vector |
| 551 | */ |
| 552 | static void pitch_sharpening(AMRContext *p, int subframe, enum Mode mode, |
| 553 | AMRFixed *fixed_sparse) |
| 554 | { |
| 555 | // The spec suggests the current pitch gain is always used, but in other |
| 556 | // modes the pitch and codebook gains are joinly quantized (sec 5.8.2) |
| 557 | // so the codebook gain cannot depend on the quantized pitch gain. |
| 558 | if (mode == MODE_12k2) |
| 559 | p->beta = FFMIN(p->pitch_gain[4], 1.0); |
| 560 | |
| 561 | fixed_sparse->pitch_lag = p->pitch_lag_int; |
| 562 | fixed_sparse->pitch_fac = p->beta; |
| 563 | |
| 564 | // Save pitch sharpening factor for the next subframe |
| 565 | // MODE_4k75 only updates on the 2nd and 4th subframes - this follows from |
| 566 | // the fact that the gains for two subframes are jointly quantized. |
| 567 | if (mode != MODE_4k75 || subframe & 1) |
| 568 | p->beta = av_clipf(p->pitch_gain[4], 0.0, SHARP_MAX); |
| 569 | } |
| 570 | /// @} |
| 571 | |
| 572 | |
| 573 | /// @name AMR gain decoding functions |
| 574 | /// @{ |
| 575 | |
| 576 | /** |
| 577 | * fixed gain smoothing |
| 578 | * Note that where the spec specifies the "spectrum in the q domain" |
| 579 | * in section 6.1.4, in fact frequencies should be used. |
| 580 | * |
| 581 | * @param p the context |
| 582 | * @param lsf LSFs for the current subframe, in the range [0,1] |
| 583 | * @param lsf_avg averaged LSFs |
| 584 | * @param mode mode of the current frame |
| 585 | * |
| 586 | * @return fixed gain smoothed |
| 587 | */ |
| 588 | static float fixed_gain_smooth(AMRContext *p , const float *lsf, |
| 589 | const float *lsf_avg, const enum Mode mode) |
| 590 | { |
| 591 | float diff = 0.0; |
| 592 | int i; |
| 593 | |
| 594 | for (i = 0; i < LP_FILTER_ORDER; i++) |
| 595 | diff += fabs(lsf_avg[i] - lsf[i]) / lsf_avg[i]; |
| 596 | |
| 597 | // If diff is large for ten subframes, disable smoothing for a 40-subframe |
| 598 | // hangover period. |
| 599 | p->diff_count++; |
| 600 | if (diff <= 0.65) |
| 601 | p->diff_count = 0; |
| 602 | |
| 603 | if (p->diff_count > 10) { |
| 604 | p->hang_count = 0; |
| 605 | p->diff_count--; // don't let diff_count overflow |
| 606 | } |
| 607 | |
| 608 | if (p->hang_count < 40) { |
| 609 | p->hang_count++; |
| 610 | } else if (mode < MODE_7k4 || mode == MODE_10k2) { |
| 611 | const float smoothing_factor = av_clipf(4.0 * diff - 1.6, 0.0, 1.0); |
| 612 | const float fixed_gain_mean = (p->fixed_gain[0] + p->fixed_gain[1] + |
| 613 | p->fixed_gain[2] + p->fixed_gain[3] + |
| 614 | p->fixed_gain[4]) * 0.2; |
| 615 | return smoothing_factor * p->fixed_gain[4] + |
| 616 | (1.0 - smoothing_factor) * fixed_gain_mean; |
| 617 | } |
| 618 | return p->fixed_gain[4]; |
| 619 | } |
| 620 | |
| 621 | /** |
| 622 | * Decode pitch gain and fixed gain factor (part of section 6.1.3). |
| 623 | * |
| 624 | * @param p the context |
| 625 | * @param amr_subframe unpacked amr subframe |
| 626 | * @param mode mode of the current frame |
| 627 | * @param subframe current subframe number |
| 628 | * @param fixed_gain_factor decoded gain correction factor |
| 629 | */ |
| 630 | static void decode_gains(AMRContext *p, const AMRNBSubframe *amr_subframe, |
| 631 | const enum Mode mode, const int subframe, |
| 632 | float *fixed_gain_factor) |
| 633 | { |
| 634 | if (mode == MODE_12k2 || mode == MODE_7k95) { |
| 635 | p->pitch_gain[4] = qua_gain_pit [amr_subframe->p_gain ] |
| 636 | * (1.0 / 16384.0); |
| 637 | *fixed_gain_factor = qua_gain_code[amr_subframe->fixed_gain] |
| 638 | * (1.0 / 2048.0); |
| 639 | } else { |
| 640 | const uint16_t *gains; |
| 641 | |
| 642 | if (mode >= MODE_6k7) { |
| 643 | gains = gains_high[amr_subframe->p_gain]; |
| 644 | } else if (mode >= MODE_5k15) { |
| 645 | gains = gains_low [amr_subframe->p_gain]; |
| 646 | } else { |
| 647 | // gain index is only coded in subframes 0,2 for MODE_4k75 |
| 648 | gains = gains_MODE_4k75[(p->frame.subframe[subframe & 2].p_gain << 1) + (subframe & 1)]; |
| 649 | } |
| 650 | |
| 651 | p->pitch_gain[4] = gains[0] * (1.0 / 16384.0); |
| 652 | *fixed_gain_factor = gains[1] * (1.0 / 4096.0); |
| 653 | } |
| 654 | } |
| 655 | |
| 656 | /// @} |
| 657 | |
| 658 | |
| 659 | /// @name AMR preprocessing functions |
| 660 | /// @{ |
| 661 | |
| 662 | /** |
| 663 | * Circularly convolve a sparse fixed vector with a phase dispersion impulse |
| 664 | * response filter (D.6.2 of G.729 and 6.1.5 of AMR). |
| 665 | * |
| 666 | * @param out vector with filter applied |
| 667 | * @param in source vector |
| 668 | * @param filter phase filter coefficients |
| 669 | * |
| 670 | * out[n] = sum(i,0,len-1){ in[i] * filter[(len + n - i)%len] } |
| 671 | */ |
| 672 | static void apply_ir_filter(float *out, const AMRFixed *in, |
| 673 | const float *filter) |
| 674 | { |
| 675 | float filter1[AMR_SUBFRAME_SIZE], ///< filters at pitch lag*1 and *2 |
| 676 | filter2[AMR_SUBFRAME_SIZE]; |
| 677 | int lag = in->pitch_lag; |
| 678 | float fac = in->pitch_fac; |
| 679 | int i; |
| 680 | |
| 681 | if (lag < AMR_SUBFRAME_SIZE) { |
| 682 | ff_celp_circ_addf(filter1, filter, filter, lag, fac, |
| 683 | AMR_SUBFRAME_SIZE); |
| 684 | |
| 685 | if (lag < AMR_SUBFRAME_SIZE >> 1) |
| 686 | ff_celp_circ_addf(filter2, filter, filter1, lag, fac, |
| 687 | AMR_SUBFRAME_SIZE); |
| 688 | } |
| 689 | |
| 690 | memset(out, 0, sizeof(float) * AMR_SUBFRAME_SIZE); |
| 691 | for (i = 0; i < in->n; i++) { |
| 692 | int x = in->x[i]; |
| 693 | float y = in->y[i]; |
| 694 | const float *filterp; |
| 695 | |
| 696 | if (x >= AMR_SUBFRAME_SIZE - lag) { |
| 697 | filterp = filter; |
| 698 | } else if (x >= AMR_SUBFRAME_SIZE - (lag << 1)) { |
| 699 | filterp = filter1; |
| 700 | } else |
| 701 | filterp = filter2; |
| 702 | |
| 703 | ff_celp_circ_addf(out, out, filterp, x, y, AMR_SUBFRAME_SIZE); |
| 704 | } |
| 705 | } |
| 706 | |
| 707 | /** |
| 708 | * Reduce fixed vector sparseness by smoothing with one of three IR filters. |
| 709 | * Also know as "adaptive phase dispersion". |
| 710 | * |
| 711 | * This implements 3GPP TS 26.090 section 6.1(5). |
| 712 | * |
| 713 | * @param p the context |
| 714 | * @param fixed_sparse algebraic codebook vector |
| 715 | * @param fixed_vector unfiltered fixed vector |
| 716 | * @param fixed_gain smoothed gain |
| 717 | * @param out space for modified vector if necessary |
| 718 | */ |
| 719 | static const float *anti_sparseness(AMRContext *p, AMRFixed *fixed_sparse, |
| 720 | const float *fixed_vector, |
| 721 | float fixed_gain, float *out) |
| 722 | { |
| 723 | int ir_filter_nr; |
| 724 | |
| 725 | if (p->pitch_gain[4] < 0.6) { |
| 726 | ir_filter_nr = 0; // strong filtering |
| 727 | } else if (p->pitch_gain[4] < 0.9) { |
| 728 | ir_filter_nr = 1; // medium filtering |
| 729 | } else |
| 730 | ir_filter_nr = 2; // no filtering |
| 731 | |
| 732 | // detect 'onset' |
| 733 | if (fixed_gain > 2.0 * p->prev_sparse_fixed_gain) { |
| 734 | p->ir_filter_onset = 2; |
| 735 | } else if (p->ir_filter_onset) |
| 736 | p->ir_filter_onset--; |
| 737 | |
| 738 | if (!p->ir_filter_onset) { |
| 739 | int i, count = 0; |
| 740 | |
| 741 | for (i = 0; i < 5; i++) |
| 742 | if (p->pitch_gain[i] < 0.6) |
| 743 | count++; |
| 744 | if (count > 2) |
| 745 | ir_filter_nr = 0; |
| 746 | |
| 747 | if (ir_filter_nr > p->prev_ir_filter_nr + 1) |
| 748 | ir_filter_nr--; |
| 749 | } else if (ir_filter_nr < 2) |
| 750 | ir_filter_nr++; |
| 751 | |
| 752 | // Disable filtering for very low level of fixed_gain. |
| 753 | // Note this step is not specified in the technical description but is in |
| 754 | // the reference source in the function Ph_disp. |
| 755 | if (fixed_gain < 5.0) |
| 756 | ir_filter_nr = 2; |
| 757 | |
| 758 | if (p->cur_frame_mode != MODE_7k4 && p->cur_frame_mode < MODE_10k2 |
| 759 | && ir_filter_nr < 2) { |
| 760 | apply_ir_filter(out, fixed_sparse, |
| 761 | (p->cur_frame_mode == MODE_7k95 ? |
| 762 | ir_filters_lookup_MODE_7k95 : |
| 763 | ir_filters_lookup)[ir_filter_nr]); |
| 764 | fixed_vector = out; |
| 765 | } |
| 766 | |
| 767 | // update ir filter strength history |
| 768 | p->prev_ir_filter_nr = ir_filter_nr; |
| 769 | p->prev_sparse_fixed_gain = fixed_gain; |
| 770 | |
| 771 | return fixed_vector; |
| 772 | } |
| 773 | |
| 774 | /// @} |
| 775 | |
| 776 | |
| 777 | /// @name AMR synthesis functions |
| 778 | /// @{ |
| 779 | |
| 780 | /** |
| 781 | * Conduct 10th order linear predictive coding synthesis. |
| 782 | * |
| 783 | * @param p pointer to the AMRContext |
| 784 | * @param lpc pointer to the LPC coefficients |
| 785 | * @param fixed_gain fixed codebook gain for synthesis |
| 786 | * @param fixed_vector algebraic codebook vector |
| 787 | * @param samples pointer to the output speech samples |
| 788 | * @param overflow 16-bit overflow flag |
| 789 | */ |
| 790 | static int synthesis(AMRContext *p, float *lpc, |
| 791 | float fixed_gain, const float *fixed_vector, |
| 792 | float *samples, uint8_t overflow) |
| 793 | { |
| 794 | int i; |
| 795 | float excitation[AMR_SUBFRAME_SIZE]; |
| 796 | |
| 797 | // if an overflow has been detected, the pitch vector is scaled down by a |
| 798 | // factor of 4 |
| 799 | if (overflow) |
| 800 | for (i = 0; i < AMR_SUBFRAME_SIZE; i++) |
| 801 | p->pitch_vector[i] *= 0.25; |
| 802 | |
| 803 | p->acelpv_ctx.weighted_vector_sumf(excitation, p->pitch_vector, fixed_vector, |
| 804 | p->pitch_gain[4], fixed_gain, AMR_SUBFRAME_SIZE); |
| 805 | |
| 806 | // emphasize pitch vector contribution |
| 807 | if (p->pitch_gain[4] > 0.5 && !overflow) { |
| 808 | float energy = p->celpm_ctx.dot_productf(excitation, excitation, |
| 809 | AMR_SUBFRAME_SIZE); |
| 810 | float pitch_factor = |
| 811 | p->pitch_gain[4] * |
| 812 | (p->cur_frame_mode == MODE_12k2 ? |
| 813 | 0.25 * FFMIN(p->pitch_gain[4], 1.0) : |
| 814 | 0.5 * FFMIN(p->pitch_gain[4], SHARP_MAX)); |
| 815 | |
| 816 | for (i = 0; i < AMR_SUBFRAME_SIZE; i++) |
| 817 | excitation[i] += pitch_factor * p->pitch_vector[i]; |
| 818 | |
| 819 | ff_scale_vector_to_given_sum_of_squares(excitation, excitation, energy, |
| 820 | AMR_SUBFRAME_SIZE); |
| 821 | } |
| 822 | |
| 823 | p->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation, |
| 824 | AMR_SUBFRAME_SIZE, |
| 825 | LP_FILTER_ORDER); |
| 826 | |
| 827 | // detect overflow |
| 828 | for (i = 0; i < AMR_SUBFRAME_SIZE; i++) |
| 829 | if (fabsf(samples[i]) > AMR_SAMPLE_BOUND) { |
| 830 | return 1; |
| 831 | } |
| 832 | |
| 833 | return 0; |
| 834 | } |
| 835 | |
| 836 | /// @} |
| 837 | |
| 838 | |
| 839 | /// @name AMR update functions |
| 840 | /// @{ |
| 841 | |
| 842 | /** |
| 843 | * Update buffers and history at the end of decoding a subframe. |
| 844 | * |
| 845 | * @param p pointer to the AMRContext |
| 846 | */ |
| 847 | static void update_state(AMRContext *p) |
| 848 | { |
| 849 | memcpy(p->prev_lsp_sub4, p->lsp[3], LP_FILTER_ORDER * sizeof(p->lsp[3][0])); |
| 850 | |
| 851 | memmove(&p->excitation_buf[0], &p->excitation_buf[AMR_SUBFRAME_SIZE], |
| 852 | (PITCH_DELAY_MAX + LP_FILTER_ORDER + 1) * sizeof(float)); |
| 853 | |
| 854 | memmove(&p->pitch_gain[0], &p->pitch_gain[1], 4 * sizeof(float)); |
| 855 | memmove(&p->fixed_gain[0], &p->fixed_gain[1], 4 * sizeof(float)); |
| 856 | |
| 857 | memmove(&p->samples_in[0], &p->samples_in[AMR_SUBFRAME_SIZE], |
| 858 | LP_FILTER_ORDER * sizeof(float)); |
| 859 | } |
| 860 | |
| 861 | /// @} |
| 862 | |
| 863 | |
| 864 | /// @name AMR Postprocessing functions |
| 865 | /// @{ |
| 866 | |
| 867 | /** |
| 868 | * Get the tilt factor of a formant filter from its transfer function |
| 869 | * |
| 870 | * @param p The Context |
| 871 | * @param lpc_n LP_FILTER_ORDER coefficients of the numerator |
| 872 | * @param lpc_d LP_FILTER_ORDER coefficients of the denominator |
| 873 | */ |
| 874 | static float tilt_factor(AMRContext *p, float *lpc_n, float *lpc_d) |
| 875 | { |
| 876 | float rh0, rh1; // autocorrelation at lag 0 and 1 |
| 877 | |
| 878 | // LP_FILTER_ORDER prior zeros are needed for ff_celp_lp_synthesis_filterf |
| 879 | float impulse_buffer[LP_FILTER_ORDER + AMR_TILT_RESPONSE] = { 0 }; |
| 880 | float *hf = impulse_buffer + LP_FILTER_ORDER; // start of impulse response |
| 881 | |
| 882 | hf[0] = 1.0; |
| 883 | memcpy(hf + 1, lpc_n, sizeof(float) * LP_FILTER_ORDER); |
| 884 | p->celpf_ctx.celp_lp_synthesis_filterf(hf, lpc_d, hf, |
| 885 | AMR_TILT_RESPONSE, |
| 886 | LP_FILTER_ORDER); |
| 887 | |
| 888 | rh0 = p->celpm_ctx.dot_productf(hf, hf, AMR_TILT_RESPONSE); |
| 889 | rh1 = p->celpm_ctx.dot_productf(hf, hf + 1, AMR_TILT_RESPONSE - 1); |
| 890 | |
| 891 | // The spec only specifies this check for 12.2 and 10.2 kbit/s |
| 892 | // modes. But in the ref source the tilt is always non-negative. |
| 893 | return rh1 >= 0.0 ? rh1 / rh0 * AMR_TILT_GAMMA_T : 0.0; |
| 894 | } |
| 895 | |
| 896 | /** |
| 897 | * Perform adaptive post-filtering to enhance the quality of the speech. |
| 898 | * See section 6.2.1. |
| 899 | * |
| 900 | * @param p pointer to the AMRContext |
| 901 | * @param lpc interpolated LP coefficients for this subframe |
| 902 | * @param buf_out output of the filter |
| 903 | */ |
| 904 | static void postfilter(AMRContext *p, float *lpc, float *buf_out) |
| 905 | { |
| 906 | int i; |
| 907 | float *samples = p->samples_in + LP_FILTER_ORDER; // Start of input |
| 908 | |
| 909 | float speech_gain = p->celpm_ctx.dot_productf(samples, samples, |
| 910 | AMR_SUBFRAME_SIZE); |
| 911 | |
| 912 | float pole_out[AMR_SUBFRAME_SIZE + LP_FILTER_ORDER]; // Output of pole filter |
| 913 | const float *gamma_n, *gamma_d; // Formant filter factor table |
| 914 | float lpc_n[LP_FILTER_ORDER], lpc_d[LP_FILTER_ORDER]; // Transfer function coefficients |
| 915 | |
| 916 | if (p->cur_frame_mode == MODE_12k2 || p->cur_frame_mode == MODE_10k2) { |
| 917 | gamma_n = ff_pow_0_7; |
| 918 | gamma_d = ff_pow_0_75; |
| 919 | } else { |
| 920 | gamma_n = ff_pow_0_55; |
| 921 | gamma_d = ff_pow_0_7; |
| 922 | } |
| 923 | |
| 924 | for (i = 0; i < LP_FILTER_ORDER; i++) { |
| 925 | lpc_n[i] = lpc[i] * gamma_n[i]; |
| 926 | lpc_d[i] = lpc[i] * gamma_d[i]; |
| 927 | } |
| 928 | |
| 929 | memcpy(pole_out, p->postfilter_mem, sizeof(float) * LP_FILTER_ORDER); |
| 930 | p->celpf_ctx.celp_lp_synthesis_filterf(pole_out + LP_FILTER_ORDER, lpc_d, samples, |
| 931 | AMR_SUBFRAME_SIZE, LP_FILTER_ORDER); |
| 932 | memcpy(p->postfilter_mem, pole_out + AMR_SUBFRAME_SIZE, |
| 933 | sizeof(float) * LP_FILTER_ORDER); |
| 934 | |
| 935 | p->celpf_ctx.celp_lp_zero_synthesis_filterf(buf_out, lpc_n, |
| 936 | pole_out + LP_FILTER_ORDER, |
| 937 | AMR_SUBFRAME_SIZE, LP_FILTER_ORDER); |
| 938 | |
| 939 | ff_tilt_compensation(&p->tilt_mem, tilt_factor(p, lpc_n, lpc_d), buf_out, |
| 940 | AMR_SUBFRAME_SIZE); |
| 941 | |
| 942 | ff_adaptive_gain_control(buf_out, buf_out, speech_gain, AMR_SUBFRAME_SIZE, |
| 943 | AMR_AGC_ALPHA, &p->postfilter_agc); |
| 944 | } |
| 945 | |
| 946 | /// @} |
| 947 | |
| 948 | static int amrnb_decode_frame(AVCodecContext *avctx, void *data, |
| 949 | int *got_frame_ptr, AVPacket *avpkt) |
| 950 | { |
| 951 | |
| 952 | AMRContext *p = avctx->priv_data; // pointer to private data |
| 953 | AVFrame *frame = data; |
| 954 | const uint8_t *buf = avpkt->data; |
| 955 | int buf_size = avpkt->size; |
| 956 | float *buf_out; // pointer to the output data buffer |
| 957 | int i, subframe, ret; |
| 958 | float fixed_gain_factor; |
| 959 | AMRFixed fixed_sparse = {0}; // fixed vector up to anti-sparseness processing |
| 960 | float spare_vector[AMR_SUBFRAME_SIZE]; // extra stack space to hold result from anti-sparseness processing |
| 961 | float synth_fixed_gain; // the fixed gain that synthesis should use |
| 962 | const float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use |
| 963 | |
| 964 | /* get output buffer */ |
| 965 | frame->nb_samples = AMR_BLOCK_SIZE; |
| 966 | if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) |
| 967 | return ret; |
| 968 | buf_out = (float *)frame->data[0]; |
| 969 | |
| 970 | p->cur_frame_mode = unpack_bitstream(p, buf, buf_size); |
| 971 | if (p->cur_frame_mode == NO_DATA) { |
| 972 | av_log(avctx, AV_LOG_ERROR, "Corrupt bitstream\n"); |
| 973 | return AVERROR_INVALIDDATA; |
| 974 | } |
| 975 | if (p->cur_frame_mode == MODE_DTX) { |
| 976 | avpriv_report_missing_feature(avctx, "dtx mode"); |
| 977 | av_log(avctx, AV_LOG_INFO, "Note: libopencore_amrnb supports dtx\n"); |
| 978 | return AVERROR_PATCHWELCOME; |
| 979 | } |
| 980 | |
| 981 | if (p->cur_frame_mode == MODE_12k2) { |
| 982 | lsf2lsp_5(p); |
| 983 | } else |
| 984 | lsf2lsp_3(p); |
| 985 | |
| 986 | for (i = 0; i < 4; i++) |
| 987 | ff_acelp_lspd2lpc(p->lsp[i], p->lpc[i], 5); |
| 988 | |
| 989 | for (subframe = 0; subframe < 4; subframe++) { |
| 990 | const AMRNBSubframe *amr_subframe = &p->frame.subframe[subframe]; |
| 991 | |
| 992 | decode_pitch_vector(p, amr_subframe, subframe); |
| 993 | |
| 994 | decode_fixed_sparse(&fixed_sparse, amr_subframe->pulses, |
| 995 | p->cur_frame_mode, subframe); |
| 996 | |
| 997 | // The fixed gain (section 6.1.3) depends on the fixed vector |
| 998 | // (section 6.1.2), but the fixed vector calculation uses |
| 999 | // pitch sharpening based on the on the pitch gain (section 6.1.3). |
| 1000 | // So the correct order is: pitch gain, pitch sharpening, fixed gain. |
| 1001 | decode_gains(p, amr_subframe, p->cur_frame_mode, subframe, |
| 1002 | &fixed_gain_factor); |
| 1003 | |
| 1004 | pitch_sharpening(p, subframe, p->cur_frame_mode, &fixed_sparse); |
| 1005 | |
| 1006 | if (fixed_sparse.pitch_lag == 0) { |
| 1007 | av_log(avctx, AV_LOG_ERROR, "The file is corrupted, pitch_lag = 0 is not allowed\n"); |
| 1008 | return AVERROR_INVALIDDATA; |
| 1009 | } |
| 1010 | ff_set_fixed_vector(p->fixed_vector, &fixed_sparse, 1.0, |
| 1011 | AMR_SUBFRAME_SIZE); |
| 1012 | |
| 1013 | p->fixed_gain[4] = |
| 1014 | ff_amr_set_fixed_gain(fixed_gain_factor, |
| 1015 | p->celpm_ctx.dot_productf(p->fixed_vector, |
| 1016 | p->fixed_vector, |
| 1017 | AMR_SUBFRAME_SIZE) / |
| 1018 | AMR_SUBFRAME_SIZE, |
| 1019 | p->prediction_error, |
| 1020 | energy_mean[p->cur_frame_mode], energy_pred_fac); |
| 1021 | |
| 1022 | // The excitation feedback is calculated without any processing such |
| 1023 | // as fixed gain smoothing. This isn't mentioned in the specification. |
| 1024 | for (i = 0; i < AMR_SUBFRAME_SIZE; i++) |
| 1025 | p->excitation[i] *= p->pitch_gain[4]; |
| 1026 | ff_set_fixed_vector(p->excitation, &fixed_sparse, p->fixed_gain[4], |
| 1027 | AMR_SUBFRAME_SIZE); |
| 1028 | |
| 1029 | // In the ref decoder, excitation is stored with no fractional bits. |
| 1030 | // This step prevents buzz in silent periods. The ref encoder can |
| 1031 | // emit long sequences with pitch factor greater than one. This |
| 1032 | // creates unwanted feedback if the excitation vector is nonzero. |
| 1033 | // (e.g. test sequence T19_795.COD in 3GPP TS 26.074) |
| 1034 | for (i = 0; i < AMR_SUBFRAME_SIZE; i++) |
| 1035 | p->excitation[i] = truncf(p->excitation[i]); |
| 1036 | |
| 1037 | // Smooth fixed gain. |
| 1038 | // The specification is ambiguous, but in the reference source, the |
| 1039 | // smoothed value is NOT fed back into later fixed gain smoothing. |
| 1040 | synth_fixed_gain = fixed_gain_smooth(p, p->lsf_q[subframe], |
| 1041 | p->lsf_avg, p->cur_frame_mode); |
| 1042 | |
| 1043 | synth_fixed_vector = anti_sparseness(p, &fixed_sparse, p->fixed_vector, |
| 1044 | synth_fixed_gain, spare_vector); |
| 1045 | |
| 1046 | if (synthesis(p, p->lpc[subframe], synth_fixed_gain, |
| 1047 | synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 0)) |
| 1048 | // overflow detected -> rerun synthesis scaling pitch vector down |
| 1049 | // by a factor of 4, skipping pitch vector contribution emphasis |
| 1050 | // and adaptive gain control |
| 1051 | synthesis(p, p->lpc[subframe], synth_fixed_gain, |
| 1052 | synth_fixed_vector, &p->samples_in[LP_FILTER_ORDER], 1); |
| 1053 | |
| 1054 | postfilter(p, p->lpc[subframe], buf_out + subframe * AMR_SUBFRAME_SIZE); |
| 1055 | |
| 1056 | // update buffers and history |
| 1057 | ff_clear_fixed_vector(p->fixed_vector, &fixed_sparse, AMR_SUBFRAME_SIZE); |
| 1058 | update_state(p); |
| 1059 | } |
| 1060 | |
| 1061 | p->acelpf_ctx.acelp_apply_order_2_transfer_function(buf_out, |
| 1062 | buf_out, highpass_zeros, |
| 1063 | highpass_poles, |
| 1064 | highpass_gain * AMR_SAMPLE_SCALE, |
| 1065 | p->high_pass_mem, AMR_BLOCK_SIZE); |
| 1066 | |
| 1067 | /* Update averaged lsf vector (used for fixed gain smoothing). |
| 1068 | * |
| 1069 | * Note that lsf_avg should not incorporate the current frame's LSFs |
| 1070 | * for fixed_gain_smooth. |
| 1071 | * The specification has an incorrect formula: the reference decoder uses |
| 1072 | * qbar(n-1) rather than qbar(n) in section 6.1(4) equation 71. */ |
| 1073 | p->acelpv_ctx.weighted_vector_sumf(p->lsf_avg, p->lsf_avg, p->lsf_q[3], |
| 1074 | 0.84, 0.16, LP_FILTER_ORDER); |
| 1075 | |
| 1076 | *got_frame_ptr = 1; |
| 1077 | |
| 1078 | /* return the amount of bytes consumed if everything was OK */ |
| 1079 | return frame_sizes_nb[p->cur_frame_mode] + 1; // +7 for rounding and +8 for TOC |
| 1080 | } |
| 1081 | |
| 1082 | |
| 1083 | AVCodec ff_amrnb_decoder = { |
| 1084 | .name = "amrnb", |
| 1085 | .long_name = NULL_IF_CONFIG_SMALL("AMR-NB (Adaptive Multi-Rate NarrowBand)"), |
| 1086 | .type = AVMEDIA_TYPE_AUDIO, |
| 1087 | .id = AV_CODEC_ID_AMR_NB, |
| 1088 | .priv_data_size = sizeof(AMRContext), |
| 1089 | .init = amrnb_decode_init, |
| 1090 | .decode = amrnb_decode_frame, |
| 1091 | .capabilities = CODEC_CAP_DR1, |
| 1092 | .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT, |
| 1093 | AV_SAMPLE_FMT_NONE }, |
| 1094 | }; |